A few weeks ago, I wrote an essay on the origins of COVID-19. I tried to keep it mostly neutral. The subject has been politicized to the point that most people paying attention have picked a side. Essays that take one side on this topic are a dime a dozen at this point.

The neutral stance partly reflects my own views on the matter. The Chinese government's obstruction of a full-fledged forensic investigation by neutral third parties has left us unable to resolve the origins question one way or the other. This is a terrible state of affairs, since both the natural spillover (zoonosis) and lab-leak scenarios present risks for future outbreaks. It sure would be good to know what actually happened in 2019 so we could focus our efforts appropriately.

Ever since the movie Outbreak, I had been interested in zoonoses and had read several popular books (Quammen, Preston, etc.) on the topic. In the waning days of the pandemic, I was reasonably sure we were facing a zoonotic scenario. I hadn’t really paid attention to the investigation into origins. Once I did, I felt that the case for a lab leak was hardly certain, but seemed very plausible, and even likely. That was more or less my view at the start of 2025. Since then, we’ve had additional scientific papers and analysis (of the meager evidence we do have). After engaging with some of those publications - the 2024 Huanan market analysis, the 2025 recombination and phylogeography paper, the 2026 selection study, as well as the debate around those papers, my views have shifted. For some time now, I’ve thought that a zoonotic origin is more likely.

This is hardly a slam-dunk case. We still have not found the animal that infected the first person or the immediate precursor virus to SARS-CoV-2. We cannot trace a particular infected mammal (or farm, trader, or truck) all the way into the Huanan market. Evidence for zoonosis that looked stronger in 2022, especially the two-lineage argument, looks shakier today.

But what tilts my opinion towards zoonosis is how the remaining evidence looks after accounting for those weaknesses. The zoonosis story has missing links. But there seems to be a reasonable amount of evidence suggesting that a natural spillover was very much plausible. The lab-leak story has a lot of suspicious context, but virtually no evidence connecting any of it to SARS-CoV-2 itself.

My initial essay went into more detail on the background of the origins debate. This essay will occasionally revisit some of that, but in less detail.

Let’s start with the market

The Huanan Market has always been somewhat awkward for the lab-leak hypothesis to deal with. If we were only reckoning with a crowded market where many people got infected, it would not be that relevant. Early epidemics tend to cluster somewhere. Once a respiratory virus is moving through a city, it’s hardly a mystery if we find viral RNA in a place with thousands of customers and workers. But the data associated with the market is more suggestive than that.

The December 2019 cases that we know about were geographically concentrated around Huanan. This includes cases where the patients had no reported market exposure. Worobey and colleagues made the geospatial argument for market centrality in 2022. [1]

As I discussed in my initial essay, there has been a serious debate over the statistical soundness of that argument. Ascertainment bias is the main worry - that is, once investigators recognized the market as a common link, they may have preferentially found cases around it, because that’s where they chose to look. Michael Weissman has made that objection in papers and articles. Stoyan and Chiu raised a somewhat different statistical criticism, arguing that the market-centroid result had been overinterpreted. Débarre and Worobey responded to the criticism, noting that even cases without a known market link centered on Huanan. Furthermore, they argued that the market remained at or near the center even under alternative spatial summaries proposed by critics. [2] [3] [4]

There’s also the separate issue of data quality. The early case records we have come from Chinese public health authorities. It’s incomplete, and we don’t know all the specifics of how the cases were added to the dataset (or possibly excluded from it).

In 2025, Andrew Levin combined dates and locations of early cases in a Bayesian model and reported a large Bayes factor favoring a lab-origin model. [5] But neither zoonosis nor lab-leak scenarios are a single model. There are many variables - e.g., how the various vendors were affected, how quickly the virus spread into surrounding neighborhoods, how cases were ascertained, where the first infection happened, etc. A Bayesian analysis can indeed tell us which of two particular mathematical models better accounts for the data. But that’s a different proposition from calculating that one historical origin narrative is literally N times more probable than the other. Ratios of that kind depend on the particular models used in the comparison. There would be a whole range of ratios if we considered all the combinations.

Nevertheless, I think there is a narrower geographic claim to be made. Which is that the earliest recognized outbreak centers on a market where live mammals susceptible to SARS-like coronaviruses were being sold immediately beforehand. In a pre-pandemic wildlife survey, researchers documented 17 shops selling live wild animals in Wuhan. Those 17 were distributed across four markets: seven at Huanan, four each at Qiyimen and Dijiao, and two at Baishazhou. Huanan had the largest concentration of documented wildlife vendors in that survey. This doesn’t mean that Huanan was Wuhan’s busiest wildlife market. Market traffic data is limited**. What we can say is that the earliest recognized outbreak centered on the wildlife market, with the most documented wildlife vendors, rather than one of the other even busier markets in Wuhan. [1] [6]

(** Worobey et al. used 2013-2014 Sina social media check-ins as an imperfect traffic proxy and identified at least 70 Wuhan markets (and 430 locations of types associated with superspreading) with more check-ins than Huanan. But that data was outdated and does not directly reflect conditions as they stood in December 2019.)

What was actually found inside the market

In addition to individual case records, we have results from several early environmental sampling efforts at the Huanan market. The sampling rounds were not all designed the same way.

The January 1 samples. On Jan 1, investigators sampled widely, but deliberately emphasized stalls associated with human cases and stalls selling livestock, poultry, or farmed wildlife. Of 515 samples, 27 were qPCR-positive. So if we looked at a simple spatial map of positive dots, that would overstate what that round alone can show. (Later spatial analyses compared positive samples with the actual distribution of negative samples and still found increased positivity in the southwest section of the market.) [7] [8]

The January 12 samples. The sampling from Jan 12 provides a cleaner comparison regarding the wildlife trade itself. Investigators took exactly 10 samples from each of 7 wildlife stalls and sequenced all 70 samples. Wildlife stall A stood out, three of its ten samples were qPCR-positive - these included samples from a cart, a hair-or-feather-removal machine, and the ground. Metagenomic sequencing found SARS-CoV-2 reads in those three samples and in two additional PCR-negative samples from the same stall. The only other Jan 12 sample with SARS-CoV-2 sequence reads came from a freezer in nearby wildlife stall B. Because the seven stalls were sampled equally on that trip, stall A’s prominence cannot be explained by the investigators simply swabbing it more often. [8]

Later drainage samples. Later drain samples repeatedly detected SARS-CoV-2 immediately in front of or downstream from stall A. I think we have to give those samples less weight. By then, the market had been closed and disinfected for weeks. So, any drainage contamination detected in later samples cannot cleanly distinguish virus shedding from animals vs. accumulated human contamination. The Jan 12 balanced stall comparison remains the more informative result. [8]

All of the sequenced SARS-CoV-2-positive samples from that stall also contained mammalian wildlife genetic material. Raccoon dogs were the most abundant wildlife species detected. Civet and bamboo rat DNA also showed up in relevant samples. On some animal-associated surfaces, mammalian wildlife DNA content exceeded human DNA. Keep in mind that this was the stall that also had the greatest number of SARS-CoV-2-positive samples. [8]

Raccoon dogs are known to be susceptible to SARS-CoV-2. In a controlled infection experiment, 6 of 9 inoculated animals became productively infected and shed virus. 2 of 3 untreated contact animals acquired infection from the infected ones. All infected animals showed little overt illness. [9]

Now, none of this is dispositive. The animals were gone by the time the relevant sampling took place. An environmental swab measures a mixture of whatever happened to leave material on a surface. A sick person could have deposited SARS-CoV-2 RNA on the same cart that previously carried an uninfected raccoon dog. So, even if we can show co-location, that’s not the same as demonstrating infection. In fact, we have no evidence that an infected animal was present at the Huanan market.

As I noted in my last essay, Jesse Bloom has discussed this issue in detail. In that market metagenomic dataset, several animal coronaviruses quantitatively track the genetic material of the animals they are known to infect. (Meaning, we can match up DNA for those non-SARS-CoV-2 coronaviruses to the animals they are known to infect.) But the dataset also shows that, for SARS-CoV-2, we do not see an equally clean abundance relationship with susceptible mammalian hosts (raccoon dogs, civets, etc.). Bloom’s conclusion was: the metagenomics cannot tell us which market species, if any, was shedding SARS-CoV-2. [10]

It’s not hard to think of possible human-to-human transmission explanations for much of what happened at Huanan. Courtier-Orgogozo and de Ribera noted that among 53 early cases with direct market exposure, 30 were fixed-stall vendors, 12 were bulk purchasers buying for hotels or restaurants, and 2 were delivery workers. Only 9 were ordinary shoppers or nearby residents. People who spent hours in the market could readily have infected one another in shared enclosed spaces such as toilets, canteens, or activity rooms. Huanan almost certainly became an early amplification site, whether or not the first successful spillover happened there. [11]

So, a large cluster of market-linked cases and widespread environmental contamination do not, by themselves, point to a specific mode of origin. The more discriminating information is really the localization data within the market. [1] [8]

West-side layout of the Huanan market showing the live-mammal selling area and SARS-CoV-2-positive environmental sampling locations. The panel also documents raccoon dogs present at the market. Reproduced from Worobey et al., Science 377 (2022), Fig. 4, doi:10.1126/science.abp8715. The source article is distributed under CC BY 4.0.

The market was not a collection of random DNA

As mentioned above, the market samples contained more than SARS-CoV-2 plus bits of animal DNA. Researchers recovered close relatives of viruses associated with raccoon dogs, bamboo rats, and civets, sometimes with enough sequence data to reconstruct most of a genome. Those animal viruses were found where the genetic material from their expected hosts was found. This indicates that the collected environmental samples can provide insight into host-virus ecology. We aren’t just looking at a scramble of genetic confetti. [8] [10]

Bloom’s abundance result still blocks the strongest inference. The samples cannot be used to identify the SARS-CoV-2 host. The narrower point that can be established is that the wildlife section was an “active” wildlife-plus-virus environment. Susceptible mammals were there. Other viruses known to be carried by those mammals were also there. And SARS-CoV-2 contamination was concentrated in that same part of the market at the beginning of the pandemic. [10]

The population genetics of the raccoon-dog DNA samples gives us another (plausibly relevant) data point. Mitochondrial sequences recovered from Huanan were genetically distinct from those of the northern Chinese market and fur-farm populations used for comparison. The sequences are, instead, compatible with central or southern Chinese populations, closer to where the virus is thought to have originated. [8]

None of this allows us to reconstruct a chain of suppliers or to trace an infected animal back to a bat cave in Yunnan (southern China). But it does show that the Huanan raccoon dogs were compatible with populations that could have arrived through long-distance wildlife trade. So we can treat that as pointing to both geographic convergence and ecological plausibility.

Something had to move the virus lineage to Wuhan

I discussed this in detail in my prior essay, but I will outline it again for completeness.

The closest known relatives of SARS-CoV-2 come from bats far from Wuhan - in southern China and northern Laos. The very closest relatives, RaTG13 and the BANAL variants, are not direct ancestors of SARS-CoV-2. But they do tell us where much of the relevant natural sarbecovirus diversity is to be found. [12]

(A lot has been written about the nature of bat caves in Southern China, where bats from multiple species co-exist in large numbers. The ecology of those caves is fascinating in itself.)

The 2025 Cell paper by Pekar and colleagues approached the genomic analysis around this with an important complication in mind: heavy recombination in sarbecoviruses. Different “pieces” of the genome can have distinct histories, so treating the whole genome as a single, clean family tree can distort the evolutionary history across space and time. The authors split the genomes into non-recombinant regions and then reconstructed the histories of each region separately. The analysis suggests that ancestors genetically close to SARS-CoV and SARS-CoV-2 were circulating only a few years before the respective human outbreaks. And, for SARS-CoV-2, closely related ancestral regions are found in southern China and northern Laos (in horseshoe bat populations). [12]

Given that the inferred time from this analysis is just a few years, we can’t physically get from those southern populations up to Wuhan in that short time. Certainly not at any speeds ordinary Rhinolophus bat dispersal can readily explain. Most likely, “something” moved the lineage up to Wuhan. [12]

Lab-leak proponents have a perfectly good candidate for that transport: researchers from the Wuhan Institute of Virology (WIV) collected bat coronaviruses in southern China and transported the samples to Wuhan.

Popular press summaries of the 2025 Cell paper describe it as “confirming” a wildlife route. It does not. But wildlife commerce does provide a second means of transport. China had a network that moved live mammals over long distances into big-city markets. We know susceptible mammals were being moved into Huanan shortly before the outbreak. The raccoon-dog genetics from the Huanan environmental samples are at least compatible with those from animals sourced from central or southern China. There is a plausible biological chain here. [6] [12]

The evolutionary reconstruction by Pekar et al. appears to require long-distance movement of the viral lineage. We know that an independently documented animal trade capable of providing it existed in China in 2019. There is also some evidence that the Huanan market was an endpoint for this trade. WIV researchers do provide a plausible alternate transport mechanism. But, so far, we don’t have any evidence of a progenitor virus having traveled that route in a research collection.

The A/B lineage argument

I discussed the early A and B lineages at some length in the first essay. There was an influential analysis by Pekar et. al. arguing that early viral diversity was best explained by at least two separate introductions into humans. Their argument was that, if two distinct lineages really spilled over around the same market, that would fit repeated exposure to infected animals much more naturally than a single laboratory accident. [13]

This two-introduction inference isn’t really secure. Jesse Bloom’s 2025 paper argues that we cannot confidently “root” the early SARS-CoV-2 tree using the data available to us. Michael Weissman has separately challenged the Bayesian analysis behind the two-introduction result. [14] [15]

However, even if Bloom is entirely right that there was only one successful introduction and the environmental swabs cannot identify the infected animal, most of the market argument still survives. We are still left with several key evidentiary elements: the early epidemic centered on Huanan, susceptible live mammals at the market, SARS-CoV-2 concentrated in its wildlife-selling section, and animal-associated surfaces tested positive for the virus. Plus, we have the broader evolutionary history of a natural bat sarbecovirus.

The furin cleavage site

Over the years, the furin cleavage site has probably generated more suspicion and debate than any other molecular feature of SARS-CoV-2.

SARS-CoV-2 carries a polybasic insertion at the S1/S2 boundary of the spike protein that’s absent from its closest known sarbecovirus relatives. This feature appears to aid the virus's entry into human cells. [27] Lab-leak proponents also argue that SARS-CoV-2 looked suspiciously competent in infecting humans right out of the gate. (As opposed to that competency increasing with mutation in humans over the course of the early spread.)

And then the infamous DEFUSE proposal surfaced. DEFUSE showed that researchers at WIV, along with counterparts at UNC and EcoHealth, had contemplated experiments involving S1/S2 cleavage sites in sarbecoviruses before the pandemic.

Natural-origin advocates tend to dismiss all of this by saying, “furin sites occur in nature, so this isn’t a big deal.” But that dismissal seems hasty. There really is a narrower comparison to be made against just sarbecoviruses (as opposed to betacoroniviruses in general). In that comparison, the SARS-CoV-2 site is indeed unusual.

It’s not just the furin site

If we take a close look at the biology around the actual furin cleavage site in the virus, it turns out to be less “modular” than the engineered-in-a-lab story assumes.

Early versions of SARS-CoV-2 depend on a “combination” of features for effective entry into human cells. This was demonstrated by Vu and colleagues in 2022 and followed up by a paper in 2025. We have the furin site itself. But there’s also the length and composition of the surrounding loop, including an upstream QTQTN motif, as well as glycosylation in that region. If we delete the QTQTN motif while leaving the furin site, spike processing deteriorates. If we alter the glycosylation, the TMPRSS2-dependent entry changes. If we disrupt the furin motif while preserving the extended loop, replication and pathogenesis also fall! [16] [17]

So it’s not the case that a lab tech could drop that inserted sequence into that S1/S site like a sort of little four-amino-acid “on switch” to yield the highly infectious SARS-CoV-2 phenotype. We’d also need the other pieces. That’s much harder to design upfront. Besides, there’s also no mention of those elements in the DEFUSE proposal.

Genomic comparison has shown that parts of that surrounding architecture are already present in sarbecoviruses found in nature. RaTG13, for example, carries the QTQTN motif without the SARS-CoV-2 furin site. Mutation, recombination, and selection in nature can go through thousands of combinations of pre-existing features much faster than any human engineering process. Nature is a very powerful “gain-of-function” engineer.

All of this additional molecular evidence does not automatically rule out a laboratory-associated origin. What it does do is make the simplest “someone inserted a furin site and made SARS-CoV-2” thesis much less impressive.

There is also no obvious pre-emergence adaptation signal

In a lab setting, there is an alternative to purposeful and deliberate alteration of the virus genome. Experimental passage can also be an evolutionary search process. If selection during passage does some of that “searching,” researchers don’t have to predict every interaction in advance. And you don’t have to generate all of the elements in the lab. If the lab started with an undisclosed natural virus, that could have been a starting point containing much of the needed background.

Simplistic furin-site insertion scenarios seem less credible following the analysis by Vu and colleagues. But, perhaps the virus was indeed adapted through cell culture or animal passage? The 2026 Cell paper by Havens and colleagues tests that idea. They examined several previous viral emergences and compared changes in natural selection regimes. They also examined lab and gain-of-function passage datasets. [18]

For SARS-CoV-2, they found no significant shift in selection intensity on the pre-emergence stem relative to related bat viruses. A detectable change appears only once sustained human circulation begins. That’s what we might expect from adaptation following a host switch. [18]

The same framework did pick up distinctive changes in some laboratory-passage datasets. It also detected a pre-emergence shift in the 1977 H1N1 reemergence, which has long been suspected of having a lab origin. [18]

The Haven et al. test is not some sort of universal lab-leak detector. A short handling episode (as opposed to prolonged serial passage) could leave almost no trace. A naturally human-compatible virus might not even require all that much adaptation. Different passage regimes could leave different signatures.

Still, the Havens et al. result does argue against one particular intuition: SARS-CoV-2 does not appear to have spent a long period being adaptively tuned in a pre-pandemic experimental environment.

The restriction map argument

(The restriction map anomaly is not one that I discussed in my last essay. I’ll only briefly mention it here.)

Some authors have argued that the “spacing” of BsaI and BsmBI restriction sites in SARS-CoV-2 looks convenient for synthetic assembly. [23] [25]

Also, the restriction-site argument could well be a case of post-hoc feature selection. If you start with the finished genome and ask what pattern looks “designed,” it’s easy to underestimate how many alternative patterns might also have seemed interesting. A 2025 preprint from Zach Hensel argues that the BsaI/BsmBI restriction-site pattern looks much less distinctive as we add more natural sarbecovirus genomes to the comparison set. This should make us somewhat reluctant to treat the restriction map as an independent forensic signature. [23] [25] [26]

Where does this leave the lab-leak scenario?

Let’s review what we know.

We know that the Wuhan Institute of Virology ran a major bat coronavirus research program. WIV researchers carried out years of field surveillance of viruses in southern China. They sequenced several SARS-related coronaviruses and isolated live bat sarbecoviruses capable of using the human ACE2 receptor. They also transported those viruses from southern China back to their lab in Wuhan. All of this activity has been documented by the WIV researchers themselves.

The 2015 UNC-led SHC014 chimera experiments provide another datapoint. WIV scientists collaborated on that project with UNC researchers. The idea in those experiments was to insert a bat coronavirus spike into a SARS backbone. The resulting virus would be tested in human airway cells and mice. (Note: even though this was a collaborative project, it’s not clear how much of the work was actually done at WIV.) [19] [20] [21]

We also know that oversight at WIV may have been imperfect. The HHS Office of Inspector General concluded that NIH and EcoHealth did not effectively monitor awards and subawards. They also documented instances of reporting failures by EcoHealth. None of that proves a biosafety accident happened at WIV, but it does make claims like “nothing risky was going on at WIV” untenable. [22]

Finally, we have the 2018 DEFUSE proposal. DEFUSE contemplated experiments involving S1/S2 cleavage sites in sarbecoviruses. DARPA rejected the proposal and didn’t actually fund it. (We should always remember that whenever DEFUSE is discussed.) But the proposal still tells us what sorts of experiments people in this research network had been “considering” right before the pandemic. [23]

All of this information is relevant when we consider the geographic coincidence of a virus showing up in Wuhan. We have a novel bat sarbecovirus emerging in Wuhan, far from the closest known bat-virus populations, in the city that hosts the institute that was going out to collect and study those viruses. An institute that had recently considered modifying coronaviruses in a way similar to features seen in the novel sarbecovirus that started the pandemic. [12] [19] [20]

All of these facts taken together are a genuine cause for suspicion. Also, serious proponents of zoonosis shouldn’t pretend these facts are irrelevant. The facts do establish both opportunity and capability, making a lab leak more plausible than it would be in a random city without a coronavirus research program.

China has not made all the records available that would allow outsiders to properly test the lab scenario. Records like sequence databases, biosafety information, occupational health records, freezer inventories, and experimental documentation. Some of the material we would most like to see remains inaccessible to this day. [24]

The missing data on both sides

Lab-leak proponents (correctly) emphasize that the zoonotic chain is incomplete. To date, no one has found an infected raccoon dog, civet, or other intermediate animal. Nor has anyone found the immediate precursor virus, or the exact route of its entry into Huanan.

Chinese teams reported 457 animal-related samples from 188 individuals across 18 species in and around Huanan between January and March 2020, all of which were negative for SARS-CoV-2. But these samples were largely from unsold goods kept in refrigerators and freezers. Or from carcasses, animal products, or stray/vector animals sampled after the market had closed. Raccoon dogs were not among the animals tested. The susceptible live mammals documented at Huanan in late 2019 had already been removed, so they could not be tested. [7] [24]

Based on subsequent broader searches, Chinese scientists also reported 32,479 animal samples from 218 species across 18 provinces. No SARS-CoV-2 was detected. That’s genuine negative evidence against an easily detectable, widespread animal reservoir.

The WHO SAGO report notes that the Chinese reports did not provide the sampling schemes, species-level sample sizes, farm locations, or animal sources needed to assess how well they covered the relevant supply chains. A published 2022 survey of 1,941 game animals included only 95 raccoon dogs and 425 masked palm civets. And those were mostly sampled after Feb 2020. Moreover, those samples were mostly from natural habitats rather than the production sites that could have supplied Wuhan. As of SAGO’s 2025 assessment, no surveillance results from the upstream wildlife value chains and wildlife farms supplying Huanan had been published or made available to SAGO despite repeated requests. [24]

The missing intermediate host remains the largest hole in the zoonotic reconstruction. But was there an adequate search for such a host? Chinese investigators did conduct broader surveillance, but they did not conduct an exhaustive contemporaneous search of the live mammals and source farms most relevant to Huanan. (Or they did, but have not disclosed that they did) [24]

Now let’s compare the corresponding chain on the research side.

So far, no pre-outbreak virus at WIV has been publicly demonstrated to be sufficiently similar to SARS-CoV-2 to serve as a plausible progenitor. No experiment involving such a virus at WIV has been identified either. No laboratory or field-collection exposure has been shown to have infected the first person. No researcher or collector has been demonstrated to have been infected by SARS-CoV-2 prior to the outbreak. Nor do we have a clear epidemiological chain from a lab-origin scenario/event into the patterns we see at the Huanan market. [24]

On the lab leak side, there is evidence of opportunity, capability, institutional weakness, and some uncomfortable coincidences. But any affirmative evidence that the proposed lab leak event “actually” happened remains sparse. Obviously, China’s refusal to provide records is one reason we cannot rule out a virus research-related accident. But that’s not evidence that the missing records contain a SARS-CoV-2 progenitor. There may have been a relevant research pathway contemplated as part of DEFUSE, but there’s no evidence that any of that research ever took place.

An unfortunate asymmetry in the debate

The evidentiary standards required by either side in the public debate have diverged. Proponents of the zoonotic hypothesis are asked to produce the entire chain from bat reservoir to intermediate host to infected animal to supplier to market to first infected human. If any one of those links is missing, that’s treated as a major strike against the hypothesis.

For zoonosis, we have a fairly detailed scene without the culprit animal. For a research-associated origin, we have somewhat relevant research without an actual virus, accident, infected worker, or transmission chain. It’s important to recognize this asymmetry as these two stories are not incomplete in the same way.

The conjunction argument for a lab leak

A key part of the lab-leak argument is that DEFUSE, the furin cleavage site, the claimed restriction-map anomaly, etc., should be considered together rather than as unrelated points. The overall package is more suggestive than any one item alone.

But this “conjunction” still lacks any biological links. DEFUSE was only a proposal. There’s still no evidence (e.g., laboratory records) showing SARS-CoV-2 being constructed. The known viruses discussed in the DEFUSE proposal are not direct progenitors of SARS-CoV-2. No experiment has ever surfaced showing that an undisclosed SARS-CoV-2-like virus was modified according to the DEFUSE plan.

And as we noted earlier, the spike biology also matters here. The furin site needs to be considered in light of the additional structural features required to fully realize the infection potential seen in the early strains. That makes any simplistic insertion story less attractive. The restriction map argument should probably not be given the weight it has been, since that site pattern may not be as artificial as once thought. [16] [17]

Why Wuhan?

The core of the lab leak argument is straightforward. What are the odds that a pandemic involving a bat coronavirus began through natural causes in the same city as the Wuhan Institute of Virology, which was known to be heavily involved in bat coronavirus research?

But, if we look at the details, the outbreak didn’t just start “somewhere” in Wuhan. Early cases centered on the market, which hosted susceptible live mammals. And the strongest environmental signal in the market lies in its wildlife-selling areas.

The lab leak story has an obvious reason for “Why Wuhan?” WIV researchers transported the virus there. But the zoonotic story has a good reason too. Wildlife commerce transported susceptible mammals into the city, and the early outbreak appeared at one of the places where those animals were being sold.

At the end of the day, the geographic observation that the pandemic started in Wuhan isn’t dispositive for either scenario. If the first recognized outbreak had appeared in some random Wuhan office building, perhaps the WIV coincidence would indeed loom much larger.

Bayesian analysis can be misleading

The natural temptation is to turn these observations into a Bayesian calculation. In theory, that seems like a compelling formal approach. We start with prior odds and update based on the market, the wildlife sampling data, WIV, DEFUSE, the furin site, etc.

But, in practice, I don’t think we can do this numerical analysis effectively.

What might be the proper prior for a lab accident? As I mentioned in my last essay, if the reference class is “new human coronaviruses,” the base rate is strongly zoonotic. If the class is “new sarbecoviruses emerging in a city with a major sarbecovirus laboratory,” the answer looks very different. (That second reference class has approximately one member.)

The likelihoods may be even harder. What is P(Huanan pattern | zoonosis) as opposed to P(Huanan pattern | lab accident)? We don’t have empirical distributions for either quantity.

Another point I mentioned in my last essay is that, in both scenarios, parts of the evidence are somewhat interdependent. The Huanan case maps, environmental contamination, wildlife DNA, and wildlife stall localization are not four independent observations that we can multiply together. Neither is WIV’s location, its research program, DEFUSE, nor the furin cleavage site.

Which is why I picked that simpler reasoning approach in my last essay: which story would make the joint pattern less surprising, and which explanation requires auxiliary assumptions chosen mainly after seeing the data? That’s Bayesian reasoning in spirit, without incorporating probabilities that we don’t really have a handle on.

Could there be more evidence out there?

It’s been close to 7 years since the Wuhan outbreak. I no longer hold out hope that any new credible, properly provenanced evidence will surface that might further change the balance of arguments. China’s refusal to provide key records may have made a clean answer to the origins question impossible.

The WHO SAGO report stated that the weight of available evidence suggested a zoonotic spillover. At the same time, WHO did hedge by stating that a laboratory biosafety or biosecurity breach could not be ruled out because the information needed to fully evaluate it was unavailable. [24]

What might new evidence look like?

On the lab leak side, we would need something like a (contemporaneous, pre-outbreak) WIV sample or a sequence of a virus extremely close to SARS-CoV-2. Ideally, in the context of relevant experiments. Or maybe we find evidence of a biosafety event at WIV. Or documentation of COVID-like illness among WIV personnel. Any of that would supply/augment the biological bridge that the lab case currently lacks.

Equivalent evidence on the zoonosis side would be data from 2019 or earlier from a market, farm, supplier, or wildlife sample containing an immediate or near-immediate SARS-CoV-2 ancestor. It would be much better if we could also link that to the wildlife supply chain leading to Huanan.

Evidence of that kind on either side would matter far more than another round of arguments about things like one codon or restriction-site spacing “looks engineered” or not.

Hence Zoonosis

It’s clear that, on the basis of all the evidence we have to date, we can’t establish a clear case for either hypothesis. But that shouldn’t lead us to conclude that they are equally supported by that evidence.

The lab leak case relies mostly on suspicious circumstances surrounding an event for which the central biological evidence has never been produced.

The zoonotic case is incomplete, but it has more evidence linking its proposed mechanism to the observed early outbreak.

So, reviewing all the material we’ve gone over so far, it seems to me that the lab leak hypothesis has more missing causal links and fewer affirmative observations linking its proposed mechanism to the actual onset of the outbreak than the zoonosis scenario.

This is why I lean towards accepting some version of the zoonosis scenario. Not with much confidence and not because the case is closed.

References

[1] Worobey M, et al. “The Huanan Seafood Wholesale Market in Wuhan was the early epicenter of the COVID-19 pandemic.” Science 377:951-959 (2022). doi:10.1126/science.abp8715. Erratum: Science (15 March 2024), doi:10.1126/science.adp1133.

[2] Stoyan D, Chiu SN. “Statistics did not prove that the Huanan Seafood Wholesale Market was the early epicentre of the COVID-19 pandemic.” Journal of the Royal Statistical Society Series A 187(3):710-719 (2024). doi:10.1093/jrsssa/qnad139.

[3] Weissman MB. “Proximity ascertainment bias in early COVID case locations.” Journal of the Royal Statistical Society Series A 187(3):720-722 (2024). doi:10.1093/jrsssa/qnae021.

[4] Débarre F, Worobey M. “No evidence of systematic proximity ascertainment bias in early COVID-19 cases in Wuhan: Reply to Weissman (2024).” arXiv:2405.08040 (2024). https://arxiv.org/abs/2405.08040

[5] Levin AT. “A Bayesian Assessment of the Origins of COVID-19 Using Spatiotemporal and Zoonotic Data.” NBER Working Paper 33428 (2025). doi:10.3386/w33428.

[6] Xiao X, Newman C, Buesching CD, Macdonald DW, Zhou ZM. “Animal sales from Wuhan wet markets immediately prior to the COVID-19 pandemic.” Scientific Reports 11:11898 (2021). doi:10.1038/s41598-021-91470-2.

[7] Liu WJ, et al. “Surveillance of SARS-CoV-2 at the Huanan Seafood Market.” Nature 631:402-408 (2024; online 2023). doi:10.1038/s41586-023-06043-2.

[8] Crits-Christoph A, Levy JI, Pekar JE, et al. “Genetic tracing of market wildlife and viruses at the epicenter of the COVID-19 pandemic.” Cell 187(19):5468-5482.e11 (2024). doi:10.1016/j.cell.2024.08.010.

[9] Freuling CM, et al. “Susceptibility of Raccoon Dogs for Experimental SARS-CoV-2 Infection.” Emerging Infectious Diseases 26(12):2982-2985 (2020). doi:10.3201/eid2612.203733.

[10] Bloom JD. “Importance of quantifying the number of viral reads in metagenomic sequencing of environmental samples from the Huanan Seafood Market.” Virus Evolution 10(1):vead089 (2024; published online 30 December 2023). doi:10.1093/ve/vead089.

[11] Courtier-Orgogozo V, de Ribera FA. “SARS-CoV-2 infection at the Huanan seafood market.” Environmental Research 214(Pt 1):113702 (2022). doi:10.1016/j.envres.2022.113702.

[12] Pekar JE, et al. “The recency and geographical origins of the bat viruses ancestral to SARS-CoV and SARS-CoV-2.” Cell 188(12):3167-3183.e18 (2025). doi:10.1016/j.cell.2025.03.035.

[13] Pekar JE, et al. “The molecular epidemiology of multiple zoonotic origins of SARS-CoV-2.” Science 377:960-966 (2022). doi:10.1126/science.abp8337.

[14] Bloom JD. “The Data are Insufficient to Confidently Root the SARS-CoV-2 Phylogenetic Tree.” Molecular Biology and Evolution 42(6):msaf118 (2025). doi:10.1093/molbev/msaf118.

[15] Weissman MB. “An Article in Science on Covid Origins Contains a Fundamental Error.” Econ Journal Watch 23(1):46-68 (2026). https://econjwatch.org/articles/an-article-in-science-on-covid-origins-contains-a-fundamental-error

[16] Vu MN, et al. “QTQTN motif upstream of the furin-cleavage site plays a key role in SARS-CoV-2 infection and pathogenesis.” Proceedings of the National Academy of Sciences 119(32):e2205690119 (2022). doi:10.1073/pnas.2205690119.

[17] Morgan AL, et al. “The furin cleavage site is required for pathogenesis, but not transmission, of SARS-CoV-2.” Journal of Virology 99(7):e00467-25 (2025). doi:10.1128/jvi.00467-25.

[18] Havens JL, et al. “Dynamics of natural selection preceding human viral epidemics and pandemics.” Cell 189(9):2762-2775.e11 (2026). doi:10.1016/j.cell.2026.02.006.

[19] Ge XY, et al. “Isolation and characterization of a bat SARS-like coronavirus that uses the ACE2 receptor.” Nature 503:535-538 (2013). doi:10.1038/nature12711.

[20] Hu B, et al. “Discovery of a rich gene pool of bat SARS-related coronaviruses provides new insights into the origin of SARS coronavirus.” PLoS Pathogens 13(11):e1006698 (2017). doi:10.1371/journal.ppat.1006698.

[21] Menachery VD, et al. “A SARS-like cluster of circulating bat coronaviruses shows potential for human emergence.” Nature Medicine 21:1508-1513 (2015). doi:10.1038/nm.3985.

[22] U.S. Department of Health and Human Services, Office of Inspector General. “The National Institutes of Health and EcoHealth Alliance Did Not Effectively Monitor Awards and Subawards, Resulting in Missed Opportunities to Oversee Research and Other Deficiencies.” A-05-21-00025 (2023). https://oig.hhs.gov/reports/all/2023/the-national-institutes-of-health-and-ecohealth-alliance-did-not-effectively-monitor-awards-and-subawards-resulting-in-missed-opportunities-to-oversee-research-and-other-deficiencies/

[23] EcoHealth Alliance et al. “DEFUSE: Defusing the Threat of Bat-borne Coronaviruses.” Proposal to DARPA PREEMPT, March 2018; declined by DARPA. https://assets.ctfassets.net/syq3snmxclc9/4NFC6M83ewzKLf6DvAygb4/0cf477f75646e718afb332b7ac6c3cd1/defuse-proposal_watermark_Redacted.pdf

[24] WHO Scientific Advisory Group for the Origins of Novel Pathogens (SAGO). “Independent assessment of the origins of SARS-CoV-2.” Report to the WHO Director-General, 27 June 2025. https://www.who.int/publications/m/item/independent-assessment-of-the-origins-of-sars-cov-2-from-the-scientific-advisory-group-for-the-origins-of-novel-pathogens

[25] Bruttel V, Washburne A, VanDongen A. “Endonuclease fingerprint indicates a synthetic origin of SARS-CoV-2.” bioRxiv (2022). doi:10.1101/2022.10.18.512756. Preprint.

[26] Hensel Z. “On the distributions of restriction sites in human and pangolin sarbecoviruses.” arXiv:2510.23833 (2025). Preprint. https://arxiv.org/abs/2510.23833

[27] Lavie M, Dubuisson J, Belouzard S. “SARS-CoV-2 Spike Furin Cleavage Site and S2′ Basic Residues Modulate the Entry Process in a Host Cell-Dependent Manner.” Journal of Virology 96(13):e00474-22 (2022). doi:10.1128/jvi.00474-22.

There is a numerical parameter that tells us how fast our universe is expanding - the Hubble constant, H0. H0 is one of the most important numbers in all of cosmology. It sets the scale of the universe’s expansion and affects the downstream calculations we make about its characteristics - attributes such as the age of the universe, distances to distant galaxies from Earth, etc. H0 is not the only factor influencing those numbers, of course, but it is the input with the most impact.

One of the biggest challenges cosmology and astronomy have faced over the past 100 years or so has been nailing down the actual value of H0. Considering the billions of dollars of investments we have made in telescopes and other instruments, you’d think we’d be very close to a solid number. But unfortunately, that’s not the case.

We measure H0 in two different (sets of) ways today. Two paths that yield distinct consensus values for H0. The error bars on each of those two consensus values are tiny. But the values themselves don’t match. The unit we use to measure H0 in is kilometers/second/megaparsec (km/s/Mpc). In those units, one path yields a value of around 73, and the other around 67. The gap between them cannot be explained by the error bars.

This mismatch is the infamous Hubble tension. In its current form, it’s been sitting at the center of cosmology for about a decade. The solution to the tension could be a boring one - e.g., the difference is the result of subtle measurement errors. Or it could be a more exciting one, suggesting that our understanding of the universe has one or more significant gaps.

In the past decade, we’ve ruled out some of the boring options, and yet the tension remains. This state of affairs has somewhat increased scientists’ confidence that our standard cosmological models may be missing key pieces. This has made the Hubble tension problem even more interesting.

Let’s go over the state of the problem as it stands today.

1. Why is the Hubble tension such a big deal

On the face of it, that 9% or so discrepancy between the two H0 measurements of 67 and 73 might seem like it’s not much of a problem. Surely this isn’t the crisis it’s often made out to be in the press, the literature, and at conferences?

The problem arises because the two measurements come from opposite ends of cosmic history. What connects them is our standard model of cosmology - called Lambda-Cold Dark Matter, or LCDM.

On the one hand, we have the early universe approach. If we start from the Big Bang, 13.8 billion years ago, and apply the LCDM model to fast-forward through the universe’s evolution, we can arrive at the current expansion rate. This gives us the 67 value.

On the other hand, we can try to measure the expansion rate directly today through astronomical observations. That gives us the 73 value.

If our current model of cosmology, LCDM, has gaps, meaning if our equations are missing elements that result in the Hubble tension, it’s unlikely to just be a small adjustment. The solution might well be the biggest advancement in theoretical cosmology in decades. But if there is a gap, why does LCDM still match our experimental and observational data so precisely? Where has this otherwise brilliant formulation sprung a leak?

Some History

The argument over H0 and its “real” value goes back nearly a hundred years. Edwin Hubble’s original measurement of H0 was off roughly by a factor of 7, largely due to miscalibrated distances. Later, for much of the 20th century, the verdict was split between two camps that differed by a factor of 2 in their measurements. The de Vaucouleurs group argued for around 100, the Sandage group for around 50. The disagreement between the two groups was, apparently, bitter and personal.

The hope, by the end of the 20th century, was that the forthcoming Hubble Space Telescope (HST) would resolve the issue once and for all. In 2001, a team led by Wendy Freedman used HST to determine a value for H0 of around 72. This was hailed as finally settling the question of H0’s value.

The point of this little rundown of history is perspective. The value of H0 has been controversial for nearly a century. The error bars have shrunk dramatically. Today’s discrepancy of 67 vs. 73 is a precision-era descendant of a long, contentious measurement discussion.

2. The Two Main Approaches

I’m going to call the two main approaches for calculating H0 the early and late routes since they measure the universe starting from two very different epochs.

The Early Route: Using the Cosmic Microwave Background

The cosmic microwave background (CMB) is the leftover glow from when the universe was about 380,000 years old, when matter cooled enough for neutral atoms to form and light to travel freely. Think of the CMB as the oldest light there is. The Planck satellite has mapped the CMB in exquisite detail. The pattern of hot and cold spots in that map encodes an enormous amount of information about the very early universe. E.g, what were its contents in terms of ordinary and dark matter and radiation, how big were the sound waves that ripped through the primordial plasma as a result of acoustic oscillations, etc.

The CMB doesn’t directly give us a value for H0. What it does give us are measurements of the content and geometry of the early universe. We then feed those values into the LCDM model and let the model tell us what today’s expansion rate “should be.” The answer we get from Planck data is about 67.4 (with an uncertainty under 1%). Lest anyone think the Planck data might be problematic, CMB-based measurements have also been made independently using the Atacama Cosmology Telescope and the South Pole Telescope. Those efforts resulted in very similar values of 66.86 to 68.6, fairly close to Planck’s. The early route is thus very “internally consistent”.

The strength of this approach is precision- a direct result of the physics involved being quite tractable. Density ripples in the early universe were tiny, so they can be modeled using a linear perturbation approach tightly constrained by theory. This does not mean that the calculations required to arrive at H0 with this early route method are simple. But we understand the underlying physics well.

Of course, stepping back a bit, there is always the question - is that really true? Do we “really” know about and understand “all” of the underlying physics? The CMB-based value for H0 is an “inference.” We measure the early universe and then run it through LCDM to get today’s H0. If there are gaps in our understanding of LCDM, then the CMB-based H0 will be off, even if the CMB map itself is flawless.

There are other early-universe approaches, such as using what we know from baryon acoustic oscillations (BAO) and Big Bang Nucleosynthesis (BBN). These results yield values closer to those from the CMB route, and do not rely on Planck’s CMB measurements or the late route.

The Late Route: The Cosmic Distance Ladder

One of the greatest achievements in the area of measurement in modern physics is the cosmic distance ladder. Honed over decades of modeling, measurement, and calculations, the cosmic distance ladder is our best attempt to nail down cosmological-scale distances in the universe.

(Here’s a fun lecture from the great Terry Tao on the cosmic distance ladder.)

The reason the ladder exists is a problem that has been known for a long time - we cannot measure the distance to a faraway galaxy in one step. So you have to build up a ladder of measurement reference points from an initial set of anchors, going up in distance rung by rung.

Here’s a brief overview of how the ladder works. The following outlines the ladder approach that was used by the SH0ES team. This team, led by Adam Riess (Riess shares a Nobel Prize in physics for the discovery of cosmic acceleration), calculated the highly regarded late-universe measurement of H0. That value is 73.49 +/-0.93 as of 2025.

Rung one (the anchors)

We start with a small number of galaxies whose distances can be reliably determined largely using purely geometric methods. (SH0ES uses 4 galaxies).

• The Milky Way: Gaia and Hubble satellite measurements of parallaxes to individual Cepheid variable stars, star by star, give us direct distance measurements in the Milky Way. Parallax involves triangulating against the Earth’s orbit. (The reason for targeting Cepheids, a particular type of star, is explained in Rung two.)

• The Large Magellanic Cloud: Late-type detached eclipsing binary (DEB) stars follow a well-understood “surface brightness to color relation.” The radii of the stars can be calculated geometrically, and the true luminosity assessment based on star color lets us calculate distance.

• The Small Magellanic Cloud: A 2024 update added the SMC alongside the LMC to the anchor galaxy set.

• NGC 4258: This is a galaxy with water masers that gives us angular and physical sizes (via orbital dynamics), which can be used to calculate distance.

The distances calculated from these galaxies affect all the subsequent ladders on the rung. They generally align with each other, but any errors here propagate upwards and affect all the other rungs.

Rung two (the calibrators)

As mentioned in the prior section, Cepheid variables are a type of star that have become invaluable in astronomy. These stars pulsate with a well-defined period that can be tightly linked to their intrinsic brightness. So we can measure the period → we know the true luminosity → compare that to how bright it looks → calculate the distance.

Once we have Cepheids in the anchor galaxies calibrated, we can then look for Cepheids in a separate, special set of galaxies. This second galaxy set has to be close enough to resolve individual Cepheids (within roughly 80 Mpc) but, crucially, also has recently hosted a Type 1a supernova.

A Type 1a supernova occurs in binary star systems where one star is a white dwarf. These supernovae produce consistent peak luminosities because white dwarfs explode at a fixed critical mass, which can be used to calculate distance from the observed brightness.

(The actual situation with Type 1a supernovae is somewhat more complicated than that. We do have characteristic light curves, but there are multiple progenitor routes. Empirical standardization is key to modeling.)

In the SH0ES dataset, there are 42 supernovae inventoried across 37 host galaxies, with each galaxy providing both a Cepheid distance and a well-observed supernova.

The Cepheids have thus handed over the distance scale baton to the supernovae, whose moment to shine comes in Rung three.

Rung three (going the distance)

Type 1a supernovae are often referred to as “standard candles” and can be used much further than Cepheids.

The moniker "standard candles" might make us think these supernovae all explode with the same brightness. They don't. Their true luminosities span a range. What makes them work is that they are “standardizable” - the brighter ones fade more slowly, the dimmer ones fade faster. There is a tight relationship, known as the Phillips relationship, between peak luminosity and the shape of the light curve. Once we account for this variance, we can use the fact that the supernovae are very bright to measure distances out to hundreds of Mpc.

These calibrated supernovae are in galaxies whose motion relative to us is dominated by cosmic expansion. This lets us compare how fast those galaxies recede against how far they are away. That final comparison gives us a value for H0.

Strengths and weaknesses of the Late Route approach

The main strength of the late route approach is that it’s a local measurement in the present, and so what happened between recombination and now is irrelevant. If there were exotic physics in the early universe, it would not affect the approach’s methodology. What matters is standard cosmological geometry and the astrophysics related to DEBs, Cepheids, and Type 1a supernovae.

Another strength is that the late-route approach uses a completely different set of assumptions than the early-route approach.

The main weakness should be obvious. All of the rungs have their own possible systematic errors. The further up the ladder you go, the more such errors compound.

In the next section, let’s take a close look at some of the ways in which the rungs could be problematic.

3. The Systematics: Where a Distance Ladder Can Go Wrong

The problems involved in building an accurate distance ladder are largely the fairly non-exotic hazards of doing astrophysics with light that has traveled long distances through a dusty, crowded universe.

Dust (reddening). Interstellar dust dims and reddens starlight. If we don’t correct for it perfectly, then the star appears fainter than it is. Meaning, we think it’s further away than it really is. There’s no simple arithmetic formula that corrects all of these errors. Different distance indicators are reddened by different amounts. The corrections depend on assumptions about the dust’s properties along the line of sight.

Crowding (blending). Distant galaxies are packed with stars. Galaxies can have hundreds of billions of stars. When we point a telescope at one star, some of the light in the measurement may actually come from nearby unrelated stars, blended into the same pixels. This works the reverse of reddening. That extra light makes the star look brighter and, hence, closer than it really is. The errors also get worse the farther out we look.

Metallicity. The chemical composition of a star affects its brightness and color. (Astronomers like to call anything heavier than helium a metal.) If a Cepheid reference has a different metal content than the others we compare it to, we need to correct for the difference, or we get a systematic error.

Supernovae progenitors. Type 1a supernovae originate from binaries involving a white dwarf. But what if their intrinsic brightness subtly depends on the age or the chemistry of the stars that produced them?

Selection effects. What if the sample we used to calibrate Type 1a supernovae locally in Rung two is not truly representative of the samples from Rung three? Discussions continue on how to address these sorts of issues.

The problem with all of these systematic effects is that it was hard to rule out whether some combination of them was causing an H0 measurement to deviate from the CMB-based approach. The hope was - perhaps once they were all corrected for, the tension would largely vanish as well. At least that was the hope, until JWST arrived on the scene and made that position harder to hold.

4. Enter JWST

The James Webb Space Telescope has a mirror about 2.7 times the diameter of Hubble’s. It can see deep into the infrared, where dust is far less of a problem. The bigger mirror can resolve finer details, the infrared reach lets it address the reddening of Cepheids, and so on.

Two different teams of scientists used JWST to gather data from the same galaxies. However, they came away with different conclusions.

The SH0ES / Riess reading

Adam Riess’s team used JWST to re-observe their Cepheids at higher resolution. The logic was that if crowding were indeed inflating Hubble's brightness measurements, then JWST would make that obvious and pull the number down. That didn’t happen. The JWST Cepheid distances came out in close agreement with the Hubble ones. The SH0ES team sees that result as strong evidence that crowding/blending is not the primary culprit behind the Hubble tension. They also (cross) checked Cepheids against other indicators (red giants, for example) calibrated to the same NGC 4258 (masers) anchor and found close agreement once more.

Net result: The SH0ES team’s position is that the systematics check out. And so, the Hubble tension is not primarily due to measurement errors and is real.

The CCHP / Freedman reading

Wendy Freedman (yes, the one who led the 2001 Hubble Space Telescope-based measurement effort that resulted in an H0 of around 72) leads the Chicago-Carnegie Hubble Program (CCHP). For some time, the CCHP group has argued that the distance ladder should not rely so heavily on Cepheids. Cepheids are, after all, young stars in dusty, crowded disks of galaxies, which are the source of a lot of systematic errors. As a result, CCHP decided to build a program around three independent indicators, cross-checking each against the others using the same JWST data.

• The Tip of the Red Giant Branch (TRGB): The Red Giant Branch is the red-giant region of the Hertzsprung–Russell diagram (and not a specific galaxy or constellation, even though it sounds like one). It’s a stage of stellar evolution in stars whose masses are similar to the Sun's. The RGB is preceded by the Main Sequence stage and immediately followed by the Horizontal Branch stage. When a low mass star is in the RGB phase, its inert helium core grows and heats until it hits the temperature where the helium abruptly ignites - the helium flash. That’s the most luminous the star ever gets on the RGB branch. The flash essentially re-establishes the star and drops it onto the fainter horizontal branch. So across a population of these stars, there’s a sharp upper cutoff in brightness, and the Tip of the RGB method involves measuring that “clean edge” of RGB evolution prior to the flash. Like all the other astrophysical things we’ve discussed above, the cutoff isn’t (literally) identical across the board - it varies a bit based on a star’s color and metal content. But that variance is modest and can be calibrated for. Additionally, the stars selected for measurement live in the dust-free outskirts of galaxies, away from the reddening and crowding that affect Cepheids.

• J-band Asymptotic Giant Branch (JAGB): JAGB refers to a type of carbon star that can be used as a relatively clean indicator. This relies on the stable, uniform near-infrared luminosities of these carbon-rich asymptotic giant branch stars in the J band.

• Cepheids, for comparison.

The CCHP numbers are more complicated to cite or explain than the SH0ES’ nice single figure. In 2024, using the TRGB method combining Hubble and JWST supernova calibrators, they announced an H0 value of about 70.4. Using only JWST data, they reported 68.8 from TRGB and 67.8 from JAGB. Their full three-method combo (TRGB + JAGB + Cepheids) is around 70.0.

Net result: irrespective of which CCHP method we use, their H0 values are consistently below SH0ES (~73). CCHP argues that their measurement is more consistent with the CMB-based measurement than SH0ES’ result.

The SH0ES vs. CCHP Discrepancy

We have two serious, competent teams, using the same amazing telescope, arriving at numbers that don’t agree. So the Hubble tension is now more than just a 67 vs. 73 debate. It’s now a 67 (CMB/LCDM) vs. 70 (CCHP) vs. 73 (SH0ES) discussion. (As someone said in a very different context: “Who ordered that?!”)

This isn’t a case of one team being sloppy in their work. The choice of method, stars, anchors, supernovae population, etc., affects the result. There is a certain amount of subjectivity in both the SH0ES and CCHP approaches.

But the supporting evidence for CCHP’s 70 vs. SH0ES’ 73 is not symmetric. If you line up the local distance indicators, most cluster around an H0 value of 73. This includes independent analyses of the very same TRBG data that CCHP uses! The core CCHP results cannot be dismissed, but as things stand, it's a conspicuously low outlier among all of the late-route/local measurements.

On the other hand, the fact that many local indicators agree on 73 should not be interpreted as “many statistically independent measurements agree on 73.” It’s true that the convergence of different distance indicators - Cepheids, the TRBG, JAGB stars, and the rest appears to be pointing to something. But they are all dependent on a shared local-distance infrastructure - the same geometric anchors, overlapping supernova samples, some common calibration choices, etc. That shared dependency could theoretically be one of the reasons that their values are being “tugged” in a particular direction (towards 73).

There are a couple of truly independent measures with separate inference chains - strong lensing and gravitational-wave sirens. But they don’t do a good job of distinguishing the 67 from the 73. I cover this in more detail below.

In any case, the complaints about Cepheid crowding seem to have died down somewhat, mostly because replacing Cepheids with other options (TRGB, JAGB) yields higher H0 values than lower ones.

5. The Miscellaneous Issues

There are ways to measure H0 in the late universe that do not rely on the distance ladder at all. These could act as independent tie-breakers. The problem is that these methods aren’t precise enough yet to be usable that way.

Strong gravitational lensing (time delays)

When a distant quasar sits behind a massive foreground galaxy, the light from the quasar gets bent into multiple images. This is because the light rays travel along different paths, so they arrive at slightly different times. That time delay depends on “absolutely distances’ in the universe, and hence on H0. No distance ladder is involved. This observational effort has been pursued by the TDCOSMO collaboration, which reports an H0 value of around 72 (+4 or -3.7).

But there’s a catch. It’s hard to separate how the lensing galaxy’s mass is distributed from the cosmological signal we are trying to observe. This problem, called the mass-sheet transformation, results in imprecision, leading to larger error bars on any measurement. As a result, right now, lensing is not precise enough to adjudicate between Planck’s 67 and SH0ES’s 73, even though the central value is closer to 73 than 67.

(It’s possible to tighten those numbers, but that would require making some strong assumptions in the model.)

Standard sirens (gravitational waves)

To the dismay or delight of astronomers, these are not the seductive sirens that Odysseus struggled with. When two neutron stars spiral into each other and merge, the gravitational waves they emit “encode the absolute distance” to the source. We can determine distances to remote galaxies, largely from “first principles”. No elaborate calibration chain or distance ladder needed here either. This is an amazing tool - probably the cleanest in the toolbox.

But sirens are in short supply. We’ve had one truly golden event, GW170817 in 2017, that came with a visible electromagnetic counterpart (to the gravitational wave) that helped pin down its host galaxy and redshift. We haven’t had another one like it since.

“Dark” sirens, which are more common, don’t have an optical/electromagnetic counterpart, just the gravitational wave. So these require statistical tricks to extract H0.

Currently, similar to lensing, siren-based numbers have large error bars. The good news is that every observing run adds events, slowly building up precision.

Other indicators

There are a bunch of other options besides lensing and sirens: surface brightness fluctuations, Tully-Fisher, Type II supernovae (modeled from their spectrum), cosmic chronometers, etc.

But with most of these options, we see a pattern. When you swap Type 1a supernovae in the distance ladder with these alternative indicators, the H0 generally comes out “high,” near the SH0ES’s 73, not Planck’s 67. This is sometimes seen as a sign that the Hubble tension is real and that better late-universe measurements won’t make the problem go away.

A different explanation: maybe it’s us

There’s an idea that doesn’t fit the tidy “measurement error vs new physics” dichotomy in most of the debate around the Hubble tension.

What if the local number is high not because of systematics in the measurement per se, but because of “where we Milky-Way-dwelling humans happen to be in the universe”? Meaning, what if the Milky Way happens to live inside a large, mildly underdense region of the universe, a cosmic “void”?

(The void in question is usually referred to as the KBC void or the Local Hole.)

If this is the case, then the extra matter surrounding that region would gently pull nearby galaxies outward. That effect could appear to us as a faster local expansion. What we thought were high measurements of H0 would then be due to a local issue in our part of the universe.

Evidence for such a void comes mainly from galaxy counts in the near-infrared. The evidence, as it stands, appears to show that our neighborhood is underdense out to a few hundred million light-years.

Nevertheless, despite some excellent papers, this idea has not gained much traction in the field.

One reason is that any such void deep and wide enough to fake a higher H0 would be in serious tension with LCDM itself. LCDM predicts a certain amount of large-scale smoothness. Any underdensity that large would be a significant anomaly. We’d trade one anomaly (Hubble tension) for another (non-smoothness on large scales).

Another reason is that any such void would naturally predict a particular pattern in the variance of the expansion rate. The inferred H0 would fall off as we look past the edge of the void to higher redshift. But when we look at the data (e.g., the supernova Hubble diagram), that sort of falloff doesn’t seem to show up.

Finally, a recent analysis using direct distance measurements to map the local velocity field found that if there were such a local void, its dimensions would be much smaller than required. About a tenth of what it would need to be.

6. Why This is a Hard Problem

The CMB and SH0ES numbers have gotten more precise over time, but have not budged from their respective anchor values of approximately 67 and 73.

JWST was supposed to help sort out the issues with the ladder. But, if we take a look at what we’ve done with JWST, it’s only really addressed a few of the ladder’s weaknesses - mostly Cepheid-related ones. And that result was mostly a vindication of the Hubble Space Telescope’s original measurements.

There are other possible issues with the ladder, as we discussed earlier. There’s the zero-point calibration, the geometric anchors and the mild tension between them, the metallicity corrections, plus the concern that our Type 1a supernovae sample used for calibration is not representative of supernovae far away. JWST has done nothing to address those. At best, we can say our thorniest systematic (Cepheid crowding) has been largely addressed.

So the systematic issue across all the local measurements remains open, especially the concern that there could be a “single correlated error in the infrastructure shared by all the indicators (Cepheids, TRGB, JAGB, etc.)”.

It may be possible that the independent tie-breakers (lensing and sirens) could be used someday to adjudicate on local measurements. But they are not precise enough to be useful today.

Fixing the theory is hard

If we assume that the local measurements are right, then perhaps we just need to fix the theory on the CMB side? Well, this is easier said than done.

Most efforts to fix the theory sort out theoretical issues related to the early route value involve the “sound horizon.” The sound horizon is the characteristic scale imprinted by sound waves on the primordial plasma. This sets the ruler used by the CMB and other early probes. If you like to better understand how these sound waves came about in the early universe and the nature of that “imprinting,” you can read section 3 of my essay on dark matter.

Shrinking the sound horizon

Before recombination, the universe was a hot plasma. Pressure waves rippled through it at relativistic speeds. The sound horizon is just how far one of those waves could travel before recombination. That is, before the plasma cooled and “froze” the pattern in place

The frozen scale is a ruler of known physical length, and the CMB lets us measure the “angle” it subtends in the sky. An angle is a length divided by a distance. If we know the angle and the ruler’s physical distance, we can solve for how far away it sits. That distance is what helps us determine H0, and the CMB measurement of it is very precise.

Now, let’s consider the possibility that the true ruler is a “little shorter” than what the LCDM assumes. To subtend the same precisely measured angle, a shorter ruler has to sit “closer” to us, resulting in a closer last-scattering surface. This, in turn, implies that the universe “expanded more” to bring it to that shorter distance in the available time. This also results in a higher H0.

So shrinking the sound horizon makes the early-universe ruler smaller, and the expansion rate implied by the CMB goes up. Most theoretical changes to the early universe that aim to fix the Hubble tension use this approach. They just achieve the shrinking in different ways.

Some possible theoretical fixes

Several theoretical fixes have been proposed to shrink the sound horizon. Let’s go over a few of them.

Early dark energy. (EDE) This option might be the front-runner. Let’s say there exists a scalar field that acts like a brief pulse of dark energy around the time of matter-radiation equality and then dilutes away. That would boost the expansion right before recombination, and the sound horizon freezes in at a smaller scale. This approach comes closest to reaching 73 of any single fix, but it needs the scalar field to be suspiciously fine-tuned to make things work out right. The pulse would have to switch on and off at just the right moment with just the right strength.

Extra relativistic species. This approach involves adding some sort of “dark” component such as a sterile neutrino or dark radiation, or an interacting relativistic sector. This would also speed up the early expansion and shrink the horizon. It’s economical and independently motivated by particle physics considerations. I discuss some of this in more detail in my dark matter essay. But, on its own, this approach cannot push H0 far enough without distorting other features of the CMB.

Varying electron mass. In the early universe, if the mass of the electron had been slightly heavier, hydrogen would have formed earlier. Recombination would occur earlier, thereby shifting the sound horizon. But… this is a theoretically “expensive” type of proposition. We’re talking about modifying the electron's mass! I can imagine outcries from even the general public, similar to what we had with Pluto being demoted. Anyway, analysis suggests that this idea is somewhat constrained by the physics of early light-element formation.

Primordial magnetic fields. It’s been hypothesized that the early universe could contain primordial magnetic fields, perhaps from cosmic inflation or early phase transitions. Such fields would make all the hot plasma in the early universe “clumpier”. As a result, recombination would complete faster on average, shrinking the sound horizon. Separately, such magnetic fields might also possibly help explain the magnetism that we see in distant galaxy clusters today. However, the relevant plasma physics is difficult to model, so the magnitude of the effect is uncertain.

Problem with these theoretical fixes

The problem with all of these fixes/approaches is that shrinking the sound horizon only gets us so far. There is a set of “no-go” arguments that show that any model that “only” relies on reducing the sound horizon runs into a wall. To raise H0 sufficiently, we’d have to adjust the matter density. If we ended up with a lower value for density, that would then spoil the data fit to the baryon acoustic oscillation (BAO) measurements from galaxy surveys (independent estimates of how much matter there is). Higher values, on the other hand, would worsen agreement with weak lensing and structure-growth data.

And let’s say we tried to fix this by making changes in the expansion rate in the late universe, instead of trying to fix things in the early universe. If we tried to change dark energy’s behavior at low redshift, that data is also constrained by those same galaxy surveys, as well as the data we have from supernovae. It also doesn’t move H0 far enough.

Sound-horizon-free measurements. Recent results have further constrained the sound horizon. If the real problem was the sound horizon size in our CMB-based calculations for H0, then any method of measuring H0 that doesn’t rely on the sound horizon should land close to the higher value of 73. In 2026, there were several approaches that tried to do these “sound horizon-free measurements.” The result? They come closer to Planck’s 67 than to 73. The error bars are large, so this is suggestive rather than settled. But if these results hold up, maybe shrinking the sound horizon isn’t going to work after all.

In any case, the shrinking of the sound horizon debate is not settled yet.

7. What Would a Real Answer Even Look Like

Let’s say we do solve this problem. What might the “shape“ of that solution look like?

Here are a few options. These are not all mutually exclusive.

Systematics in the distance ladder. As we noted, there’s still room for possible errors in the distance ladder. If someone finds that error in the supernova calibration, or even in the Cepheids, maybe the local number will then drift down to 67. (Remember, this was the possibility JWST was supposed to confirm and didn’t.)

Systematic in the CMB. As I mentioned earlier, we have three separate measurements of the CMB-based number: Planck and the two ground-based ones. These aren’t totally independent measurements - there are some shared foreground models and assumptions. But potential fixes seem scarcer on the CMB side.

New early-universe physics. This would likely involve successfully shrinking the sound horizon without breaking anything else. If we could do this, the ruler would be changed everywhere. That should also leave the late-universe distance measurements internally consistent.

New late-universe physics. This could be a correction to dark energy’s behavior or perhaps corrections to certain characteristics of supernova modeling. A dark energy fix would change the expansion history in a specific, redshift-dependent way. We should be able to test that, especially with supernova data.

The Void: This was the local environment issue we looked at earlier. The KBC void or the Local Hole the Milky Way is in. The inferred value of H0 would fall off if we look past the void’s edge to regions of higher redshift.

Changing Dark Energy?

In the last couple of years, results from the Dark Energy Spectroscopic Instrument (DESI) suggest a possible change to how we understand dark energy today. We understand dark energy today as the cosmological constant, the L in our current LCDM model of the universe. This is the energy of empty space that accelerates the universe's expansion. Its equation-of-state parameter, called “w”, is fixed at -1 for all time. Also, the density of dark energy does not change.

DESI measures the positions of a large number of galaxies. It also uses the BAO as a ruler to trace how the universe has expanded over time. As of the second data release from DESI in 2025, it had looked at the redshifts of around 14 million galaxies.

DESI’s data, especially when “combined” with CMB and supernova data, seems to prefer a w that isn’t quite -1. The data also seems to suggest that w might be changing with time. The exact result varies depending on which supernova dataset is used in the combo analysis, but the preference for this “dynamical dark energy” over a vanilla cosmological constant seems to be around 3-4 sigma. So, not yet a discovery, but strong enough to be addressed one way or another.

So what does that mean for the Hubble tension? Well, if dark energy is changing, perhaps we could use that to model a fix for H0. But when we do the math, it appears that the inferred H0 remains low, near the early-universe value.

So if the DESI result holds up, that’ll be a second piece of pressure on the standard LCDM model, in addition to the Hubble tension. Personally, I think the DESI result is just a red herring. Once we get better results and sort out things like the discrepancies between which supernova dataset DESI uses, it’s quite probable that we’ll go back to a w = -1 situation and constant dark energy.

The other cosmological tension: S8

As I discussed in my essay on dark matter, there is another similar tension in cosmology called S8. This is a disagreement about how “clumpy” the universe is. The early universe (as inferred from the CMB) predicts a single level of clumpiness. Several late-universe weak lensing surveys have historically found the universe slightly less clumpy than that.

As our accumulated data improves, the S8 tension also seems to be improving. The 2025 “Legacy” analysis from the KiDS survey has shifted upwards to agree with the CMB-based results. The e-ROSITA X-ray survey data also seem to point in the same direction. On the other hand, DES Year 6 continues to find some residual tension.

So, unlike the Hubble tension, the S8 seems to be dissolving.

8. What the next few years might look like

There’s a lot of new data that is arriving. Some of this may well move the needle on the Hubble tension.

Combined indicator local measurement. Work by a large collaboration, the H0 Distance Network, combined a variety of distance indicators (parallaxes, eclipsing binaries, masers, Cepheids, TRGB, Miras, JAGB stars, supernovae, etc.) into a cross-linked network using covariance weighting. The idea was to come up with a local measure that wouldn’t be affected by removing any of those indicators. This approach yielded an H0 value of 73.5 +/- 0.81, which is broadly consistent with SH0ES. Because of the way the indicators were weighted and combined, the collaboration reported that dropping Cepheids or TRGB, for example, barely affects the result. Swapping supernovae for galaxy-based indicators changes the value by a small amount and increases error bars somewhat, but the core measurement seems stable.

Based on the collaboration’s testing of potential fixes on the theory side, Early Dark Energy (EDE) seems the most promising so far, with extra-radiation and late-time models faring badly. Note that these results do not favor any of these fixes over LCDM at this stage of the project.

Updates from DESI. DESI continues to gather data. Future data releases will hopefully shed light on the possibility of dynamical dark energy and the value of w.

Euclid. The European Space Agency's Euclid space telescope launched in 2023. It’s surveying billions of galaxies and beginning to release data that will tighten the picture of the late universe from different angles. This is likely to help with the S8 tension and shed more light on dark energy.

Simons Observatory. This is a new CMB-targeted facility in Chile. It’ll add high-precision early-universe maps that will help further pin down the sound horizon.

Vera Rubin Observatory. Rubin’s LSST survey will significantly increase our dataset of supernovae and lensed systems. That will provide a wealth of data to help reduce systematic errors that might be plaguing supernova measurements.

Nancy Grace Roman Space Telescope. Roman, NASA’s new flagship, will help refine the distance ladder with a larger, more uniform sample of Cepheids and supernovae. And, it’ll do this with a single instrument, reducing cross-calibration pain points in the current ladder framework.

Gravitational wave data. This is more speculative, as we have to wait for new events. LIGO-Virgo-KAGRA will continue to record more merger events. And if we get lucky and observe more neutron-star mergers with a clear electromagnetic counterpart, those events could eventually deliver a clean, ladder-free H0.

Sirens. Remember when we talked about sultry yet imprecise sirens. If we find a couple of lucky bright sirens that come in at 73, that would indicate that the local measurement value is not as high merely due to measurement error. If it comes in at 67, well then, that would be good for the CMB and LCDM, but it would not tell us where we went wrong with the distance ladder.

So where does that leave us?

We still have a 67 vs. 73 discrepancy in the H0 measurements, and so the tension continues. We don’t precisely know where the issue arises or how to fix it. The problem could arise from a correlated calibration issue or from incomplete physics.

Historically, cosmological tensions have eased or even gone away as we get new data and updated models. The Hubble tension seems persistent and recalcitrant, refusing to give way to sustained theoretical and experimental efforts. Maybe we’ll catch a break in the next few years. As an optimist about these sorts of things, my viewpoint is - surely, something has got to give.

The volume of the entire Earth contains, at any given moment, about one kilogram of a certain substance we refer to as “dark matter” (DM). The local density of this substance in our neighborhood works out to about one proton mass per three cubic centimeters. This is an inconsequential and negligible amount as far as life on earth is concerned. Over our whole lifetime, only a fraction of a milligram’s worth of DM will ever stream through our body. All of that will pass through as if though us weren’t even there. Which, from the point of view of DM, we basically aren’t.

This substance is about 85% of all the matter in the universe. DM is a vital component of the current LCDM model of the universe. And yet, we have never detected a single particle of it.

So that’s the strange place DM research sits in 2026. The case for DM based on a study of the universe is strong, but the thing itself remains unidentified. Critics of the idea of DM sometimes treat the many decades of “null detection results” as if “the whole idea is wrong and everyone’s being stubborn.” But that criticism is confusing two different questions: whether the phenomenon (of missing mass) itself is real, and whether we are making progress in identifying its microscopic source(s).

First of all, the very shape of what we’ve been able to exclude via our DM search efforts is quite informative. For example, we may have failed to find WIMPs (a possible type of DM particle) in our searches. But those search failures happened in specific, well-characterized regions of the space of possibilities. Those failures increasingly “constrain” what DM can “still” be.

There’s a second point worth keeping in mind before we go any further. The question is broader than “what is the DM particle?” There is no good reason to expect there to be just one particle, as opposed to a family of particles, fields, or other effects. This is an important distinction - the evidence can tell us that some additional gravitating component is there without yet telling us what that thing is microscopically.

Why should we care? Well, without something like DM, we would not exist. As in, you and me for starters, but also galaxies, stars, and planets in general. The reason is structure formation. When we do the math on how the early universe grew into the lumpy universe around us, ordinary matter alone does not get the job done.

There is an important milestone in the history of our universe that we call “recombination.” This was about 380,000 years after the Big Bang. Protons, neutrons, and light nuclei had existed long before this. What could not form stably were neutral atoms: the universe was still a hot ionized plasma, and energetic photons kept knocking electrons free. Recombination was when electrons could finally remain bound to nuclei, and the baryonic matter then decoupled from the radiation field.

(Technical note: The baryon-only formation problem is actually worse than what I stated above: photon diffusion, called Silk damping, actively “erased’ baryon fluctuations on small scales before recombination, so a baryon-only universe doesn’t just grow too slowly, it starts with its small-scale seeds wiped out.)

DM, on the other hand, works differently. Unlike ordinary baryonic matter, it doesn’t couple to photons. Hence, it was free to start “clumping” together early on. By recombination, dark matter had already dug a scaffolding of gravitational wells. Once neutral atoms formed and baryons decoupled from the photons, ordinary matter could fall more freely into those wells, cool by radiating energy away (which, by the way, DM cannot do), collapse into disks, fragment into stars, and eventually make heavier elements like carbon and oxygen.

The idea of DM was originally motivated by the observation that visible matter in the universe (e.g., stars, gas, and dust) does not contain nearly enough mass or gravitational pull to account for the high speeds at which stars move within galaxies or how galaxies cluster together. Our models of cold dark matter (CDM) also suggest that around 80-90% of it exists in DM halos around galaxies. Halos are invisible regions of matter that surround a galaxy. DM halos extend far beyond the visible edges of stars and gas, and provide the extra gravitational mass needed to keep spinning galaxies from flying apart.

In the rest of this essay, I’m going to work through where the DM research program actually stands: why DM is so hard to detect, what we already know it isn’t, what the evidence for it looks like, what kinds of things it could still be, and which candidates remain interesting. I’ll also look at modified-gravity alternatives like MOND, some of the newest data and current problems, and the experiments that might move things forward. At the end, I’ll speculate a bit about where this all goes, and I’ll flag that speculation when I get there.

Some important/relevant terminology

Baryon: A baryon is a subatomic particle, such as a proton or neutron, made of three quarks.

Cold Dark Matter: CDM is dark matter that was moving slowly enough in the early universe to have a very small free-streaming scale, allowing structure to form from the bottom up. In our baseline models, CDM is usually approximated as collisionless, but “cold” does not imply that it interacts only through gravity.

Cosmic microwave background: CMB is the oldest light in the universe, representing leftover thermal microwave radiation from the Big Bang that fills all of space.

Galaxy rotation speed: The rate at which stars, gas, and other matter orbit around the center of a galaxy.

LCDM: Lambda CDM is the standard model of cosmology that describes the universe using Lambda (the dark energy/cosmological constant) and CDM.

MACHO: Massive Compact Halo Object. A heavy, dark astronomical body (e.g., black hole, neutron star, or brown dwarf) that emits little to no light and is theorized to make up a portion of galactic DM. (Named in response to “WIMP” by folks with a sense of humor.)

MOND: Modified Newtonian Dynamics. MOND is a hypothesis proposing that gravity itself departs from standard Newtonian laws at ultra-low accelerations, acting stronger than predicted. The theory attempts to explain galaxy rotation speeds without DM.

WIMP: Weakly Interacting Massive Particle is a hypothetical subatomic particle that has mass and interacts only via gravity and the weak nuclear force (hence “weakly interacting”), making it a leading candidate to explain invisible DM.

1. Why The Search for Dark Matter Is Hard

Virtually every instrument we use to observe the universe is electromagnetic in nature. Our eyes are photon detectors. Optical telescopes, radio dishes, X-ray satellites, and gamma-ray observatories - all photon detectors too. They just differ in the range of frequencies they target. Astronomy is, partly, the science of catching photons that something else emitted, absorbed, scattered, or blocked.

However, this strategy doesn’t really work in the search for DM. DM has no electric charge, and so it doesn’t interact electromagnetically. It doesn’t emit, absorb, reflect, or scatter light. Light does “bend” around DM. This is important for us in detection efforts. But that’s just gravity curving the light’s “path”, not an actual “interaction” with DM.

This challenging situation is not because our investigative toolkit is “insufficiently sensitive” and needs improvement. The toolkit is effectively blind to the thing we are trying to find. (Like a radio receiver trying to detect a smell.)

Detector design choices. Given the detection challenges, even conceptualizing a new DM detector involves guesswork. Specifically, we need to guess what DM might couple to and how. For example, let’s consider a particular type of detector that is currently used in the DM search. A xenon time projection chamber is optimized to find particles in a certain mass range with a specific interaction model by seeing how they scatter (off nuclei). But, if real DM happens to be a thousand times lighter (than what the detector is optimized for), or if it doesn’t couple to nuclei, etc., that detector will never see any signs of DM, no matter how long it runs.

Comparing DM to neutrinos. As we know, neutrinos are part of the Standard Model and are also “invisible” to electromagnetism. Yet, we do detect neutrinos regularly. Ray Davis started looking for them in the 1960s with a tank of cleaning fluid. So how are neutrinos different from DM? Well, neutrinos still feel the weak force. The detection rate this enables is tiny but nonzero. But “tiny but nonzero” just implies that we have a hard engineering problem to solve. What we usually do for neutrinos is build something enormous (a giant tank of something appropriate), put it underground, and count the number of detections over a long period of time. The “WIMP” detection program was, in fact, based on the notion that DM “might” be like neutrinos in this respect. (WIMPs are a candidate type of DM particle that we assumed might couple with the weak force just like neutrinos. That’s what the “weakly interacting” part in the WIMP acronym means.)

Since we are comparing DM and neutrinos, we should note that neutrinos are technically a form of DM. There’s not all that much neutrino mass out there in comparison to the more elusive type of DM we’ve been searching for. But neutrinos are a real, confirmed, non-baryonic gravitating component that emits no light. And there are almost as many of them as there are photons in the cosmic microwave background (CMB). (We’ll come back to the question of neutrinos as a DM possibility later.)

So the main research question we face is not really “can invisible non-baryonic matter exist?” We already know that it can. The question is “what makes up the dominant part of dark matter?”

Using gravity for detection. If DM doesn’t interact with electromagnetic, strong, or weak forces, then gravity is our only option left. But, gravity is weaker than electromagnetism by about 10^36. The only reason effects from gravity are detectable is that they add up cumulatively and never “cancel”. Electric charge comes in two signs, and +/- electric charges neutralize each other. Mass comes in only “one sign” (from a gravitational perspective) and thus adds up. This means that, over large scales, gravity wins by default. But it also means that gravity presents its own challenges for the more subtle types of searches.

2. What Dark Matter Isn’t and How We Know

Before getting into the topic of what evidence we currently have for DM, let’s clear away a few possibilities that were ruled out a long time ago. From what I can tell, there seems to be a good bit of public confusion about DM that stems from treating these older possibilities as still open options. Ruling things out is actually a large part of what the field has accomplished so far.

DM isn’t antimatter: antimatter annihilates with matter, producing a specific gamma-ray signature. Over time, we’ve mapped the gamma-ray sky in great detail. What we see in the data is nowhere near what a universe with similar amounts of both matter and antimatter would produce. So the antimatter isn’t actually there, or we’d have seen it. Especially since antimatter isn’t actually “dark” - it interacts electromagnetically “exactly” as strongly as matter.

DM isn’t ordinary matter that’s just too dim to see: this is a possibility that might pop into an intuitive mind - maybe DM is just ordinary matter, but for some reason, it’s too “dim” for us to see using our instruments. This is a valid scientific hypothesis and, in fact, eliminating this possibility took some effort. Let’s look at the evidence from two different sources.

First, let’s look at what we get from Big Bang Nucleosynthesis (BBN). This was the process that created the initial light atomic nuclei in the first few minutes after the Big Bang. The abundances of deuterium, helium, and lithium from BBN depend on the “baryon-to-photon ratio”. We have measured those abundances, which lets us calculate the total baryon density in the universe. The answer turns out to be around 5% of the critical density. (Measurements tell us that our universe is mostly flat. Critical density is the average amount of matter needed for that “flatness” to happen.) If we then compare that number against the total matter density, we arrive at two conclusions: most baryons emit no light, and the dominant component of the total matter out there (the remaining 85% or so) cannot be baryonic matter. That is, it has to be something other than ordinary matter.

Second, the MACHO searches. One DM possibility was that the DM halos we see around galaxies are “dim compact objects” (e.g., brown dwarfs, old white dwarfs, neutron stars, stellar black holes). If this were the case, those objects would occasionally pass in front of background stars and “gravitationally microlens” them. Over the years, scientists from the MACHO, EROS, and OGLE collaborations have monitored millions of stars in the Magellanic Clouds, looking for clues from lensing surveys. They found some matches, but nowhere near enough. This doesn’t rule out MACHOs per se. But it does constrain them - compact objects (from roughly 10^-7 to 10 solar masses) can be, at most, a small fraction of DM halos.

So the somewhat boring possibility that DM might have just been ordinary matter too dim to see was tested in two different ways and rejected. Whatever supplies most of the missing mass must be something else.

DM isn’t (mostly) black holes: primordial black holes have been considered to be a viable DM candidate for a long time. These are far smaller and much less massive than the regular black holes we usually hear about. Black holes of stellar-mass and larger have been ruled out as DM possibilities by the microlensing surveys we just talked about, as well as by the CMB imprint of their accretion, and by reviews of LIGO’s merger population (these are mergers between DM black hole binaries). Besides, the black holes that form from collapsing stars are made of baryons, so they were already counted in the BBN census (three minutes after the Big Bang). The only black holes that are not included in that census are the primordial ones. We’ll cover PBHs in more detail in a later section.

DM isn’t ordinary neutrinos: we briefly discussed neutrinos being a possible DM candidate above. Neutrinos fit the specs for DM and were briefly a leading DM candidate in the 1980s. But neutrinos fail as the main source of DM because they’re “hot”. Neutrinos were relativistic when structure began forming, so they free-streamed out of small perturbations and erased them. So, if DM was mostly neutrinos, that would have given us top-down structure formation (superclusters condensing first and then fragmenting into galaxies), whereas we know that our universe assembled bottom-up (small things merging into bigger ones). Additionally, we can examine the estimated mass of neutrinos to see if they would suffice as the source of most of DM. Based on oscillation experiments, we can set a floor for the mass of a neutrino of about 0.06 eV. We can also use cosmological observations to cap the mass from above at roughly 0.1 eV. So, this leaves a narrow window of possibilities for the mass of a neutrino that is nowhere near enough for neutrinos to be a major source of DM.

3. The Evidence

Let’s look at the best evidence we have so far for the existence of DM.

The CMB acoustic peaks

To understand a key part of the evidence for DM, we need to understand an aspect of the early universe called “acoustic oscillations”.

(The next few paragraphs are a little technical, but they lay out an important part of the evidence for DM.)

Primordial acoustic oscillations. Before recombination (until about 380k years or so after the Big Bang), the universe was an ionized plasma. In that state, photons and baryons were tightly coupled through frequent scattering with free electrons. Gravity kept trying to pull the baryons into overdense regions. Radiation pressure from the photons kept pushing back. This interaction-based “tug of war” between gravity and radiation pressure set up acoustic oscillations - literal standing sound waves rippling through the primordial fluid. Then, when the universe cooled enough for neutral atoms to form, the coupled photons abruptly decoupled and streamed free. This “froze” the oscillation pattern into a snapshot. The angular power spectrum of the cosmic microwave background (CMB) is “that snapshot”.

The peaks in the CMB spectrum. When we take a look at the peaks in the spectrum, we can see characteristics of the sound waves as they existed at the time of the snapshot. The odd-numbered peaks correspond to maximum compression, the even ones to maximum rarefaction. (Compression and rarefaction are components of longitudinal waves.)

Unlike baryons, DM does not interact with photons. So it neither experienced radiation pressure nor participated in the oscillations. Instead, DM “clustered” and helped create and sustain the gravitational potential wells in which the photon–baryon fluid oscillated. That created an extra, steady gravitational pull that broke the symmetry between compression and rarefaction in the oscillations and enhanced the odd peaks relative to the even ones.

Also, since DM didn’t oscillate, its steady gravity “drove up” the amplitude of the acoustic peaks. This dynamic shows up most clearly in the third peak, which we see is nearly as tall as the second one instead of decaying away (as it might have in a universe of baryons alone).

The impact of ordinary matter shows up differently from that of DM. The inertia of the baryons “loaded” the primordial fluid without adding pressure. As a result, the compression phase inside each potential well was deepened. This enhanced the compression peaks (the odd-numbered ones) relative to the rarefaction peaks.

We could not “fake” the DM signal in the CMB data with more baryons. Baryons change the CMB peaks differently. If there had been more baryons in the mix (as opposed to DM), we would have gotten more inertia in the oscillating fluid. That would have resulted in a sharper odd–even contrast, a lower sound speed which would shift the whole peak pattern, and a small-scale damping signature (the Silk damping we discussed earlier).

What we have here is the use of a theoretical framework to predict specific quantitative fingerprints for ordinary vs. dark matter, and the predictions exactly match what we see in the CMB data. The contributions from ordinary and dark matter are statistically separable, and we can use data from the Planck experiment to separate them.

Results from Planck. Within the standard LCDM model, Planck provides simultaneous measurements of the baryon and CDM densities (at about 1% accuracy or better). The baryon density we observe also agrees with what we (theoretically) get from BBN in those first few minutes after the Big Bang. This really is an amazing consistency check across totally different physics at totally different epochs.

(We shouldn’t think of baryons vs DM as two independent knobs. The effects are physically entangled, but they affect the spectrum differently enough that the densities can be disentangled in the data. The baryon loading primarily controls the odd-even contrast, while DM strongly affects the gravitational potentials and therefore the “overall” peak envelope.)

A correct prediction from a no-CDM cosmology

In 1999, before the CMB peaks were measured, Stacy McGaugh worked out what the spectrum would look like with no CDM at all. He used the BBN baryon density and (otherwise conventional) general relativity as a benchmark. The prediction he got was a first-to-second peak amplitude ratio of 2.4, following from the BBN baryon density. And, when the WMAP subsequently measured it, the answer was 2.34 ± 0.09.

Well, that was a correct theoretical prediction that did not come from a model with CDM. The second peak came in considerably smaller than LCDM models had anticipated before the data existed. The no-CDM benchmark got it right with no free parameters. This was a strikingly successful cosmological prediction made in the MOND context, even though it was not a prediction derived from the MOND force law itself.

But the same 1999 model also made a second prediction, and this one it got wrong. Without CDM, there’s nothing but baryonic damping to generate the spectrum. So each peak should have been lower than the one before, and the third peak in particular should have fallen well below the second. But that’s not what we observe. We see that the third peak is nearly as tall as the second. This is exactly what we would expect if the steady forcing of a non-oscillating component (like CDM) was holding it up, and what a baryon-only universe would not produce. McGaugh, to his credit, did concede that the third peak falsifies the simple no-CDM model.

So what’s the scorecard on this test? One successful and one falsified prediction each for no-CDM. Moreover, the falsified one is the one that directly tests for DM. McGaugh’s position is that this only kills the simple ansatz, not MOND itself. This is a fair and reasonable conclusion. But that also means the CMB is now evidence “for” a DM-like component in “any” framework, even in MOND. I’ll come back to this in section 6.

(The MOND ansatz is that at very low accelerations, gravity behaves differently in its strength.)

Structure formation and the matter power spectrum

We discussed large-scale structure formation earlier, including how a baryon-only setup would not result in the universe we see today.

Large-scale structure in the early universe is generally thought to have arisen through the growth of density fluctuations. We can model the linear growth of those density fluctuations and quantify the overall fluctuating density field. We call this the matter power spectrum. (The first 10 slides or so in this presentation explain some of the calculations in a straightforward way.)

Beyond just a yes/no prediction of DM, the detailed shape of the matter power spectrum encodes the properties of DM itself. The spectrum has a characteristic “turnover scale” that corresponds to maximum amplitude and represents an epoch when matter and radiation had equal density in the universe. The position of the turnover (which corresponds to large-scale features) in the spectrum measures the total matter density. If instead we look at the very smallest scales (dwarf galaxy size and below), we see different effects. At those scales, if the DM had been warm, its particles would have moved fast enough to stream out of small clumps and smooth away structure there. So a missing-power cutoff at those really small scales measures how warm DM might have been. That is what the “Lyman-alpha forest” looks for.

The Lyman-alpha forest is the pattern of absorption by intergalactic hydrogen in the light from distant quasars. A spectrograph of the forest lets astronomers probe some of the smallest cosmological scales. Galaxy surveys measure “clumpiness” on large scales, and the forest helps analyze things at the scale of individual small halos. These measurements place strong limits on warm DM. For standard thermal relic DM, the particles generally need to be heavier than a few keV (roughly 3 to 6 keV) to behave as “cold enough”.

Finally, let’s review what we get from baryon acoustic oscillations (the same primordial sound waves that show up as CMB peaks). They show up imprinted as a preferred galaxy separation scale of about 150 megaparsecs. Astronomers have measured this with precision using BOSS, eBOSS, and DESI as a ruler. The fact that baryons participated in the oscillations and DM didn’t is reflected in the amplitude of the wiggles. What we see is a much stronger wiggle feature than what we might (otherwise) have seen with a baryon-only universe. This implies that most of the matter did not oscillate at all and could not have been just ordinary matter.

Galaxy Clusters

So far, the evidence we’ve looked at has largely been at cosmological scales. Let’s move down in scale to galaxy clusters.

Cluster weight. The DM story actually began with galaxy clusters in 1933, when Fritz Zwicky found the Coma cluster far too massive to be held together just by the galaxies he could see. Clusters have now been weighed in three independent ways: member galaxy velocities, X-ray gas temperature, and gravitational lensing. The results from the three methods more or less agree. We get a makeup of 85% dark matter, with the remaining sixth or so making up the rest.

We can also estimate the fraction of the mass that is baryonic. In galaxy clusters, the hot X-ray gas outweighs the stars by about ten to one. So most of a cluster’s ordinary matter is gas we can weigh directly from its glow. If we count it all, the baryon fraction comes out to around 15%.

These percentages obtained using multiple methods are consistent with the cosmic BBN value and far short of the lensing mass. So something like CDM must exist to account for the additional mass we detect via lensing.

Colliding clusters. In the Bullet Cluster, two galaxy clusters passed through each other in the distant past. The gas in those clusters, representing most of the baryonic mass, slowed. The lensing mass sailed through and ended up spatially “offset” from the gas, sitting out ahead of it. This case isn’t a one-off. We have other clusters (MACS J0025.4-1222 and the “Musket Ball” cluster) that show the same type of separation. This is harder for a MOND-style approach to explain.

(There’s another example - Abell 520, dubbed the “train wreck,” which is messier and genuinely anomalous in any framework. I mention it because it’s messy for both MOND and CDM models to explain as things stand. Also, as of 2026, the highest-resolution mass maps of the Bullet Cluster show that the collision was more complex than a clean two-body smash, with at least three mass clumps and a merger that was more lopsided than once thought.)

Despite what many CDM advocates say, the Bullet Cluster case does not, by itself, kill MOND, because MOND already requires extra matter in clusters, and that extra matter can supply the offset. What it eliminates is the simplest modified gravity explanations. It also forces any surviving models to carry a collisionless mass component. As I said earlier, MOND needs something like DM, too.

Why separability is important. There is a powerful “separability argument” here. If ordinary matter can end up in a galaxy with little or no inferred DM, while other systems show the opposite spatial pattern, this is naturally explained by a collisionless, gravitating component that can be dynamically separated from the gas and stars. A purely modified force law has a harder time explaining this. We don’t want to overstate that into a proof that DM “has to be a substance”: nonlinear modified-gravity theories and extra gravitational degrees of freedom can, of course, produce more complicated effective mass distributions. But separability is exactly the kind of behavior a particulate or field-like DM component predicts naturally, and any modified-gravity alternative has to match that level of explanation.

Galaxies that lack dark matter

Let’s talk about a newer set of discoveries.

We’ve discussed that we need something like CDM to explain galaxy formation. But could there be galaxies without CDM that are nevertheless consistent with the CDM model? LCDM has long predicted one class of DM-free galaxy - “tidal dwarfs.” Tidal dwarfs are formed from baryonic debris “stripped out of” interacting galaxies. The debris comes from the DM-poor disk, so the resulting dwarfs are born without DM halos. Known tidal dwarfs do show roughly baryonic dynamics (meaning no DM).

Anyway, recently, the situation got more interesting. In 2018, van Dokkum and collaborators reported that the ultra-diffuse galaxy NGC 1052-DF2 contains essentially no DM. Its dynamical mass, measured from the velocities of its globular clusters, matches its stellar mass alone. Another galaxy, DF4, followed with the same property.

(There was a debate over the distances of those galaxies from Earth that was settled in favor of the original claim by deep Hubble tip-of-the-red-giant-branch measurements.)

DF2 and DF4 then turned out to be part of a tight, statistically significant linear alignment of about a dozen faint, kinematically connected galaxies, with velocities increasing linearly along the trail. The discovery team’s interpretation is a dwarf-scale analog of the Bullet Cluster. They see this as a result of a single high-speed “bullet dwarf” collision that separated the DM from the gas. The gas then formed a string of DM-free galaxies. The scenario also accounts for DF2 and DF4’s unusually luminous, nearly identical globular clusters and their large sizes. And in 2026, a third galaxy on the trail, DF9, was identified as also lacking DM.

(Note: the bullet-dwarf picture is the discovery team’s “interpretation,” and so not yet a consensus. But the underlying dynamical measurements seem to have survived every challenge so far. And the prediction that other trail members would also lack DM has now been borne out once.)

The bottom of the evidence chain: rotation curves

Galaxy rotation curves, established by Vera Rubin and Kent Ford in the 1970s, are historically important. They are also quite rhetorically vivid and thus easy for people to grasp. They provide extraordinarily strong evidence for a mass discrepancy phenomenon in galaxies. What they provide is comparatively weak evidence for the more specific claim that the discrepancy must be due to a collisionless DM halo rather than modified dynamics. They also happen to be the place where modified gravity is most competitive with CDM. In fact, if galaxy dynamics were all we had to go on, something like MOND would be a serious contender vs DM as a theory. But it isn’t, so the choice between CDM and MOND isn’t close once the cosmological evidence is included.

The specification of what CDM is like

Let’s see if we can characterize DM based on everything we have so far. Given the state of the evidence, the data constrain the types of models that could resolve the galaxy rotation curve anomalies. It also provides us with some “specifications” for what DM looks like:

• Abundance: DM is about 27% of the energy density (5% ordinary matter, 68% dark energy). About 85% of matter.

• Cold: DM was non-relativistic when structure formed. The Lyman-alpha bound quantifies exactly how cold (slow).

• Nearly collisionless: DM’s self-interaction is below roughly 1 cm²/g based on cluster collisions. But it’s not zero.

• Mostly dissipationless: DM can’t efficiently radiate energy away. Baryons can cool and collapse into disks, while the dark component stays spread out in large, puffy halos.

• Stable: DM has existed for over 13.8 billion years (it doesn’t decay).

• Feeble non-gravitational coupling. Doesn’t seem to couple to anything other than gravity.

4. What Kind of Thing Could Dark Matter Be?

When we think about what kind of thing DM could be, the right question to ask is broader than just “what’s the right DM particle?” A more flexible set of questions would frame this in terms of a landscape or a sector. That is, what does the “dark sector” contain, what are its internal interactions, and how (if at all) does it talk to ordinary matter? This distinction matters. Decades of null searches have refuted some particular models of DM, but haven’t really impacted the astrophysical and cosmological inference that something DM-like exists.

Physicists have a term for the ways a hidden/dark sector can couple: “portals.” (A terrible nomenclature choice, if you ask me.) There are only a few simple options: through the Higgs, through “kinetic mixing” with the photon, through neutrinos, and through axion-like couplings. Thirty years of null experimental results have been “bounding” the parameter space for these portals.

Why expect just one particle? Well, we shouldn’t. The visible sector has six quarks in three colors, six leptons, twelve force carriers, a Higgs, and three gauge groups. Not to mention a rich structure of bound states, composites, chemistry, etc. The visible sector accounts for only 5% of the energy budget. DM contributes 27%. Should we really expect that the larger component is a “single inert species” while the smaller one is like a particle zoo? If we drop that assumption, then a whole landscape opens up.

Let’s explore some possibilities.

A Dark Sector

Dark photons and dark radiation. We could, for example, give the dark sector its own hypothetical unbroken U(1) gauge symmetry (an invariance of the physical system under local phase rotations of quantum fields). This would mean that DM has a long-range force acting only on itself, carried by a light “dark photon.” (“Light” here simply means low or no mass.) This would also produce self-interaction, possible dark recombination, and extra radiation density that would be visible to CMB experiments. As of today, this idea is constrained, but not fully excluded, and is directly testable by CMB-S4 (or would have been testable had CMB-S4 not been shut down by the Trump admin).

Atomic dark matter. Two dark fermions with opposite dark charges could form bound states, so we’d get dark atoms with dark chemistry. If that sounds complicated, well, it’s only complicated in the same way as ordinary matter is, which we know exists.

Partially interacting dark matter. What if we consider the possibility that only a fraction of DM has rich interactions? In that case, a dissipative subcomponent could cool into a thin, dark disk within the galaxy (coplanar with the visible one), while the bulk remains in a spherical halo. This is testable using data from Gaia. (Gaia results have not been kind to this idea. As of today, current constraints permit, at best, a small dissipative fraction.)

Self-interacting dark matter. There is also the possibility that DM only interacts with itself (SIDM), and that this effect is potentially velocity-dependent, so that it matters in slow dwarf galaxies but switches off in fast clusters. In that scenario, scattering would turn the normal CDM-predicted steep central cusps into flatter cores. In dense systems, this can trigger a runaway gravothermal core collapse, producing anomalously “dense” centers that we can look for. (I’ll come back to this in Section 7.)

Questions every dark matter model must answer

There are two questions that every candidate DM model must be able to answer.

First: “Why is it dark”? There are only a few credible answers. DM might be:

• a Standard Model singlet carrying none of our charges (the simplest but also the hardest to detect).

• a pseudo-Goldstone boson of a broken symmetry, which is the axion’s answer and explains both its lightness and feeble coupling.

• a neutral composite of a confining dark sector, the way a neutron is neutral.

• a gravitational relic with no Standard Model coupling by construction.

Second: “Why is it stable”? Surviving 13.8 billion years is remarkable. To make that happen, DM would require something like:

• an unbroken symmetry (e.g., R-parity is what stabilizes the SUSY neutralino - more on this in the next section), or

• a fortunate accident of quantum numbers (this is what prevents the proton from decaying), or

• It’s so light and feebly coupled that there’s nothing to decay into.

Any candidate particle that can’t address both concerns cleanly isn’t really a candidate.

A coincidence worth thinking about. The DM density is about five times the baryon density. When we consider thermal freeze-out, that ratio is seen as an accident, since the two values are set by unrelated processes. But asymmetric DM takes the coincidence seriously and ties its abundance to the baryon asymmetry via a shared mechanism. Most models naturally predict a mass of around 5 GeV and no annihilation signal today (no antiparticles left to annihilate with). Searching for something like that is a genuinely different research program from the WIMP one (and underexplored as things stand today).

In any case, if the dark sector is rich, a position like “we looked for a WIMP and didn’t find it” is not very informative. It’d be like an alien civilization concluding there’s no life on Earth after scanning one radio frequency.

5. The Candidates

Time to meet the contestants. Let’s take a at look at five of the top candidates for what DM might be. There are others, but these are the big 5.

WIMPs

WIMPs were the primary DM candidate for thirty years because of the “WIMP miracle.” The idea behind WIMPs was as follows: a new type of particle was created in large numbers after the Big Bang, but it underwent self-annihilation due to the weak force. That dynamic stayed balanced until, eventually, the expansion of the universe diluted things enough that annihilation couldn’t keep up. So, these particles “froze out.” The leftover density would depend almost entirely on the annihilation cross-section. If we consider a weak-scale annihilation cross-section, that would give us essentially the DM density we observe in the universe today. So we get a cosmological answer to a problem from (seemingly) unrelated particle physics.

One reason WIMPs were compelling as a DM candidate was that supersymmetry (SUSY) independently predicted the existence of a stable weak-scale particle (the neutralino) for reasons unrelated to cosmology. In 2012, the LHC found the Higgs in the mass range that precision electroweak data had favored, but also found no evidence of supersymmetry or any of the conjectured superpartners, including the neutralino. This hasn’t killed the idea of SUSY or WIMPs outright, but it did eliminate the two-birds-one-stone argument that made WIMPs feel inevitable as the main constituent of DM.

The thermal freeze-out model also remains alive. What freeze-out fixes is the “annihilation” cross-section. Underground detectors, however, measure the “scattering” cross section. The two are tied together only in the simplest models, where the DM annihilates to Standard Model particles through the very mediator it scatters off. In so-called “secluded” DM models, the DM freezes out by annihilating into dark states (dark photons, dark Higgses). Calculations show that the relic abundance comes out right, and the scattering cross section can be arbitrarily small. So the null results we’ve had so far weigh mainly against “minimal WIMP models”, not against thermal freeze-out as a production mechanism.

The status of the search. Direct detection has improved by roughly eight orders of magnitude over four decades. The leaders are multi-tonne dual-phase liquid xenon time projection chambers: LZ in South Dakota, XENONnT at Gran Sasso, and PandaX-4T at China Jinping. LZ’s December 2025 analysis set the world’s most stringent limit across the main WIMP range and, for the first time, extended its reach below 9 GeV, still with no evidence of WIMPs.

Since 1997, the DAMA/LIBRA experiment at Gran Sasso claimed an annual modulation in its event rate that was interpreted as the DM wind from Earth’s orbital motion. Efforts by others to replicate it have effectively ruled out a DM origin for that data. (Nobody has fully explained what DAMA sees, which is unsatisfying, but it isn’t DM.)

So the classic, natural SUSY WIMP is mostly gone as an option. Corners of the parameter space survive, but they’re finely tuned, and avoiding tuning was the whole appeal in the first place.

Result: not dead yet, but no longer special. Worth exhausting the search because we’re close to being able to. Wouldn’t bet on it though.

Axions

The axion is a hypothesized particle that arises from the Peccei-Quinn solution to the strong CP problem. Within the Standard Model, there is a parameter in QCD called theta that could have taken any value. Actual measurements show that it’s below about 10^-10, which is very small. This is not a contradiction, but it does pose the question of why the value is so suspiciously close to zero. The Peccei-Quinn fix promotes theta to a dynamical field that relaxes to zero. The quantized excitation of that field is the axion. (Frank Wilczek named it after a laundry detergent, on the grounds that it “cleaned up the problem”.) Later, someone noticed that a very light axion produced through the misalignment mechanism would be cold, stable, and could have the right relic density to be a candidate for DM. Another case of two problems, one solution - similar to WIMPs.

Cosmology impacts the mass prediction. How much axion DM you get depends on whether Peccei-Quinn symmetry broke before or after cosmic inflation.

Before inflation: our observable universe inflated from one patch with one initial misalignment angle, which is a free parameter, so a wide range of masses work, and the theory predicts little (though CMB isocurvature limits do constrain it).

After inflation: the misalignment angle averages over many causally disconnected patches, the abundance becomes calculable, and we get a real mass prediction. The complication is that this scenario also makes cosmic strings and domain walls whose decay produces more axions, and computing that requires hard lattice work whose answers have moved around. Pinning down the number means simulating the tangle of cosmic strings and domain walls that radiate axions, and that simulation is brutally hard to extrapolate to the real world. Current post-inflationary estimates cluster in the tens to hundreds of micro-eV, but they still scatter across more than an order of magnitude of spread. This spread is why people talk about “the axion window”.

How we hunt for axions. Axions convert to photons in a strong magnetic field (the Primakoff effect). A haloscope is a tunable microwave cavity (a closed structure that traps and resonates electromagnetic waves at microwave frequencies) inside a powerful magnet at millikelvin temperatures, with a quantum-limited amplifier listening for excess power at a single frequency. We can tune this setup slowly and listen. Imagine scanning an AM radio dial across the whole spectrum, one hair-width at a time, for a station broadcasting at unimaginably low power.

Two benchmark model classes help us define “goalposts” for sensitivity. In KSVZ models, the axion couples to photons only through heavy exotic quarks. In DFSZ models, the axion also couples to ordinary Standard Model fermions, and its photon coupling comes out about 2-3 times weaker. So, “KSVZ sensitivity” is the easier target, and “DFSZ sensitivity” is the demanding one that covers both benchmarks.

The status of the search. ADMX has excluded DFSZ axions of masses between 3.27 and 3.34 μeV at 90% confidence, assuming a standard halo made entirely of axions. A more recent run excluded the more strongly coupled KSVZ axions, even at fractional DM densities, though it did not reach DFSZ sensitivity at an acceptable scan rate. Upgrades are underway to push the searches even further.

In summary, only a few percent of the interesting range for axions has been covered in the search so far. Progress is slow, but squeezed-state receivers, better magnets, and multi-cavity designs are pushing scan rates up fast.

(There is also the broader family of axion-like particles (ALP). ALPs share some properties with axions but may have different masses, couplings, interactions, etc. than traditional axions. ALPs are hunted with helioscopes (CAST, then IAXO), light-shining-through-walls (ALPS II), and astrophysical constraints from stellar cooling and SN1987A.)

My read: personally, if I had to bet on one candidate today, it’d be axions. Not necessarily because evidence favors it, but because the motivation is independent (CP problem), the searches are cheap relative to the payoff, and the parameter space is being systematically covered.

Ultralight and fuzzy dark matter

If we conceptually push the axion mass down to around 10^-22 eV, the equivalent (de Broglie) wavelength grows and becomes kiloparsec-scale (dwarf galaxy size!). The resultant DM would then behave as a coherent classical wave, with quantum pressure resisting small-scale gravitational collapse. Called fuzzy dark matter, this possibility was seen as attractive because quantum pressure naturally produces flat cores in galaxy centers. So this was also seen as a way to solve multiple problems at once.

Where it stands: badly, at the canonical mass. If we assume that most of the DM is ultralight and fuzzy, this imposes some mass constraints. The Lyman-alpha forest would require a mass above roughly 10^-21 eV. Analyses of the smallest galaxies seem to exclude the entire 10^-22 to 10^-20 eV range. This constraint is also insensitive to baryonic physics uncertainties because it uses observed DM densities rather than predicted ones. Ultrafaint dwarf observations push the bound as high as 3 × 10^-19 eV.

My read: the whole idea of fuzzy DM was largely proposed to better explain small-scale structure, and small-scale observations have now squeezed the most attractive mass range for fuzzy DM very hard. There is still (model-dependent) room, but the simple version of the idea has lost much of what originally made it appealing.

Sterile neutrinos

Another natural DM candidate arises from the fact that neutrinos have mass. Besides axions, sterile neutrinos had been an option that I had (previously) found most promising.

The Standard Model does not account for neutrino mass. The most popular hypothesized mechanism for neutrino mass generation also naturally implies the existence of right-handed neutrino states. (In the Standard Model, we only have left-handed neutrinos.) Those states would be “sterile,” feeling neither the weak force nor anything else except gravity and a slight mixing with ordinary neutrinos. A keV-scale sterile neutrino would be a warm DM candidate, sitting right at the Lyman-alpha boundary we discussed earlier.

We had a bit of experimental drama related to sterile neutrinos in 2014: an unidentified X-ray emission line showed up in stacked spectra of galaxy clusters - exactly what a 7.1 keV sterile neutrino decaying would produce. Twelve years of contested follow-up experiments and analysis ensued! The best current data comes from XRISM’s microcalorimeter. Stacking 3.75 megaseconds of early XRISM observations of ten clusters, no unidentified lines were detected, with a 3-sigma upper limit on the 7.1 keV decay rate of about 1.0 × 10^-27 per second: three to four times below the Hitomi limit, though still about five times above the original XMM-Newton detection level. XRISM may reach the required sensitivity by stacking several more years of cluster observations.

(Notes: XRISM was looking for spectral lines that might have shown the decay signature of a sterile neutrino. Hitomi was a Japanese X-ray telescope that was lost after producing some exciting new results.)

My read: No evidence so far, but probably not yet dead as a proposal. Independently, the simplest production mechanism is strongly constrained by combining X-ray non-detections with structure formation limits. So, as a particle to explain “all of DM”, sterile neutrinos are in serious trouble. But just as a component of DM, less so.

Primordial black holes

Primordial black holes (PBHs) require no new particle physics, which is either a good thing or a bad sign (depending on your temperament). Large enough density perturbations in the early universe could have collapsed directly into black holes (before any stars existed). The resultant black holes would also be cold, collisionless, dissipationless, dark, and stable. Which matches the DM specs we want. (And they form before nucleosynthesis, so they’re “not” counted in the BBN baryon census.)

Almost every mass is excluded. Below about 10^17 g, Hawking radiation would have evaporated them or lit up the gamma-ray sky as they died. Above about 10^22 g, the microlensing surveys should have caught them. In between those limits sits the stubborn so-called “asteroid-mass window”, hard to constrain because the relevant Schwarzschild radii there are sub-Angstrom (smaller than an atom despite the asteroid-scale mass!). The magnification from microlensing switches off when the lens’s Einstein radius drops below the apparent size of the source star. Accounting for realistic extended mass functions, the window narrows but remains open: all of the DM could still be PBHs there.

My read: still a possibility that deserves more attention. But if this is the answer, then every particle detector built for the purpose of finding DM is looking in the wrong place. Resolution would have to come from MeV gamma-ray astronomy or high-cadence microlensing instead.

Everything else

What are some miscellaneous options? Well, there are dark photons and hidden sectors, asymmetric DM, composite DM from dark quarks, ultraheavy candidates up to Planck-scale masses, freeze-in dark matter too feebly coupled to ever thermally equilibrate, pure gravitational relics, gravitinos, Q-balls, and Kaluza-Klein states. Most are much harder to test than WIMPs or axions, which is precisely why the field spent so long on the two most testable options. (That’s a bit of a selection effect in the DM search program, but such is life.)

6. Modified Gravity: What It Gets Right, and Why I Don’t Think It’s the Answer

So let’s talk about modified gravity as an alternative to DM. The logic behind modified gravity is straightforward. Every piece of DM evidence is an inference of the form “there’s more gravity here than the visible matter accounts for.” Those inferences assume that we know the “correct” law of gravity. If the law we currently have is wrong at “very low accelerations,” then the inferences collapse, and there’s no missing matter to hunt for.

(Note: This is the sort of reasoning that let Le Verrier successfully predict Neptune’s existence from Uranus’s anomalies. He also tried (and failed) to predict Vulcan from Mercury’s perihelion precession - that required Einstein to formulate GR to explain the anomaly.)

Modified Newtonian Dynamics (MOND)

MOND is Milgrom’s 1983 proposal for modified gravity. Milgrom had conjectured that below a characteristic acceleration a0 of about 1.2 × 10^-10 m/s², the acceleration scales as 1/r rather than 1/r². The theory has just one universal parameter, the a0 acceleration scale.

MOND’s successful predictions. MOND predicted flat rotation curves from the distribution of visible matter alone. It predicted the baryonic Tully-Fisher relation (link between a galaxy’s visible mass and its rotation speed) with less scatter than CDM naturally gives. It predicted low surface brightness galaxy rotation curves before they were ever measured. The radial acceleration relation, the tight one-to-one mapping between observed and calculated acceleration of visible matter across galaxies, is exactly what MOND says must happen. And, as we discussed in section 3, a no-CDM cosmology advocated in the MOND context made a strikingly successful a priori prediction of the CMB’s first-to-second peak ratio. These successes are why MOND has been seen as a credible alternative to DM.

What MOND doesn’t do. It helps to keep three things separate here: Milgrom’s empirical MOND law, relativistic MOND-inspired theories, and no-DM cosmologies. The original MOND relation is not a complete theory in the sense that LCDM is a theory. It is a phenomenological relation that summarizes a particular type of regularity in galaxy dynamics. (The word “ansatz” is sometimes used to refer to this sort of theorizing.) MOND has one parameter and fits rotation curves beautifully. But by itself, it is silent about the CMB, nucleosynthesis, structure formation, clusters, and lensing. Relativistic completions can address those things, but then the relevant question becomes whether they can do so without introducing extra gravitating fields or matter that behaves cosmologically like DM.

So MOND has real successes at galaxy scales. The problem is what happens when we ask it to do everything else a complete cosmological theory has to do.

The third peak. Remember when we talked about the CMB third peak in Section 3. The no-CDM ansatz made a falsifiable prediction, and it failed in the direction that would indicate a non-oscillating gravitating component (like DM). (More on this below.)

Clusters. MOND still faces a residual missing gravity problem in galaxy clusters by roughly a factor of two. This has persisted through decades of continuously improving X-ray and lensing data. MOND advocates patching this by adding actual DM to the theory, typically 11 eV sterile neutrinos, which is awkward since not needing DM was the selling point for modified gravity in the first place.

Separability. Colliding clusters, tidal dwarfs, and the NGC 1052 trail (section 3). These types of systems are meant to accommodate cases where your modification is a property of the law and cannot be switched off. (Since the modification is a law that applies throughout the system.)

Relativistic completions are hard to formulate. The plain version of MOND is non-relativistic. The leading relativistic version for years was TeVeS, and in 2017, the gravitational waves from a neutron star merger GW170817 arrived with its gamma-ray burst just 1.7 seconds behind, pinning the speed of gravitational waves to the speed of light at one part in 10^15. That killed TeVeS and a large number of its cousins overnight.

Relativistic MOND today

There are modern and more functional relativistic versions of MOND.

Skordis and Złośnik’s Aether Scalar Tensor theory (AeST), published in 2021, is one. AeST can now explain the CMB power spectrum while still allowing for gravitational waves that travel at the speed of light. AeST was also the first extension of general relativity to reproduce MOND in galaxies while agreeing with the standard cosmological model on the full CMB spectrum.

On the other hand, in AeST, the time-dependent term in the action behaves like gravitating dust. That is what allows the theory to reproduce the CMB power spectrum and the LCDM matter power spectrum on large scales. So, AeST theory fits the CMB by containing a component that gravitates like CDM.

We can see a common pattern emerging across all contemporary modified gravity proposals. The third peak demands something that behaves like a non-oscillating gravitating component, in any successful framework. It is proving very hard to write down a theory that fits the peaks without an additional degree of freedom that behaves cosmologically like pressureless matter. That does not, by itself, tell us whether the fundamental ontology belongs in the “matter sector” or the “gravity sector.” But observationally, the burden is becoming quite specific: whatever the underlying theory is, it needs something that clusters and gravitates very much like dark matter.

The wide binary test for modified gravity

There is one unusually clean place to test low-acceleration gravity laws directly - widely separated binary stars.

If we take binary stars separated by more than a few thousand AU, their mutual accelerations fall below the MOND a0 acceleration scale. Local DM density is negligible, so the test isolates the modified gravity law with no halo modeling required. In this scenario, MOND predicts roughly a 20% velocity boost over Newton.

In principle, this should have been one of the nicest discriminators between MOND and Newtonian gravity, and a “decisive test” for MOND. In practice, however, the current situation is a mess. Gaia has measured hundreds of thousands of systems. Instead of a clean answer, we get conflicting reports, some favoring MOND (Hernandez and collaborators, Chae) and others favoring regular gravity (Pittordis and Sutherland, Banik and collaborators). Chae’s most recent work maintains that the anomalous deviation from standard gravity is real and confirmed, while detractors continue to find the data consistent with standard gravity.

So, even though wide binaries were supposed to be a decisive test, what we have instead is an argument over real, messy data.

More precisely, the disagreement is largely about systematics, not data. Undetected close binaries, chance alignments, and improper treatment of projection effects can conspire to mimic MOND-like signals. A wide binary in which one component is secretly itself a tight pair will show anomalous velocities unrelated to gravity. The fraction of such contaminants is on the order of a few tens of percent, and the modeling approach largely determines the result. The pro-MOND camp responds that the null analyses compare actual noisy data to noise-free theoretical models, biasing the result toward regular gravity.

My read: hard to tell really. Future Gaia data releases - DR4 (2026) and DR5 (2030?) - should hopefully give us better data to make a judgment call. The study of wide binaries is a powerful tool because it probes the gravity law without astrophysical modeling in between.

Emergent Gravity

I thought I’d mention this in passing, since it’s often brought up in conversations on X - the alternative to LCDM called “emergent gravity”. This is a topic all by itself, but I’ll briefly look at one specific flavor, which is probably the most famous of the lot - the one from Verlinde.

Verlinde’s emergent gravity ties a0 to the cosmological constant with no free parameters, which is genuinely attractive, but fails against the radial acceleration relation without implausible mass-to-light ratios, faces internal consistency criticisms, and applies only to isolated static systems, so it can’t address the CMB.

The scorecard for modified gravity

Even if modified gravity turns out to be right, something DM-like is still very likely to be real. The acoustic peaks (including a falsified no-CDM prediction), BBN, structure formation, cluster gas fractions, colliding clusters, and DM-free galaxies form an interlocking case for DM with multiple converging lines of evidence that modified gravity has not been able to reproduce without effectively containing an additional gravitating component.

Galaxy-scale MOND phenomenology remains strikingly successful. What it has struggled with is the attempt to turn that phenomenology into a complete cosmology without introducing extra degrees of freedom that behave much like DM. So, even with modified gravity, we should expect that something with DM-like gravitational behavior is out there.

7. Some Other Issues With DM

(Note: DM and dark energy are separate problems. DESI’s hints of evolving dark energy are fascinating and have essentially nothing to do with what DM is. Conflating them is an error that seems to pop up often in popular discourse.)

That said, there are additional issues with CDM as a theory. Let’s start with a trio of problems that come from mismatches between simulations of DM in large galaxies like the Milky Way and actual astronomical observations:

Missing satellites. Simulations predict: large galaxies should be surrounded by thousands of small DM clumps (subhalos). This problem seems to have been (largely) resolved. Deep surveys have found dozens of ultrafaint dwarf galaxies, and modeling updates (of which halos actually light up) seem to have closed the remaining gap. Status: mostly solved.

Too big to fail. Simulations predict: large galaxies should be surrounded by massive, ultra-dense clumps of DM that have so much gravity they should have been forced to form stars (“too big to fail”), yet the corresponding bright satellites aren’t seen. Baryonic feedback in modern hydro simulations has substantially eased this concern, but some analyses argue it persists for field dwarfs. Status: somewhat solved, but contested.

Cusp-core. Simulations predict: a galaxy’s DM density should sharply spike at its center (a “cusp”). Actual data from dwarfs and galaxies with low surface brightness instead show a flat, evenly distributed center. This problem has been partly addressed by including supernova-driven fluctuations in the gravitational potential, flattening out the central cusps. Whether that works for the smallest, most DM-dominated dwarfs is still contested. Status: somewhat solved?

There is another unresolved issue in this area, called the diversity problem: the observed scatter in inner rotation curve shapes among similar galaxies is much larger than that produced by CDM simulations, while the radial acceleration relation is much tighter than that produced by the simulations. (Simultaneously too diverse and too regular!) Status: the analysis is ongoing.

Satellite planes. Satellites of the Milky Way, M31, and Centaurus A galaxies appear to lie in thin rotating planes (as opposed to being scattered isotropically). These are neat configurations that our CDM simulations rarely produce. Potential solution: Sawala and collaborators argue that with Gaia’s proper motions and proper accounting for the look-elsewhere effect, the Milky Way’s plane is a transient chance alignment (rather than a lasting structure), and so not in significant tension with CDM. Status: the discussion continues.gala

Strong lensing anomalies, the newest and most interesting. Galaxy-galaxy strong lensing in clusters shows an excess of small-scale lensing events compared to simulations (Follow-ups confirmed persistence of the issue). Real cluster substructures appear more compact and centrally concentrated than those produced by CDM. Status: too new to assess.

New anomaly, new possibilities

In 2025, there was an exciting new development. Powell, McKean, and Vegetti used gravitational imaging to detect a million-solar-mass object at a cosmological distance. This was a completely dark clump revealed purely by its perturbation of a lensed radio jet. It was also roughly a hundred times lighter than anything found this way before.

Even more striking are two lensed systems that don’t fit.

The first is SDSS J0946+1006, where pixel supersampling reveals a dark substructure at about 17 sigma, with a projected mass within 1 kpc of a couple of billion solar masses and a density profile steeper than a log-slope of −1.75. This is a clear outlier against models of cold, collisionless DM.

The second result came in 2026, when the same broad collaboration went back to the million-solar-mass object Powell had found in a different lens and tested 2 dozen models for it. Its structure is also incompatible with cold, collisionless DM but matches a self-interacting halo whose core has collapsed toward a central black hole. (The fit demands a very large cross-section, so it’s suggestive rather than settled.)

The relevance of these discoveries. Cores in some systems and anomalously dense centers in others are signatures of velocity-dependent self-interacting DM. If the pattern holds, this increases the likelihood that the dark sector has “internal physics,” discovered entirely through gravitational observations, without detecting any particles. This may currently be the most promising direction in the field.

Miscellaneous issues

The S8 tension. S8 is a cosmological parameter that measures how much cosmic matter has clumped together under gravity. For many years, data from weak lensing surveys consistently indicated that the low-redshift universe is less clumpy than extrapolations of Planck data suggest. This mismatch was dubbed the S8 tension. But the recent picture has become murkier rather than firmer. The KiDS-Legacy analysis in 2025 came out consistent with Planck, which substantially weakens the lensing side of the tension. Similarly, ACT’s CMB lensing measurements also find clustering consistent with Planck. So the S8 tension may be dissolving, but if it firms back up, dark sector explanations like decaying dark matter or dark matter-neutrino scattering are possible solutions.

Indirect detection. In the search for DM, indirect detection essentially looks for debris from DM particles that find each other and annihilate. For some years, observations suggested a tantalizing hint of something interesting - the so-called Fermi-LAT Galactic Center excess. This was a glow of GeV gamma rays from the inner galaxy. It had the right spectrum, roughly the right shape, and a relevant (implied) cross-section that looked just like annihilating DM. However, when the photon statistics were actually considered, it turned out this was probably something else - namely, pulsars. While DM would produce a smooth glow, what was observed in the excess was clumpy. Which is what we’d get if the signal were from a crowd of faint, unresolved millisecond pulsars. This is not a closed issue yet. But it’s an example of how the astrophysics under consideration can change the interpretation of a DM signal.

The Totani claim. Here’s another ambiguous indirect detection case. In November 2025, Tomonori Totani published a peer-reviewed analysis reporting a halo-like glow of gamma rays around the Milky Way from 15 years of Fermi data, peaking near 20 GeV. The glow had a spectrum that matched WIMP annihilation into bottom quarks or W bosons. (You have to model and subtract the known foregrounds, including interstellar gas emission, cosmic rays, and the Fermi bubbles, to actually get that.) To his credit, Totani himself notes the result must be verified through independent analysis and verification, and that even then, scientists will want proof the radiation isn’t from some other astronomical phenomenon. Press coverage around the paper said “dark matter finally seen.” The field’s reaction was much cooler, for a variety of reasons. This was far from a new claim, and numerous prior studies of the same data have failed to find evidence for it. The core problem here is that everything rides on the author’s foreground model. The scenario is much like the Galactic Center excess we just discussed, which looked equally WIMP-like for years before the pulsar interpretation took over.

My read: worth watching for independent re-analyses, currently far more likely to be foreground mismodeling than DM. In any case, a useful reminder that in indirect detection, the signal is only ever as good as the background astrophysical model.

What would actually undermine the case for CDM?

This brings us to a criticism that often comes up. If DM candidates can keep changing as experiments rule things out, is CDM even falsifiable? Yes, it is, but let’s actually spell out some basic falsification criteria.

The dark matter component of LCDM would be in serious trouble:

• If a modified gravity theory reproduced the acoustic peaks (including the third peak), BBN, cluster dynamics, and galaxy dynamics without any component gravitating like pressureless matter.

• If colliding systems consistently showed the lensing mass tracking the gas (rather than sailing past it).

• If DM-free galaxies turned out to be ubiquitous rather than products of rare stripping events.

Strictly collisionless CDM would be in trouble if the lensing anomalies solidify into a coherent velocity-dependent self-interaction signal. That would not falsify “cold dark matter” in the broader sense per se, since self-interacting DM can still be cold. It would instead tell us that the collisionless approximation is incomplete.

And while I’m on this topic, MOND would be falsified if the wide binary test, as we discussed previously, converges cleanly on Newtonian behavior with fully understood systematics. That test was designed to be decisive (which is to MOND’s credit as a scientific program).

It helps to think of the claims related to DM as a hierarchy. “Some additional gravitating component or degree of freedom exists” is extremely robust. “DM is predominantly non-baryonic and cold” is also very robust. “DM is effectively collisionless” is strong but directly testable. “DM is a particular particle, such as a WIMP with a detectable Standard Model coupling,” is much weaker and has been heavily constrained.

8. The Experimental Program

Let’s talk about the actual hardware that we use to run the DM research program.

(Feel free to skip this section if the experimental side of this is not that interesting to you as the theory.)

There are three main legs:

• production (making DM in colliders)

• scattering (direct detection - catching it recoiling off a nucleus underground)

• annihilation or decay (indirect detection - spotting particles it produces out in space).

Plus astrophysical probes, which I’d call a fourth leg and possibly the most productive recently.

Colliders and the intensity frontier

The relevance of colliders is often skipped in popular coverage of DM. But colliders are where the “portals” get probed directly.

Missing energy at the LHC. At the LHC, DM produced in a collision “escapes invisibly”, so we look for visible particles recoiling against nothing: mono-jet, mono-photon, mono-Z, mono-Higgs searches, etc.

Invisible Higgs decays. If DM couples through the Higgs portal and is lighter than about 60 GeV, the Higgs can decay to it. The Standard Model predicts an invisible branching ratio of about 0.1%. The ATLAS Run 2 combination sets an upper limit of 10.7% at 95% confidence. If we look at the relevant models, we get spin-independent scattering limits below about 60 GeV that are competitive with or better than the underground experiments in that mass range.

The intensity frontier. This is where most of the dark sector searches take place: high luminosity rather than high energy, to hunt feebly coupled particles. Belle II has an excellent program. A recent search using 365 femtobarns-inverse of data targeted inelastic dark matter with a massive dark photon, looking for displaced decay vertices that had missing energy. LHCb hunts low-mass dimuon resonances. Dedicated experiments like NA64, SHiP, and the forward LHC detectors like FASER cover long-lived, very weakly coupled regimes that nothing else reaches. If the dark sector story of section 4 is right, this is where a discovery would most plausibly come from.

Direct detection

When we talk about the search for DM, xenon detectors usually come to mind, so..

How xenon detectors work. A DM particle scatters off a xenon nucleus, producing a flash of scintillation light along with a spray of ionization electrons. Those electrons then drift up through the liquid to a gas layer where they generate a second, delayed flash. What’s important is the ratio “between” the two signals. The ratio can be used to cleanly distinguish the nuclear recoils (from the electron recoils that make up most of the background). Also, the precise timing lets us reconstruct the event’s position in 3D. The hardest part of the setup is probably the background suppression: xenon has to be purified of radioactive krypton and radon to almost absurd levels, and every material has to be screened for radioactivity. We need kilometers of rock overhead to stop cosmic rays and layers of veto detectors wrapped around everything. All of that for catching a (literal) handful of events per tonne per year!

Why xenon, specifically? Spin-independent scattering (interaction probability doesn’t depend on intrinsic spin states of particles) is coherent across the nucleus, so the cross section scales roughly as the “square” of the atomic mass number. A heavy nucleus like xenon (mass number around 131) gets a huge free enhancement over light targets. This single fact explains the choice of technology. It also explains the jargon: a “xenon-phobic” (nice pun) model is one whose couplings happen to interfere destructively for xenon’s proton-to-neutron ratio.

The sub-GeV frontier. This seems to be where the frontier has moved to (given all the WIMP null results). Below about a GeV, nuclear recoils fall under the detection threshold, so we go after “electron recoils with skipper-CCD experiments like SENSEI and DAMIC-M. Or with the xenon detectors’ ionization-only channels. Or we exploit the Migdal effect, in which a nuclear recoil shakes off an electron, giving us sensitivity below the nominal threshold. Cryogenic bolometers (SuperCDMS, CRESST, EDELWEISS) push the threshold below 50 eV. The theoretical side of this, light DM with light mediators, pairs naturally with the intensity frontier searches that we discussed above.

Coming soon. DarkSide-20k is under construction at Gran Sasso: 51 tonnes of underground-sourced argon, 20 of them fiducial, with first filling expected at the end of 2026. Operations start in 2027 - the setup can reach an order of magnitude beyond DarkSide-50 for light DM. (Argon’s pulse-shape discrimination rejects electron recoils at the 10^8 level.) And XLZD (merged LZ-XENON-DARWIN collaboration) plans a 60-tonne xenon target (80 if the market allows), with 3-sigma evidence potential down to 3 × 10^-49 cm² at 40 GeV. A kilotonne-year xenon exposure would do double duty: characterizing a signal if found, or reaching the neutrino fog across the GeV-to-TeV range if not. (The xenon market issue is not a joke. Global annual production is tens of tonnes, and we compete with semiconductors and medicine.)

Axion searches

Axion searches are done by a whole ecosystem of clever instruments (rather than a couple of flagships). ADMX is upgrading its scan rate. HAYSTAC pioneered quantum squeezing to beat the standard quantum limit (on amplifier noise). CAPP runs multiple cavities at high field. DMRadio targets lower masses using lumped-element resonators (instead of cavities). MADMAX (cool name) uses a dielectric haloscope (stack of discs in a mag field) for higher masses, where cavities become impractically small. IAXO is the next-generation helioscope (that is, aimed at axions from the Sun). CASPEr hunts axion-induced oscillating nuclear dipole moments with NMR. One challenge here is that sensitivity scales with mag-field squared, cavity volume, and quality factor, whereas the resonant cavity’s volume scales inversely with frequency cubed. Hence, the MADMAX dielectric and CAPP multi-cavity workarounds at high mass (since the cavities shrink to nothing).

Astrophysical probes

Finally, the astrophysical probes, which may be the most productive route of all, because they can constrain the “properties” of DM even without any lab detection.

Rubin’s LSST (the survey, not the previous name of the Rubin observatory, both of which are named “LSST”) will find new ultrafaint dwarfs, map microlensing (including PBH-relevant regimes), and deliver weak lensing over huge areas (including the primordial BH-relevant regimes). Euclid is measuring two billion galaxy shapes, with the first cosmology results expected around 2026 to 2028. So far, its strong lensing haul is already quite amazing. The Q1 quick release (on its own) found 497 strong lens candidates in just 50 square degrees, doubling the total number of known lens candidates with space-based imaging. Extrapolating to the full survey implies over 100,000 high-confidence candidates, including more than a thousand rare double-source-plane systems, which are exactly the J0946-type configurations we discussed that pin down dark substructure. Roman extends the search to higher redshift. Simons Observatory sharpens constraints on dark radiation and DM interactions. (CMB-S4 would have done this too, had it not been shut down.) CTAO (Cherenkov array) is the big one for indirect detection - it should reach below the thermal relic cross section at TeV masses, space nothing else currently touches. Strong lensing with JWST and eventually Rubin’s lens sample will map substructure down to 10^7 solar masses, exactly (the scale) where cold, warm, fuzzy, and self-interacting DM types differ.

9. The Neutrino Fog

Let’s talk about a new constraint in the search for DM that changes the character of direct detection going forward.

What is the fog? As detectors improve and get more sensitive, they eventually detect neutrinos from background sources - solar, atmospheric, diffuse supernova. These neutrinos scatter off nuclei through coherent elastic neutrino-nucleus scattering (CEvNS), producing recoils essentially identical to what we’d get from WIMPs. Boron-8 solar neutrinos, for example, mimic a WIMP of about 5 to 6 GeV, which is why the low-mass frontier hit this wall first.

(CEvNS itself has a nice history: predicted in 1974, it went unobserved for 43 years until the COHERENT experiment finally caught it just recently at Oak Ridge in 2017.)

The wall is here. LZ’s improved calibration enabled it to observe a solar neutrino signal that can mimic DM, marking its first glimpse of the neutrino fog and the strongest evidence yet of boron-8 neutrinos scattering off xenon through CEvNS. XENONnT and PandaX report the same. This was the collective message from the noble-liquid experiments at ICRC 2025 - they now operate in the fog, and future progress depends on extending sensitivity below it.

Fog, not floor. The old term “neutrino floor” implied a hard stop, and the terminology was deliberately changed because it was wrong. Above the fog, sensitivity “improves linearly” with exposure (for an essentially background-free experiment, which is what these detectors are engineered to be). Inside the fog, we’re discriminating against an irreducible neutrino background whose rate we know only to finite precision, so sensitivity improves roughly as exposure to the power 1/(2n), with n growing in the foggy regions. (Meaning, painfully slowly.) So progress can continue - it just costs a lot more per unit.

How we fight through the fog. This is mostly an instrumentation problem. Astronomers are using two methods to distinguish DM recoil from neutrino recoil, along with a supporting effort to better understand the background.

• Directionality. Solar neutrinos come from the Sun, whereas the DM “wind” blows from the direction of our motion through the halo toward the constellation Cygnus. A detector that reconstructs the direction of each recoil can distinguish between them. This is a strong approach because it works even when the two signals have identical energy spectra. That’s why we build low-pressure gas TPCs like CYGNUS, even if we have to accept tiny target masses.

• Annual modulation. The DM signal peaks in early June, when Earth’s orbital velocity adds to the Sun’s motion through the halo. The solar neutrino rate peaks in January, when Earth is nearest the Sun. The phases and amplitudes of the signals differ, so the two are separable (in principle). But the DM modulation is only a few percent of the rate, making it difficult to exploit.

• Knowing the background. A better understanding of the neutrino background improves characterization. It doesn’t separate the DM signal from the neutrino one, but it lets us subtract the neutrinos more confidently. To do this, dedicated CEvNS measurements are pinning down the interaction cross-section: COHERENT at a spallation source and, now, reactor experiments like CONUS+. Also, the solar boron-8 flux that sets the low-mass floor is being studied separately by solar-neutrino experiments and by DM detectors themselves as they start to see boron-8 directly.

So, that’s where the direct detection program currently stands. The era of cheap, order-of-magnitude gains appears to be over. Progress can and will continue, but it’s increasingly expensive and relies on clever methods to separate DM signals from neutrino backgrounds. Even if this is primarily an instrumentation challenge, it’s also a strategic funding problem, given the science-funding climate we have experienced since 2025.

10. Where This Goes

Here’s a subjective assessment of what the future of the DM search program might look like. Something we should recognize is that, even if we have had thirty years of null searches, that should barely move our credence downward that a DM-like gravitating component exists. That case rests on solid results from the CMB, BBN, structure formation, cluster gas fractions, colliding clusters, DM-poor galaxies, etc. The null results should really only change our credences in particular DM candidates and interaction models.

Having said that, what should be our credence that “a non-baryonic gravitating DM component exists?” I think it’s very high - maybe around 95% or more.

Let’s look at some other scenarios. These are all mutually exclusive.

The axion turns up: maybe 10-30%. This might be the most likely and clean discovery. We have independent motivation, rapidly improving technology, and actual coverage of the parameter space of interest. If this scenario panned out, it would solve strong CP and DM at once.

A WIMP-like signal at the edge of the fog: maybe 5-15%. Even if we manage to detect this, confirming it inside the neutrino fog would require proof of directionality and multi-target agreement. So even if this happens, it probably means we’ll get a decade or so of argument and debate!

A dark sector signature from astrophysics or the intensity frontier: maybe 15-35%. This wouldn’t be a particle detection per se, rather a coherent pattern suggestive of the dark sector’s existence. Something like velocity-dependent self-interaction across many lensing systems, a dark photon at Belle II, etc. This would mainly be confirmation that DM has internal structure. But that knowledge would also tell us where to point everything else. Given that payoff, this possibility might offer the highest expected value in the field.

Primordial black holes in the asteroid window: maybe 2-10%. This scenario is unlikely, but this is also a weird type of search. If DM is essentially comprised of PBHs, no particle detector will ever find it.

The Long Dark Grind

This may well be the most likely outcome over the next 15-20 years. We get no clean detection at all. XLZD continues pushing through the accessible WIMP space well into the 2040s. Axion searches cover more ground and find nothing. Astrophysical measurements continue to tighten constraints on the properties of DM, but its microscopic identity remains unknown.

Would this outcome count as failure? If the dark sector’s only coupling to us is gravitational, then it may be undetectable by anything we know how to build. Nothing in physics “requires” a hidden sector to talk to the visible one. But, even in this frustrating scenario, we would still know a great deal about DM: its abundance, clustering behavior, bounds on self-interaction, etc. This result might be disappointing, but it would not be a failure.

One last prediction. This one is structural and low-confidence. It’s quite possible the DM field will bifurcate. On one side, we would have a few extremely expensive flagships like XLZD, Argo, and CTAO continue to grind through the remaining accessible parameter space. On the other we would have a proliferation of smaller, clever experiments: quantum sensors, atomic clocks, levitated nanospheres, NMR setups, the intensity frontier, etc. It’s also possible that the second category produces more surprises, partly because the cost of probing genuinely new territory would be lower and partly because our theoretical guidance about where to look has not been very good so far.

The Bottom Line

So, where does that leave us? The gravitational evidence for a DM-like component appears to be overwhelming and redundant. What we are missing is the identification of what that component fundamentally is.

To summarize what we’ve covered:

• Something like DM was needed to build the universe we see. Growth after recombination gives baryonic fluctuations only so much time. Silk damping would have already erased the small-scale baryon seeds anyway. A component that could begin clustering before recombination must have supplied the gravitational wells into which ordinary matter later fell. If there were no such component, there’d be no galaxies, stars, or us.

• The strongest case for DM is cosmological. The evidence from the CMB, BBN, structure formation, BAO, cluster gas fractions, lensing, etc., converges to provide a solid case for something like CDM.

• Galaxy rotation curves are evidence for a mass discrepancy, but they are less decisive about whether that discrepancy specifically comes from a DM halo. That is also where modified gravity competes best.

• McGaugh’s 1999 no-CDM benchmark, developed in the MOND context, correctly anticipated the CMB’s first-to-second peak ratio better than the CDM models then in use. But the same benchmark predicted a third peak lower than the second. The observed third peak is much higher, just where a non-oscillating gravitating component would hold it up. Given that, a universe with no DM-like component at all seems very unlikely. Modern relativistic MOND completions get around this by introducing additional degrees of freedom that themselves gravitate, rather like dust.

• We know what the dominant DM component is not. BBN leaves far too few baryons. MACHO, EROS, and OGLE rule out dim compact objects (MACHOs) over broad mass ranges. Antimatter would have lit up the gamma-ray sky. Ordinary neutrinos are too hot and too light.

• Then, there is the separability evidence. In colliding clusters, tidal dwarfs and the NGC 1052 system, ordinary matter and the inferred gravitating component do not always stay together. DF2, DF4, and DF9 apparently lack the usual inferred DM component altogether. That behavior is very natural if there is an additional gravitating component that can be dynamically separated from baryons. A pure modification of the force law has a much harder time explaining that, although elaborate theories with extra gravitational degrees of freedom are not ruled out by this argument alone.

• Meanwhile, all the null searches over the years have done real damage to the old default picture. WIMPs have been squeezed by roughly eight orders of magnitude in direct-detection sensitivity, supersymmetry has not appeared at the LHC, invisible Higgs decays are tightly constrained, etc.

• The remaining space of possibilities is still enormous. Axions remain a possibility, with the caveat that the preferred mass depends on when Peccei-Quinn symmetry breaking occurred relative to inflation. The asteroid-mass primordial black hole window remains interesting. And beyond those sits the much larger dark-sector landscape: dark photons, atomic dark matter, self-interactions, etc.

• The most interesting current anomalies may already be pointing towards the existence of a dark sector. Overconcentrated lensing subhalos and excess cluster substructure lensing could point toward velocity-dependent self-interacting DM. Meanwhile, the S8 tension appears to be dissolving.

• Finally, direct detection is entering the neutrino-fog era. The searches are not over, but future gains will be harder, slower, and more expensive.

If we put all of that together, we end up in a strange epistemic position. Most of the matter in the universe appears to belong to a component we cannot see, touch, make, or catch, yet we can measure its cosmic abundance with remarkable precision. We know it had to cluster early. We know it is close to collisionless on large scales. And in some systems, we can map where the inferred gravitating component is, even when the ordinary matter has gone somewhere else. What we still cannot do is point to the fundamental microscopic thing underlying DM and say, "That’s what it is."

Thanks for reading! Subscribe for free to receive new posts and support my work.

If we ask a random person in the Western world what happens when they die, we usually get a fairly stable answer. Your soul, the real immortal you, separates from your body. It goes up to heaven to be with God, or down to hell to be punished. There’s a sorting that happens based on how you lived and what you believed. Good people go to heaven; bad people go to hell. And generally speaking, this happens more or less immediately at death.

Almost none of that is in the Hebrew Bible/Old Testament (HB/OT). And a surprising amount of it isn’t clearly in the New Testament (NT) either. Well, at least not in the shape that most people assume.

A composite picture built over time. The picture I just described above is a composite. The concept of the immortal soul that survives the body is largely Greek, originating with Plato. The resurrection of the dead is something else entirely (there’s an ongoing debate over how much the concept of resurrection owes to Persia). Heaven, as the final destination of the righteous dead, is a relatively late arrival that displaced older ideas in which all of the dead “just went down” to a shadowy underworld. Similarly, hell as a place of eternal conscious torment is a construction assembled across centuries out of several ingredients - a cursed ravine outside Jerusalem, Greek ideas about the underworld, a handful of apocalyptic visions, and a lot of later theological elaboration. Purgatory is an even later concept. And the specific architecture of these realms - levels, the fire, judgment at the moment of death, etc.? Well, that largely postdates the Bible entirely.

My goal in this essay is try and outline the “evolution” of these concepts over time.

Methodology. I’ve tried to approach this essay the same way I did my brief biography of the devil. I try to address certain assumptions that people have about these concepts - assumptions that the concepts have been stable over time and derive from scripture. Whereas, in actuality, they are really a stack of different ideas from different centuries that have been harmonized into a single picture over time. Multiple times, in fact. At least once, for example, when Judaism moved from ideas about Sheol to ideas about resurrection. And, at least once more, when Christianity fused ideas about resurrection with ideas about the Greek immortal soul.

I try to trace those developments in roughly chronological order. We look at:

• The older HB/OT, where we get a picture of the afterlife that’s quite bleak;

• The broader Ancient Near Eastern (ANE) world and its influence on those texts;

• The appearance of resurrection in the Second Temple crisis literature, along with some possible Persian influences;

• A collision (conversation?) with Hellenic ideas about the immortal soul;

• The messy situation that the NT inherits and doesn’t fully resolve;

• A brief look at what happened on the rabbinic side;

and then, for the second half of the essay,

• The Christian development of heaven, hell, and purgatory across the patristic and medieval periods,

ending with

• How modernity has hollowed out a good portion of these concepts.

1. Some Ground Rules

I am going to use the same analytical toolkit I used in my essay on the Satan - the standard historical-critical method. This is an academic approach that sets aside any theological premises.

A few rules specific to this particular essay:

• I treat afterlife beliefs as ideas with a datable history. Modern scholarship has demonstrated that the various texts in the HB/OT and NT were written at a particular time by particular people concerned about particular things. As such, those texts reflect the conceptual world of their composition. The fact that these texts later got bound into a single canon does not mean they share a single doctrine of the afterlife. They do not. For example, the HB/OT contains texts that deny the meaningful survival of death, and those texts coexist with other texts in the canon that affirm resurrection.

• I have tried to be careful about vocabulary, since translations sometimes sneak in theological concerns. For example, the Hebrew word nephesh is routinely rendered as “soul,” but it does not refer to an “immortal incorporeal substance”. It means something closer to “life” or “living being” or even “appetite.” (Genesis uses it for animals, for example.) Sheol, often translated “hell,” at least in older English Bibles, is not really “the Hell”. Rather, it’s the grave, the underworld, the pit that everyone goes to when they die. (I go over this in some detail in the essay.) Psyche in Greek and anima in Latin carry some Platonic baggage that’s not there in the original Hebrew. Gehenna, the word behind a lot of Jesus’s “hell” language, is a literal valley outside Jerusalem with a specific (and grisly) local history. Treating all of these as if they were early references to the Christian concepts of heaven and hell would be a category error.

Some things I do not consider (at all):

• Any apologetic stances that read the full Christian afterlife back into the whole canon, such that the patriarchs already believed in the Christian heaven, Job was really talking about the resurrection, etc. A historical-critical reading shows that the texts don’t support this. The effort to “make” them support it obscures the actually interesting story this essay is about - how the beliefs changed over time.

• The (deflationary) tactic, popular in some skeptical corners, that “it’s all just Egyptian” or “it’s all just Zoroastrian,” and Judaism and Christianity contributed nothing. Flattening everything into mere wholesale borrowing from one or more foreign sources, while denying the internal dynamics within Israel is just the perennialist error all over again. (In the guise of skepticism.)

Here’s a list of scholarly or summary works that I draw upon. I do this in a broad, general fashion, so you are not going to see precise citations to specific pages. If anyone is interested in discussing my use of these books and how they’ve influenced my thinking and the content of this essay, please send me a message.

• Alan Segal’s Life After Death: A History of the Afterlife in the Religions of the West (2004). A big, dense, and essential scholarly treatment. (One of the best books on this topic. But also, big, dense, and scholarly.)

• Bart Ehrman’s Heaven and Hell: A History of the Afterlife (2020). This is a very accessible one-volume overview and covers a good portion of the subject matter covered in this essay. I also occasionally refer to his more scholarly follow-up Journeys to Heaven and Hell (2022) for the later afterlife-tour material.

• Jon D. Levenson’s Resurrection and the Restoration of Israel (2008), which is an excellent account of how resurrection grows out of native Israelite ideas rather than being purely imported.

• For some of the older layers, Matthew Suriano’s A History of Death in the Hebrew Bible (2018), which is close to becoming a standard work these days, has reshaped how the field discusses Sheol, the dead, and related topics.

• N.T. Wright’s The Resurrection of the Son of God (2003). This is a long book, notable in this list of books for being confessionally Christian. And yet it’s still an excellent treatment of the resurrection material (even if I discount NTW’s theological conclusions).

• One of my favorite scholars, John J. Collins, on apocalypticism generally. (For example, his 1998 volume Apocalyptic Imagination. He also has excellent video discussions with Joel Baden on a variety of HB/OT topics on YouTube.)

• For a discussion of medieval developments about purgatory, I use Jacques Le Goff’s The Birth of Purgatory (1986), alongside Isabel Moreira’s Heaven’s Purge (2010), which revises and updates Le Goff’s thesis quite a bit.

• Caroline Walker Bynum’s The Resurrection of the Body in Western Christianity (1995) on how the physical-resurrection idea fared across time in the Christian context.

• There are a few more that I will note as I go through the essay, but the above list comprises my primary inputs.

A quick note about Segal. Segal’s organizing idea, if I may call it that, is that beliefs about the afterlife are a kind of “mirror”. Meaning, a society’s picture of what survives death “encodes” its picture of what a person fundamentally is. So, changes in the afterlife track changes in the concept of the self. I’ll utilize this notion a couple of times. Largely in places where I think it clarifies why a particular group might have reached for a particular concept of the afterlife. But I do want to note that - Segal’s idea is really an interpretive frame. I have otherwise tried to stay anchored in datable evidence rather than in any single organizing theory.

I’ll also try to flag disagreements between my sources. I do want to point out - I’m not a formal scholar of any of this, so if something’s off, assume it’s me garbling the words and ideas of people who know a lot better.

If any of you want to engage with these references, Ehrman’s Heaven and Hell is a good starting point. I disagree with some of his specific conclusions about annihilationism vs. eternal torment in the NT, but it’s a very accessible volume. Segal is probably where you’d want to go when you want (and are ready for) the real thing.

2. Sheol: Everybody Goes Down

For most of the Hebrew Bible/Old Testament (HB/OT), there is no heaven for the dead, no hell for the wicked, and no immortal soul that survives death. There’s just death and a shadowy destination called Sheol. Everyone ends up in Sheol regardless of how they lived.

Sheol is not a place of punishment, nor a place of reward. In fact, it’s generally portrayed as being barely a place at all. It’s the pit, the grave, the underworld. The HB/OT depicts Sheol as a dim region below the earth where the dead persist as faint, weak, shadow-versions of themselves, cut off from life and, worse, “cut off from God”. The dead in Sheol don’t really do anything. They don’t praise God (the texts are explicit and repeated about this - the dead are silent, God gets no worship from the pit). They also don’t experience much of anything.

The Psalms invoke this characterization of Sheol frequently: “For in death there is no remembrance of you, in Sheol who can give you praise?” The dead go down into silence. Down there, they cannot praise God. Ecclesiastes is bleaker and somewhat more philosophically direct - “but the dead know nothing; they have no more reward, and even the memory of them is lost. Their love and their hate and their envy have already perished; never again will they have any share in all that happens under the sun.”, and “Whatever your hand finds to do, do with your might, for there is no work or thought or knowledge or wisdom in Sheol, to which you are going”. This seems to place the dead in Sheol as being just barely above non-existence. Job, in one of his many despairing moments, describes death as the land of gloom and deep darkness from which no one returns, and treats the grave as a release from suffering “precisely because” nothing happens there.

This, I think we can all agree, is quite a different worldview from that held by most religious people in the Judeo-Christian world today. The reward for righteousness in most of the HB/OT is not heaven. Instead, the righteous get to live a good, long life, have many children, prosperity, and a peaceful death at a ripe old age, followed by being “gathered to their ancestors.” The punishment for wickedness, conversely, is a bad life and an early or dishonorable death. The whole moral economy that is depicted here operates “within” the period of life. There’s no afterlife to balance the books in. This is why the problem of the righteous who suffer is so acute in a book like Job. If there’s no afterlife where an innocent sufferer gets compensated, then any injustices have to be resolved “here in the regular world or not at all.” Job’s whole crisis is partly that it isn’t being resolved here.

Sheol is also an egalitarian leveler of the dead. Kings and paupers, righteous and wicked - all go to the same place and get the same treatment. There are occasional cases where a prophet or leader gets bodily raised to heaven, but this is exceptional (Moses, Elijah, Enoch are the usual examples we read about in the secondary literature). There’s also no sorting mechanism of any kind.

Early Israelite religion clearly also had some kind of “cult of the dead”, ancestor veneration, as well as necromancy. All of that happened in the background, since this was precisely the sort of thing the biblical writers were actively trying to suppress. The famous episode where Saul consults the medium at Endor, and she calls up the shade of Samuel should be thought of as the exception that “reveals the rule”. The necromantic practice is forbidden, but the story assumes that it works. Samuel comes up, and is annoyed at being disturbed, but is also recognizably himself. So the dead are down there in Sheol, and under the right circumstances, they can be reached. But also, the official religion very much wants you to stop trying to do that.

Maybe the dead weren’t simply nothing

I just described Sheol as being bleak and inert. This is how much of the older scholarship framed it. However, the best recent work on the topic complicates that a bit. Matthew Suriano’s A History of Death in the Hebrew Bible (2018) is slowly becoming a standard. (Here’s a short review of the book.) It comes at the topic of early afterlife beliefs in the HB/OT from the “archaeology up” rather than the “theology down”. Suriano bases a lot of this on how the Judahites actually buried their dead. They did so in rock-cut family tombs, with a primary burial followed by a “secondary” burial. In the secondary burial, bones were gathered and placed with the ancestors already in the tomb. If we read the HB/OT against what we know of the material practices of the time (gleaned from archeology), the biblical language of “being gathered to your ancestors” turns out to be more than just a euphemism for oblivion. It likely describes the cultural understanding of a real, ritually managed “transition.” Suriano’s term for what these activities and rituals secured is “functional immortality.” This is less like the survival of immortal souls and more like the persistence of the dead as remembered and venerated ancestors.

Francesca Stavrakopoulou’s work on ancestor veneration points in a similar direction. The broader consensus (surveyed recently by Kristine Garroway) is that Israel’s death culture was “similar but distinct” from its neighbors. It was organized around a cult of “dead kin” that the biblical writers only partially managed to excise from the canon.

None of this changes the fact that the older parts of the HB/OT formally have no heaven, no hell, no moral sorting, and no immortal soul. That’s still the case. But a position like “the dead are nothing, and Sheol is a meaningless void” is also inadequate as a characterization of what was happening in Israelite culture of the time. It was just the case that heaven was not the answer to mortality in this period. That role was served by a life lived well, the ancestral tomb, and the continuity of lineage.

3. The Neighbors: Egypt, Mesopotamia, and the Ancient Near Eastern Backdrop

To get a sense of how strange the Israelite position is, it helps to look at some of their neighbors. On this particular subject of the afterlife, Israel is the odd one out. The surrounding cultures had elaborate beliefs about the afterlife. Early Israel did not.

Egypt is probably the best counterexample. Egypt was the civilization most obsessed with the afterlife in the entire ancient world. The Egyptians built an enormous mortuary technology (mummification) to preserve the body. There were also elaborate tombs, grave goods, spells, etc. The Egyptian Book of the Dead (a collection of funerary texts) maps the soul’s journey through the underworld. Egypt also had a moral judgment of the dead. Most of you reading this will recognize the following scene: the heart of the deceased is weighed on a scale against the feather of Ma’at, the principle of truth and justice. Pass the weighing test, and the dead go on to a blessed afterlife in the Field of Reeds. Fail, and the monster Ammit devours the heart - leading to annihilation, the “second death”, and no afterlife at all. So, here we have a genuine post-mortem moral reckoning, a sorting of the dead by how they lived. Note that all of this is in place roughly a couple of millennia before anything like it shows up in Israel.

It might be tempting to draw a straight line from Egypt to the later Jewish and Christian judgment of the dead. After all, the Egyptian material is old, and it’s there to be absorbed or used by Egypt’s neighbors. And we know that Israel certainly knew Egypt. But any direct line of transmission is hard to demonstrate. The resurrection-based Jewish scheme that eventually develops in Israel looks structurally quite different from the Egyptian soul-journey.

A more interesting question might be the reverse one. Which is, not why Israel borrowed, but rather why it didn’t. After all, Israel had the Exodus story, placing its origins in Egypt, along with many centuries of contact. One (leading) line of explanation ties this to the same program that produced Israel’s other distinctive characteristics - the drive toward worship of “Yahweh alone.” Accompanied, of course, by strict admonitions against reverence of images and the dead. Anything like an elaborate cult of the dead, with tombs, offerings, and rituals addressed to ancestors, was exactly the kind of rival devotion that the Yahwist program worked to suppress. Stavrakopoulou does a good job synthesizing the scholarly work on this in the volume I mentioned earlier.

Mesopotamia is closer to the Israelite picture and probably more directly relevant. The Mesopotamian underworld - called Irkalla, or the “land of no return” - is grim, dusty, and dark. Irkalla is a place where the dead eat clay and dust and persist as miserable shades. If you think about it, that sounds a lot like Sheol. The Epic of Gilgamesh is a meditation on exactly that. Gilgamesh, terrified by the death of his friend Enkidu, goes in search of immortality. The whole point of the story is that he “doesn’t find it”. Death is the lot of humans. The one man granted eternal life (the flood hero Utnapishtim) is a unique exception. The message in Gilgamesh is similar to the one Ecclesiastes will deliver much later: you will die, so enjoy your life, eat and drink, and be with the people you love, because the underworld is dust.

The similarities between the Mesopotamian underworld and Sheol are close enough that most scholars see the Israelite picture as a regional variant of a shared ancient Ancient Near Eastern conception of the dead, rather than something unique to Israel. So what’s different about Israel isn’t Sheol itself. That distinction, as previously noted, goes to the increasingly strict insistence that “Yahweh alone is God”. Israel’s ideas about Sheol was a particular development, likely tied to its particular Yahwist theology.

4. The Hinge: Resurrection Appears Under Pressure

Somewhere in the Second Temple period, between roughly the 6th and 2nd centuries BCE, this bleak picture of Sheol starts to change. By the 2nd century BCE, we have Jewish texts confidently asserting that the dead (or at least some of them) will be raised, judged, and given either everlasting life or everlasting contempt. Something that isn’t in the older layers of the Bible is now unmistakably present in later ones. So, how did it get there?

The problem that resurrection solves

The old system, where righteousness is rewarded within life, works fine as long as the righteous actually prosper and the wicked actually suffer. It breaks down catastrophically when bad things happen to good people in spite of their goodness. Or worse, when bad things happen “because” of their goodness. And that’s the situation Judaism faced in the crisis period that produced the first (unambiguous) resurrection text.

Under the Seleucid king Antiochus IV Epiphanes, in the 160s BCE, there was a brutal persecution of Jews who refused to abandon the Torah. People were tortured and killed “for their faithfulness”. This is the Maccabean crisis, and it created an unbearable theological problem. If God is just, and the most righteous people are being murdered for their righteousness with no earthly reward, then either God is not just, or the reckoning must happen “somewhere other” than in this life. This is the age-old problem of theodicy, and the old answer to this problem (die in peace at a ripe old age surrounded by your children) must have seemed like a savage joke to a mother watching her seven sons tortured to death for merely keeping kosher.

The book of Daniel, whose final form dates to right around this crisis, contains what is essentially universally recognized as the clearest, and probably the earliest, unambiguous statement of individual resurrection in the HB/OT. In Daniel 12, in the context of all this persecution, we’re told that many who sleep in the dust of the earth will awake, some to everlasting life and some to shame and everlasting contempt, and that the wise will shine like the brightness of the sky. This is new conceptual material. The dead in the dust (the old Sheol vocabulary) are going to wake up, and they’re going to be sorted. The persecuted faithful will be vindicated, and their persecutors will face contempt. Note that resurrection here is framed as an instrument of justice, meant to balance the ledger left unbalanced during life, and not (primarily) as a comforting doctrine about going to a better, happier place.

The Second Book of Maccabees, covering the same crisis, makes it more concrete and also more physical. The martyrs being tortured explicitly declare that the “King of the universe” will raise them up to an everlasting renewal of life. And, quite pointedly, their tormentors will have no resurrection to life. One of the brothers being tortured sticks out his tongue and holds out his hands to be mutilated, saying he received them from Heaven and hopes to get them back again. So, the resurrection imagined here is not just spiritual, it’s a physical, bodily one. That physicality will stay central to the Jewish and then Christian mainstream for a long time.

The native roots

We discussed some pressure points above - these explain “why” the idea was wanted. But we have not discussed “where” the raw materials came from. Here, Levenson’s argument is important and quite poignant.

The older scholarly move was to say something like: resurrection appeared suddenly and must therefore have been imported wholesale from Persia. Levenson pushes back against this narrative. In his view, the materials for resurrection were sitting in Israelite religion all along - in the form of national imagery that was “ready to be literalized”. The prophets had long spoken of the Israelite nation dying and being reborn. The most famous instance is Ezekiel’s valley of dry bones. This is a vision in which the bones of the whole house of Israel are knit back together, given sinew and flesh and breath, and stand up alive. Those actual verses are quite moving to read. Of course, in its original context, this is a metaphor for national restoration from exile - the Israelite nation is dead or dying, God will revive it. But if you step back and look at this through the Levenson framing, those passages in Ezekiel really appear to be a metaphor built entirely from resurrection imagery. Hosea has similar language about God reviving and raising up the people. Isaiah has a passage about the dead living and the corpses rising that may be transitional between the metaphorical and the literal. Levenson’s point is that once Israel had that imagery in its scriptural vocabulary, the step to imagining “literal” individual resurrection is short.

The Persian question

Zoroastrianism, the religion of the Persian empire that ruled the Jews for two centuries after the exile, had a developed eschatology of exactly the kind Judaism eventually acquires. There is a final judgment, a resurrection of the dead, a renovation of the world, the righteous rewarded and the wicked punished, a cosmic movement toward a final good end, etc. The structural resemblance to what shows up in Daniel and other Jewish apocalyptic works is pretty close. The timing seems to line up as well. The new ideas appear during and after the period of Persian rule and cultural contact.

Given the resemblance, how much influence did Persia actually have on Israelite ideas about the afterlife? I covered some of this in my essay on Satan, but it’s worth addressing it here in the afterlife context.

The strong-influence position says that Persian influence is decisive. The Jews had no developed afterlife. They spent centuries under Zoroastrian rule. They emerged with a Zoroastrian-shaped eschatology. The resurrection, the judgment, the two destinies - all of these look like Persian imports. Mary Boyce, the great scholar of Zoroastrianism, argued for a version of this. Norman Cohn’s Cosmos, Chaos and the World to Come (2001) lays out a broader case that Jewish and then Christian apocalyptic eschatology owes its basic architecture to the Zoroastrian vision.

The weak-influence position, on the other hand, says - not so fast. We essentially have a major problem here - that of “dating”. (Note: this is also where a lot of contemporary scholarship has been moving.) The Zoroastrian eschatological texts that show the closest parallels (the Avesta in its fuller form, Pahlavi works like the Bundahishn, etc.) were written down very late. Likely in the Sassanian period and after, some as late as the 9th and 10th centuries CE. That’s a thousand-plus years after the supposed moment of influence. That makes it hard to prove that a fully developed Zoroastrian resurrection predated the Jewish one. (Maybe it was the reverse? Or perhaps the two developed in parallel?) The Jewish materials, as Levenson showed, have similar raw materials. And similar pressures (persecution, theodicy, the failure of the old within-life reward system, etc.) could (plausibly) independently produce similar solutions.

As I noted, this skeptical change in attitude has become common in the field. Bart Ehrman, for example, used to accept Persian influence on Jewish resurrection and apocalyptic. By the time he was writing Heaven and Hell (2020), he had changed his mind, calling the evidence for direct Zoroastrian influence thin.

What kind of afterlife is this anyway?

The afterlife that appears here is one involving resurrection, not the immortality of the soul. These are different, possibly even opposed, ideas. Confusing them is an error similar to confusing Sheol with hell.

Resurrection is bodily, future, and collective. It’s tied to a coming transformation of the whole world at the end of the age. If you die, you’re dead (really dead, in the ground). Then at the end, God raises you up, body and all, to face judgment and enter the renewed creation. There’s a gap there, where you’re really dead in the interim. But the payoff is a physical, embodied life in a restored world, not some disembodied spirit floating off to heaven.

Immortality of the soul is a very different kind of thing. The soul is a separable, naturally deathless part of you that survives the body automatically and continuously. Death is a release of the soul from the prison of the body. The real you then goes on without interruption. There’s no gap and no need for a future divine act. The soul never died in the first place.

Resurrection, at its root, Jewish and apocalyptic. Immortality of the soul is Greek and Platonic. Of course, neither of those worlds was uniform. Greek thought held shadowy Homeric shades, reincarnation, and flat materialism alongside Platonic soul-immortality. Nevertheless. much of the subsequent history of the Western notion of the afterlife is the (entangled) story of these two models.

5. Athens: The Immortal Soul Arrives

While Judaism was developing the idea of resurrection, the Greek world had, for centuries, been developing something entirely different.

The early Greek picture, in Homer, is a lot like Sheol. The dead go down to Hades, a gloomy underworld, where they persist as feeble, gibbering shades. Achilles’ ghost famously tells Odysseus he’d rather be a landless serf among the living than king over all the dead. The underworld is just that bleak. The dead are diminished, and being alive is better than any afterlife.

What changed the Greek picture was philosophy, and above all, Plato. In dialogues like the Phaedo, Plato has Socrates (on the day of his execution) arguing that the soul is immortal, that it preexists the body and survives it, that death is nothing to fear because it’s simply the liberation of the soul from the body. For the philosopher, the body is a distraction and a prison. Death frees the soul to contemplate the eternal truths directly. Plato also has judgment-of-the-dead style myths in which souls are rewarded or punished and then reincarnated according to how they lived.

The key contribution of the Platonic tradition, for our purposes here, is the idea of the soul as a “naturally immortal, separable substance.” The soul is the true self, temporarily housed in a body, released at death. This is really the intellectual pedigree of what most modern people mean by “soul.” It’s Greek, with strong philosophical roots. It’s also fundamentally at odds with the Hebrew concept of nephesh, which was never a separable immortal thing in the first place.

These two worlds (Israelite and Greek) collided in the Hellenistic period after Alexander, when Judaism was thoroughly exposed to Greek thought. We can “observe” this collision happening inside Jewish texts from the period. The Wisdom of Solomon, a Jewish text in Greek from around the turn of the era (the same text, incidentally, that first identified the Eden serpent with the devil), talks about the souls of the righteous as being in the hand of God. These souls are at peace, with an immortality that sounds thoroughly Greek. Philo of Alexandria, the great Hellenistic Jewish philosopher, had an essentially Platonic take on the soul.

So by the 1st century CE, Judaism contained at least two options that were not (obviously) compatible: the apocalyptic resurrection of the body at the end of the age and a Greek-styled immortality of the soul. There was also, of course, the old Sadducean position that held to the older biblical view and denied any afterlife at all.

(Note: I would like to briefly reference Alan Segal’s framing here. Josephus describes the Jewish “schools” as differing considerably on this point of the afterlife. The Pharisees affirmed resurrection, the Sadducees denied any survival of death, and the Essenes leaned toward the immortality of the soul. Alan Segal makes the point that these differences aren’t random - they broadly track “social location”. The Sadducees were the temple aristocracy, invested in the present order and its priestly hierarchy. Any type of future reckoning that would overturn that order held little appeal for them. The Pharisees were a more populist movement for whom resurrection promised vindication and reversal. The Essenes were a separatist community with the most otherworldly orientation of the three. Thus, afterlife beliefs at the time seemed to correlate with where a group sat in the social structure and what it wanted from the future.)

In any case, there was no single Jewish doctrine of the afterlife in the 1st century. In the middle of all of that, a new religion is then born out of the Jewish-Greco-Roman matrix: Christianity.

6. The New Testament Inherits a Mess

Something that’s not obvious to the average Christian: the NT does not hand down a single clear doctrine of the afterlife. It inherits the whole messy Second Temple situation, Resurrection, soul-language, apocalyptic judgment, and various images of the fate of the wicked, all of which co-exist in different parts of the 27 canonical NT books. Neither is there an attempt within the NT itself to fully systematize any of that. The univocality that many Christians take for granted from the NT is a harmonization that is not really there in the text. (Dan McLellan, the academic Biblical scholar who possibly has the strongest social media presence, speaks about this problematic presumption of univocality frequently in his videos.)

Jesus and the Kingdom

The historical Jesus, as best we can reconstruct him, stood in the apocalyptic-resurrection stream (as opposed to the “immortal soul” stream).

Jesus’ central message was the coming Kingdom of God. This was a future, imminent, transformation of the world in which God would set things right. The dead would be raised, and there would be a judgment. When Jesus argues with the Sadducees about resurrection, he takes the Pharisaic side (that is, the dead are raised). His framework is the apocalyptic one. The reward isn’t “your soul goes to heaven when you die.” It’s the coming Kingdom and resurrection “into it”. John the Baptist and the Apostle Paul seemed to have shared that apocalyptic vision.

(For further reading on this, I recommend John J Collins writing on apocalypticism in the Second Temple Period, which is excellent and thorough. Ehrman’s popularizations are probably more accessible for the average reader.)

The earliest layers of the Jesus tradition are really not about the modern “going to heaven” type of afterlife. It’s about the coming end of the age and the vindication of God’s people in a transformed world. More importantly, this transformation is expected “soon”.

This has been the dominant view in critical scholarship since Schweitzer, running through E.P. Sanders, Dale Allison, Paula Fredriksen, and (in his popular books) Ehrman. Jesus, as an apocalyptic prophet, preached about an afterlife that was collective, future, and bodily. (As opposed to individual, immediate, and spiritual). Like anything in the academic biblical field, this characterization isn’t unanimous, of course, but the apocalyptic reading remains the majority position, and I think it’s the better-supported one.

(The formal study of the “historical Jesus”, as opposed to the theological figure of Jesus, has been going on for over a century now. As noted above, the apocalyptic characterization is the most common one in the scholarly literature, but it’s not the only one. There’s also the proto-socialist Jesus, the quasi-feminist Jesus, Jesus who’s quite the magician, Jesus the revolutionary, Jesus the sapiential wisdom sage, Jesus the pious Galilean Jewish holy man etc.)

Now, let’s look at the Gehenna language. Jesus talks about Gehenna, which older Bibles translate as “hell,” more than anyone else in the NT. Gehenna is the Valley of Hinnom, a real ravine outside Jerusalem with a dark reputation. In earlier tradition, it was viewed as the site of child sacrifice to Moloch. Jeremiah pronounced an oracle over it, renaming it the Valley of Slaughter, so it became a byword for divine judgment and destruction. That’s also where its infernal reputation comes from.

(You’ve probably heard that Gehenna was Jerusalem’s perpetually-burning garbage dump, which is why Jesus used it as an image of hellfire. This is almost certainly false. There is no literary or archaeological evidence from the Second Temple or rabbinic periods of a burning trash dump in the Valley of Hinnom. The claim apparently traces back to a single medieval source - Rabbi David Kimhi’s commentary on Psalm 27 from around 1200 CE. This got copied and shared over time as if it were an established fact. It isn’t. The valley’s reputation comes from Moloch and Jeremiah, not from refuse. Lloyd Bailey’s 1986 article is a really good read on this.)

When Jesus warns about Gehenna, where “the fire isn’t quenched, and the worm doesn’t die”, he’s leveraging a vivid, concrete local image of judgment and destruction. Whether he meant eternal conscious torment or the destruction (annihilation) of the wicked is debated (see Ehrman’s Heaven and Hell, for example). The texts can be read either way. The “unquenchable fire” language can mean fire that fully consumes as easily as fire that torments forever. The “worm that doesn’t die” comes from Isaiah, where it’s feeding on corpses. Second Temple Judaism itself held both views. So the question can’t be settled by appeal to background alone. There’s a serious strand of scholarship (Kim Papaioannou’s The Geography of Hell in the Teaching of Jesus, Edward Fudge’s older work, Ehrman’s Heaven and Hell, etc.) arguing that the imagery leans toward annihilation rather than eternal torment. My own thinking on this is that, if we reject univocality in the NT, then it’s possible there are contradictory portrayals of Jesus’ views on this and other issues in the canon.

There’s also the parable of the rich man and Lazarus in Luke, which is the one Gospel passage that most clearly shows an “immediate” post-mortem sorting. In the parable, the poor man is carried to Abraham’s bosom, and the rich man ends up in torment in Hades. It’s a great parable that “seems to draw upon” wider Hellenistic and Egyptian folktale tradition about reversals of fortune after death (the Egyptian tale of Setne and Si-Osiris is the parallel most often cited). The key point to note is the immediate sorting after death.

Paul and the resurrection body

Paul is our earliest Christian writer, and he is very much in the resurrection camp. Still, Paul does something new with all of this. Because Paul has to deal with a novel theological problem - Christians are dying before the expected end, and the Kingdom hasn’t come. So what happens to them in the meantime?

Paul’s great treatment of this is in 1 Corinthians 15. Paul insists that resurrection is what happens. The Pauline formulation of this is as follows: Christ was raised bodily, and believers will be raised bodily. If the dead aren’t raised, then the whole faith is worthless. But the resurrection body, Paul says, is not the same as the current body. It’s transformed - sown perishable, raised imperishable; sown a natural body, raised a “spiritual body.” (That concept is beautiful but has also launched a thousand arguments. The majority reading of it has come to mean a body animated and transformed by the Spirit, not a disembodied ghost). Flesh and blood, Paul says, cannot inherit the Kingdom. Paul is essentially preaching a bodily resurrection into a transformed, glorified physical existence at the end when Christ returns.

This is thoroughly Jewish and apocalyptic. It also means that the original Christian hope was not “die and go to heaven.” It was “be raised from the dead when Christ returns to a transformed embodied life in a renewed creation.” N.T. Wright emphasizes this point throughout his book. Even if, like me, you don’t share Wright’s theology, the observation is sound. That is, the earliest Christian view of the afterlife is later resurrection, not going to heaven. In fact, the popular modern soul goes up to heaven picture likely reflects a later shift away from Paul’s actual view.

Revelation and the cosmic finale

The Book of Revelation gives the NT its most elaborate eschatology. There’s no doubt that Revelation is firmly apocalyptic. Along with parts of Daniel, Revelation is the most famous example of the entire apocalyptic genre. There’s a final battle, a first and second resurrection, an in-between millennium, a last judgment where the dead are judged according to their works out of the books, a lake of fire, etc. Classic elements of the apocalyptic genre. Revelation also includes, crucially, a “new heaven and a new earth,” with the New Jerusalem coming “down out of heaven to earth.” Note the direction there. The Christian consummation in Revelation is “heaven coming down,” and not souls escaping “up to heaven”. God dwells with humanity in a renewed material creation. The “lake of fire” also becomes a major source image for the later hell.

So if we review what the NT actually contains, we have the following: An apocalyptic “Kingdom of God” promise from Jesus. Imagery about Gehenna. A parable (Laz) with an immediate afterlife scene. Paul’s idea of a transformed resurrection body. Revelation’s cosmic renewal and lake of fire. Concepts about immortal souls that were drawn from the Hellenistic environment. The components of later Christian doctrine are mostly here: resurrection, judgment, differentiated destinies, paradise, punishment framed as eternal in some places and as destruction in others, heavenly ascent, a renewed creation, etc. What’s absent is their reconciliation. The pieces sit in the NT texts as discordant and unsystematized. So when I say the tidy modern system where your immortal soul goes at once to heaven or hell “isn’t in the NT”, I mean that many of the parts are there, but the assembled whole is not. That task of assembly is left to theology in the centuries to come.

But first, a quick look at what the other heir of Second Temple Judaism did with the same inheritance.

7. The Rabbinic Path (A Brief Detour)

I’ll keep this short, but it’s worth seeing that Judaism took the same raw materials and built something different. We also see that the Christian construction was a choice among options rather than the only possible outcome.

Rabbinic Judaism, emerging after the destruction of the Temple in 70 CE and developing through the Mishnah and Talmud, settled upon resurrection as a core doctrine. As we noted earlier, the Mishnah goes so far as to say that one who denies the resurrection of the dead has no share in the “World to Come”. Resurrection became a defining belief and even entered the daily liturgy, where God is praised as the one who revives the dead. So the Pharisaic side won decisively within Judaism.

But the rabbis also developed a rich (and characteristically underspecified) set of overlapping concepts that they never fully systematized. These include Olam Ha-Ba - the World to Come, Gan Eden - the garden/paradise for the righteous, Gehinnom - a place of punishment or purification for the wicked that most rabbinic opinion treated as temporary - a purgation lasting a limited time (the famous figure is twelve months) for most, rather than eternal torment. They also held, somewhat in tension, both the immortality of the soul (absorbed from the Hellenistic environment, echoed strongly in Philo and later thinkers) and the bodily resurrection. The way it works is that the soul survives death and is then reunited with a resurrected body at the end.

The features I’d like to point out, from a comparative angle, are as follows. Jewish thought largely kept punishment temporary and purgative rather than building an eternal hell. They also kept the whole subject relatively undogmatic and speculative. Meaning, there’s a lot of classic rabbinic “some say this, some say that” and remarkably little insistence on precise afterlife details. And, of course, they kept resurrection central.

So, the same Second Temple inheritance, when processed through the rabbinic tradition, produced a comparatively restrained, this-worldly-focused, “purgation, not eternal torment” version of an afterlife. Christianity, working with the same materials, went in a very different, far more elaborate and harsher direction.

8. The Soul Goes to Heaven: The Great Fusion

The single biggest transformation on the Christian side in this story is likely the one that fused the two incompatible models (Jewish resurrection and Greek soul immortality) into a single system. In the process, the center of gravity shifted away from future resurrection to the immediate fate of the soul. This development, which happens across the patristic period, produced the modern Christian concept of the afterlife.

One major factor driving this development was the same type of pressure Paul had already faced: the expected end didn’t come. And, as we might expect, this problem just got sharper over time. The apocalyptic expectation of an imminent return of Christ (a “very soon” resurrection and end of the age) faded as generations passed and the world kept going. The answer that emerged to address this issue was as follows: the soul separates from the body at death and goes at once to its provisional reward or punishment, to await the final resurrection and judgment at the end. So we get immediate soul destiny plus eventual bodily resurrection.

It was really the Greek philosophical apparatus that made this synthesis possible. The Church Fathers were educated in a thoroughly Platonized intellectual world. For them, the immortal, separable soul was simply the common sense educated view about human nature. Early Christian thinkers, especially the Alexandrians, Clement and Origen, who were steeped in Platonism, imported the idea of the immortal soul more or less wholesale and married it to the idea of biblical resurrection. Origen pushed furthest and speculated most freely about the soul’s journey and purification. But it was Augustine who gave the fused system its authoritative Western form: the soul is immortal, it faces its particular fate at death, and it will be reunited with the resurrected body at the last judgment.

This was a very consequential but also fairly quiet shift. Officially, Christianity never really dropped the idea of bodily resurrection. It’s still in the creeds, after all. ”I believe in the resurrection of the body”. This position was also defended fiercely. Caroline Walker Bynum’s book is about how tenaciously Western Christianity clung to the resurrection of the actual physical body across the medieval period, against every philosophical pressure to spiritualize it. But in practice, and increasingly in the popular imagination, the immediate destiny of the soul became the thing people cared about and thought about. The moment of death, not the end of the age (which could be far away or even nebulous), became the decisive moment. Heaven became a place souls go, rather than (primarily) some sort of future renewed earth we’ll be raised into. So, resurrection stayed in the doctrinal architecture, but got deemphasized.

As such, the modern person who believes their soul goes to heaven when they die is holding on to the Greek half of a synthesis whose Jewish half (future bodily resurrection into a renewed creation) has quietly faded into the background. Even though that Jewish half was the original Christian hope. This is one of the great ironies in the history of the idea. The thing most Christians now mean by “the afterlife” is closer to Plato than to Paul.

Something else that gets built up during this period (and after) is the architecture of heaven. Heaven grows from a relatively parsimonious “being with God” or “being with Christ” into an increasingly elaborated realm. For example, heaven attains the “beatific vision” (the direct sight of God) as its highest good in developed theology and, eventually, in the popular and artistic imagination. It also picks up the whole apparatus of clouds, heavenly gates, reunion with loved ones, etc. McDannell and Lang’s Heaven: A History is a great account of how the content of heaven has been reimagined over the centuries. This didn’t always happen uniformly. Sometimes heaven is characterized as God-centered and contemplative. At other times, heaven is more human-centered and social - a place of reunion, love, and even activity. So, heaven, like everything else in this story, also has a history.

9. Building Hell

Hell. The very word conjures up images of a place of eternal, conscious torment, with fire, presided over by devils with pitchforks, awaiting the wicked immediately at death. This, however, is very much a construction. The NT materials, as we saw, are ambiguous about whether they refer to torment or destruction/annihilation. The systematic doctrine of hell was developed after the Apostolic Age, in the patristic and medieval periods (that is, 2nd century CE and later).

The choice of eternal torment over annihilation. This specific choice wasn’t inevitable and was contested. Some early Christians held forms of “conditionalism” (the wicked are “ultimately” destroyed, not tormented forever). Some, following Origen, subscribed to “universalism.” That’s the doctrine of apokatastasis - the eventual restoration of all things, in which even the damned (and possibly even the devil) are ultimately purified and saved. Origen’s universalism was influential, but was later anathematized in the mid-6th century. Once conditionalism, universalism, and (largely) annihilationism were pushed to the margins, eternal conscious torment became the Western default. Augustine defended it forcefully in City of God, arguing against the “merciful” who wanted to soften or empty hell. Augustine insisted that the fire and the punishment were everlasting and literal enough, and his authority largely settled the Western question for a millennium.

The mapping of the underworld. I traced some of this development when discussing Satan’s “territory” in my other essay. Similar to heaven, hell got architecturally elaborated through a long tradition of vision literature and “apocalyptic tours”. The Apocalypse of Peter (2nd century) is an early and lurid one, matching specific torments to specific sins in a way that will become standard. The Apocalypse of Paul also supplied this widely copied tour model. Comparable visions in the medieval period (Drythelm, Tundale, etc.) built out the gates, divisions, rivers of fire, and sin-matched punishments. Ehrman’s Journeys to Heaven and Hell (2022) is a recent scholarly treatment of exactly this tour genre and how it manufactured the geography of hell. So, by the medieval period, hell had a structure and geography, a staff of tormenting demons, and a precise correspondence between specific crimes and matching eternal penalties.

Dante is one of the key synthesis points here, exactly as he was for Satan. The Inferno takes the whole accumulated tradition (moralized torments, the sin-by-sin structure, the geography, etc.) and fixes it into images so complete and so compelling that they effectively became the Western imagination’s hell. The nine circles, the punishments fitted to the sins (the contrapasso), the whole descending funnel of the damned - Dante came up with all of that. And as I mentioned in my essay on Satan, after Dante, hell is largely “illustrated and interpreted” rather than constructed.

So the modern hell is: the eternal-torment option (which, as we noted, beat out annihilation and universalism in the patristic controversies) plus centuries of vision literature mapping it out, plus Dante’s literary synthesis, plus, later, the fire-and-brimstone preaching tradition that kept it vivid. The Gehenna of Jesus became, over a thousand years, a fully architected realm of everlasting conscious punishment. Quite a long way from a controversial ravine outside Jerusalem.

10. Purgatory: The Middle Place

(If you don’t want to read about purgatory, feel free to skip this section. It turned out to be somewhat longer in length than I had originally planned.)

Purgatory deserves its own short section because it’s the clearest single case of an afterlife doctrine whose development we can actually “watch” (in retrospect).

The most famous modern account is from Jacques Le Goff in The Birth of Purgatory (1981), which argued (somewhat provocatively and brilliantly) that Purgatory as a “place”, a distinct third location in the afterlife, was essentially a high-medieval creation that emerged around the 12th century. Le Goff relies quite a bit on the term’s historical usage. The noun “purgatorium” doesn’t appear until the late 12th century, and Le Goff treats the coining of the noun as roughly the birth of the thing. Before that, in Le Goff’s view, you have “talk of purgation”, purifying fire and prayers for the dead, etc., but not the place itself.

The most important corrective to Le Goff’s narrative is Isabel Moreira’s Heaven’s Purge: Purgatory in Late Antiquity (Oxford, 2010). Moreira pushes the real roots of purgatorial thinking back centuries earlier, into late antiquity and the early Middle Ages. Moreira’s criticism of Le Goff is essentially that he mistook the naming of a concept for its existence. In her view, the concept of a purgatory (a real, intermediate, purgative post-mortem state that the living could affect) was up and running well before anyone coined the word purgatorium for it. (Peter Brown’s work on a 7th-century shift in attitudes toward the dead makes similar points.) So post-Le Goff scholarly views preserve his genuine insight that Purgatory was sharply defined, mapped, and systematized in the high Middle Ages, while rejecting his claim that the concept was “invented around 1200 CE.”

If we look at the source material for the concept, some of it’s quite old. For example, 2 Maccabees has Judas Maccabeus making atonement for dead soldiers. Later theologians even cited this as proof that the dead can be helped. The idea of a purifying fire that cleanses lesser sins goes back to hints in Paul (a passage about being saved “as through fire”). This is further developed by Augustine and, influentially, by Gregory the Great, who wrote about a cleansing fire for minor faults. So the raw materials (prayers for the dead, purgation, a fire that purifies rather than merely punishes, and even proto-narratives of a middle place) were part of the tradition well before the high Middle Ages.

What happened in the 12th and 13th centuries was that these disparate elements got consolidated and sharpened into a definite third place in the afterlife picture, along with its own logic and its own name. Between heaven (for the perfected) and hell (for the damned), Purgatory received the (vast?) middle. Thus, Purgatory was for the ordinary saved, who died in grace but with unpaid debts of sin and thus “must be purged” before entering heaven. Crucially, the living could help them through prayers, Masses, and eventually indulgences. Indulgences are where the doctrine became enormously socially and economically significant, and eventually a flashpoint of the Reformation. In any case, the Councils of Lyon (1274), Florence, and finally Trent, formalized Purgatory as Catholic doctrine.

This section illustrates an important theme of this essay and my prior one about Satan. Many of the concepts we’ve gone over have developed over time, usually building on ancient traditions and practices. But, when we can name the moment a scattered set of materials gets consolidated into a defined, named, institutionally-backed structure, perhaps the use of the word “invented” is not inappropriate? This becomes especially applicable to the concept of Purgatory.

In fact, the Reformation rejected the concept of Purgatory wholesale. Luther and Calvin threw it out, precisely because they judged it a construction resting on thin scriptural support and heavy institutional interest. The Protestant afterlife switched back to a stark heaven-or-hell binary.

11. Modernity: The Slow Hollowing

The last part of this story is what happens to all of this under the pressure of the modern period. What we see here is less construction or additions and more of a long, slow erosion, softening, and redecoration.

The Reformation, as we noted in the last section, eliminated Purgatory for Protestants and reemphasized resurrection. (In practice, Protestant popular belief drifted toward the same “soul goes to heaven” picture as everyone else.) The Enlightenment and the scientific revolution made the cosmology underlying the old afterlife hard to hold literally. For those who accepted modern viewpoints, heaven wasn’t obviously “up there" once up there was understood to be outer space. Hell wasn’t obviously “down there” once down there was best explained by geology. For many educated believers, the physical locations for heaven and hell dematerialized into spiritual locations, states, or even just metaphors.

Hell, in particular, underwent a striking softening across the 19th and 20th centuries. For a lot of modern Christians, the hard-hitting, fire-and-brimstone hell you’d find in most Puritan sermons became an embarrassment. Eternal conscious torment came to seem morally intolerable to many - disproportionate and incompatible with a loving and benevolent God. There was a real revival of the older, marginalized alternatives. Annihilationism (the wicked simply cease to exist) gained respectable defenders. Universalism (everyone eventually makes it to heaven) came back into play. Survey data on this is patchy and varies widely by country and phrasing, but the broad pattern is well attested. Belief in heaven has held up far better than belief in hell. A good many people in the modern West now affirm the first while quietly dropping the second. It’s a theologically lopsided position (heaven with no hell) that would have puzzled almost everyone before about 1800, but it does reflect real modern moral sensibilities.

Heaven too got increasingly humanized and sentimentalized. The modern popular heaven is heavily a place of reunion, of seeing loved ones again, of comfort, often with God and/or Jesus in the background. McDannell and Lang characterize this as the triumph of the “human-centered” heaven over the older “God-centered” one.

And what about the soul, that immortal separable substance that was a key part of the whole system since it was imported from Plato (and somewhat modernized for an enlightened audience by Descartes)? Well, the concept of the soul came under pressure from a different direction. Namely, from within. Specifically, from biblical scholars and theologians who noticed (correctly, as we’ve seen) that the immortal soul is Greek, not Hebrew. And that the original biblical hope was bodily resurrection, not soul-immortality. A whole modern theological movement emerged to “recover” the resurrection and demote soul immortality. This movement insists that Christianity’s real hope was always the transformation of embodied creation, not the escape of the soul from the body. N.T. Wright is a prominent, popular voice for this recovery. It’s a fascinating late twist, really - modern scholarship reaching back around two thousand years of Platonism to try to hand the tradition back its original, more Jewish, more physical hope.

The Bottomline

The afterlife most Westerners carry in their heads - immortal soul, immediate sorting, up to heaven or down to hell at the moment of death, etc. - is not in any single Biblical text and is also not as old as it feels. It’s not the view of most of the HB/OT, which had no such thing. It’s not the original hope of Jesus or Paul, who expected a bodily resurrection at the end of the age. Rather, it’s a composite of pieces laid down across more than two thousand years. That’s the history that this essay has tried to lay out.

None of what I’ve argued in this essay is an argument that the phenomenon of death isn’t profound, or that the human need to make sense of it isn’t real, or that these ideas haven’t carried enormous moral and imaginative weight. What the historical tracing offers, instead, is “clarity” about what kind of thing the afterlife is, as a concept. The afterlife is an evolved set of (competing) answers to the oldest and hardest human questions about life and death. The underlying concepts were built (and rebuilt) across cultures and centuries in response to specific pressure. Those pressures include persecution, the delay of the end times, the problem of the ordinary dead, the moral revolt against eternal torment, etc.

The traditional Jewish or Christian believer wants the afterlife to be a stable, revealed fact, faithfully recorded from the beginning and handed down intact. The historical record suggests that what we have instead is a stack of concepts whose layers can be assigned to specific historical moments. And we can do that assignment with reasonable confidence and reconstruct the development over time.

History can’t, of course, tell us whether there’s actually a heaven or a soul. The genealogy of a belief is not evidence for or against its truth. But what the genealogy can establish is profound enough on its own. The genealogy shows us that the beliefs themselves have no fixed, original, timeless form to be faithful to. There are only successive human constructions, each shaped by its moment, each displacing the last. These are “our” constructions, built and rebuilt to carry the meanings “we” needed them to carry. Whatever authority they hold is the authority “we” grant them, not something baked into the cosmos and merely transcribed.

That doesn’t prove there’s nothing on the other side of death. It does mean these particular traditions give us no independent reason to treat their afterlife claims as certain knowledge as opposed to what they demonstrably are: human answers to hard questions, built over time, out of the materials each age had on hand.

I’m going to try something a little unfashionable in this essay. I’m going to adopt a mostly neutral stance regarding the question of origin. I’m going to lay out what we know, walk through the evidence for the two leading hypotheses, while occasionally noting some rebuttals along the way.

I had spent quite a bit of time on the topic until about 2024 or so. My primary interests in molecular biology are mostly on the evolutionary side (bacteria, eukaryotic origins, etc.) not virology. Before writing this essay, I did try to catch up with some of the additional data that has come out over the past 2 years or so, as well as the most influential studies and official assessments on the topic. This includes the now well-known May 2025 Cell dating paper, the June 2025 WHO assessment, and the 2026 selection analysis.

This is hardly a systematic review, so I’ve focused on what I think is most important. I’ve also tried to be precise about which claims I think are reasonably solid, which are contested, and which got overstated in press coverage and then repeated until they hardened into “facts.” (There are a few of those, on both sides.)

I usually use inline links to references in my Substack essays. I do have some of this in this one, especially for definitions or simple references. But I also added a reference list at the end.

How to read this

I’ve tried to lay out each of the two leading hypotheses as a numbered list of key points, with relevant rebuttals/caveats attached to many of the points “inline.” Not every rebuttal is equally strong. Sometimes I think the rebuttals make a strong case against the point they are rebutting. In other cases, they only dent it. Also, don’t grade the two hypotheses by counting points. (I certainly don’t do that.) What matters is the “kind” of evidence each case rests on.

I did try to think through the epistemic issues involved here up front. In many cases, a fact supports one hypothesis over another only to the degree that it’s “more expected” under that hypothesis than under its rival. This has a consequence that is often missed: a fact can be dramatic, suspicious, and true, and still “carry little weight” if it’s roughly as expected under both stories. China’s secrecy is a good example of this - unsurprising whether the origin is natural or a lab accident. So the question I tried to keep asking was never just “is this true?” but also “how much more does one hypothesis predict it than the other?”

I’d do want to note one bit of orienting context. For a novel coronavirus such as SARS-CoV-2, zoonotic spillover is the base-rate expectation. For example, this is how SARS originally emerged in 2002 (via civets and raccoon dogs). This was also how MERS crossed over (via camels). There are many more such examples. (Quammen’s Spillover is a good and prescient book to read on this.) But we should also not be simplistic about those precedents - the reference class we pick matters. For example, if we consider the class of “all emerging respiratory viruses”, well, that results in a strong zoonotic prior. But if we consider the much narrower class of “novel coronaviruses that first surfaced in a city with a major sarbecovirus research program,” we end up with a much weaker one. The reference class for that latter scenario is nearly empty - the main class member is SAR-CoV-2.

Some Additional Reading Ideas

In addition to Quammen’s books on this topic, I recommend two popular books with opposing views that have received quite a bit of press (and accolades/criticism). Both are very readable.

Pro-natural-origin: Markolin’s “Lab Leak Fever: The COVID-19 origin theory that sabotaged science and society”

Pro-lab-leak: Ridley-and-Chan’s “Viral: The Search for the Origin of Covid-19”

The Two Stories in Brief

Before getting into the specifics, here’s a broad overview of the essential scenario each side is sketching. And please do read these as sketches, not established sequences.

The zoonotic scenario. In late 2019, a sarbecovirus that had been circulating in bats in southern China (or perhaps northern Laos) gets swept into the wildlife trade. The wildlife trade, in turn, involves a supply chain that moves live civets, raccoon dogs, and other sarbecovirus-susceptible mammals toward big-city markets. Somewhere along that chain, the virus “spills over” into a person (possibly more than once). Wuhan, a dense Chinese megacity with a major live-animal market (Huanan), is the epicenter of the initial outbreak, where the virus starts to spread rapidly among humans. In this narrative, SARS-CoV-2 is just another bat-coronavirus spillover. The third such spillover in just twenty years (SARS, MERS, COVID). An unlucky spillover with terrible global consequences, for sure, but not particularly mysterious or even suspicious. If anything, the spillover and outbreak are indictments of the wildlife trade and wet market practices.

The lab-leak scenario. Wuhan is also home to the Wuhan Institute of Virology (WIV), a major research center in China that has a documented history of collecting and experimenting on bat coronaviruses. In this narrative, a virus capable of infecting humans that was being studied at WIV (i.e., a wild strain WIV had collected, or perhaps one modified in the lab) escaped as a result of an accidental breach of biosafety. This resulted in an outbreak of the virus among humans. Early cases seemed to cluster near the Huanan market for the ordinary reasons that any early outbreak usually clusters “somewhere”. I’d like to note here that I do not intend to consider bioweapon or deliberate leak possibilities, as I think they are highly unlikely. What we are talking about here would have been just a lab accident of the kind that has happened before with other pathogens.

Comparing the scenarios The two scenarios I’ve laid out above agree on a good portion of the same set of facts. But they do diverge on a key question. Namely, was Wuhan the site of a natural spillover, or of a laboratory accident? My brief overviews of these scenarios are somewhat bare and linear. In reality, there are variants of each story. For example, the zoonotic scenarios doesn’t have to mean that the first spillover happened at the Huanan market. The market could well be the spillover site. Or maybe it served as an amplification site downstream of an earlier spillover. Or maybe, it was just one node in a long wildlife trade chain. “Laboratory-associated” is even broader in terms of possible variables. Many discussions tend to flatten most of that to “the lab built it.” For example, even reviews by intelligence agencies in the US often count all of the following as lab-associated: a researcher gets infected during field collection, a lab biosafety breach happens during handling of an “unmodified wild sample”, a serially passaged natural virus escapes from a lab, a genetically modified virus escapes from a lab, and so on. But those variations differ quite a bit in what’s required for the event to happen, not to mention the sort of traces the event would leave. Still, when I say “the lab-leak case,” I am referring to the whole family of variations on the theme. We just need to bear in mind that a fact that argues against the likelihood of an engineered-virus scenario often says little about the field-sampling-accident version.

A Very Rough Timeline: November 2019 to February 2020

Before we delve into the arguments for either scenario, let’s briefly look at the sequence of events related to the initial outbreak in Wuhan in late 2019 and the immediate response from the medical and scientific communities, as well as the press and authorities.

Somewhat problematically, many popular timelines that we encounter in online discussions are riddled with dates that sound authoritative, but on scrutiny trace back to a single unverified newspaper line. So, I’ve tried to separate what’s documented from what’s contested.

If any of you are interested in a more detailed timeline, here’s a more elaborate version from Congress that I reference occasionally in this essay. Here’s a more picturesque one from the WHO.

Let’s start with an interesting event that happened prior to our November 2019 starting point for this timeline. On 12 September 2019, the WIV took its main viral-sequence database offline. The WIV’s official position is that this action was a defensive response to hacking attempts targeting their databases. Lab-leak proponents view the whole thing as suspicious timing, considering that it happened around the time of initial outbreak.

November 2019 (contested?). The earliest claimed case seems to have been a 55-year-old from Hubei (with a reported onset of symptoms) around 17 November 2019. But we should note that this report comes from a South China Morning Post report citing unpublished government data. It’s never been published in any peer-reviewed dataset [23]. As such, we should really treat it as unverified. I included it since it’s a frequently cited “fact” in timelines from other writeups.

Moving on to firmer ground...

Early December 2019. There is a scientific report in The Lancet that the “earliest symptom onset” (among a cohort of 41 patients) was 1 December [24]. The WHO-China joint report has a date of 8 December for this [25]. Both of these dating attempts are retrospective. Honestly, I don’t think we really have a sense of “patient zero” for COVID-19. In fact, molecular-clock estimates place the spillover as early as mid-October [31].

Mid-to-late December 2019. By now, cases have begun to accumulate in Wuhan. Initial diagnoses have been independently made across several Wuhan hospitals between 18 and 29 December. Note that this was before anyone flagged the Huanan market as a common thread. Many of the earlier cases weren’t found by looking near the market. The earliest known cluster (a husband and wife, admitted by 26 December) had no known market link. A sick market worker was known to have been admitted by 27 December [26].

27 December 2019. Dr. Zhang Jixian, at Hubei Provincial Hospital, reports a cluster of four cases to health authorities. This is the first formal report up the chain [22].

30–31 December 2019. The Wuhan Municipal Health Commission issues internal “urgent notices” about the atypical pneumonia cluster. These notices leak online the very same day. (Physicians, including Li Wenliang, begin issuing warnings. Li is summoned by police for “making false comments” and “spreading rumors”). On 31 December, (a machine translation of) some Chinese reporting hits ProMED, the open disease-surveillance forum. The Health Commission issues its first public statement, reporting 27 cases. This date will become the one that WHO’s own timeline treats as the start of the outbreak [22][26][27].

1 January 2020. By now, the fact that cases have been clustering around the Huanan market has been recognized. The market is closed, cleaned, and disinfected. Environmental sampling begins that morning (a second round follows on 12 January) [2][26].

5–11 January 2020. Zhang Yongzhen’s group at Fudan University obtains a complete viral genome from an early patient sample (5 January). The group releases it publicly via the virological.org forum on 11 January. This is the very first SARS-CoV-2 genetic sequence made publicly available [28]. China’s CDC and other institutions (including the WIV) also submit sequences to GISAID over the following day or two. The question of who exactly was “first to share” eventually becomes its own internecine dispute (mostly academic/bureaucratic jockeying). In any case, the world has the virus’ genetic sequence by mid-January.

20–30 January 2020. On 20 January, Chinese authorities publicly confirm human-to-human transmission and infections among medical workers. This is the first pivotal acknowledgment. Wuhan is placed under a cordon on 23 January, the first lockdown of the pandemic. WHO declares a Public Health Emergency of International Concern on 30 January [22][27].

3 February 2020. The scientific publication Nature publishes a paper from Zhou et al. (from the WIV). This paper introduces RaTG13 into the discussion, a bat coronavirus that’s around 96% identical to SARS-CoV-2. It appears to be the closest known relative to SARS-CoV-2 [18]. The WIV paper presents RaTG13 as a newly reported virus.

11 February 2020 The virus gets a name: SARS-CoV-2. The disease gets assigned its own nomenclature: COVID-19.

Note: Even though the Huanan market shows up early (and centrally) in the case record, the earliest known cases are a “mix” of market-linked and unlinked. This ambiguity is something both natural origin and lab-leak hypotheses have tried to accommodate in their narratives.

Further note: A November 2020 addendum from the same WIV group that published details about RaTG13 clarified that RaTG13’s genome had mostly been sequenced back in 2018. RaTG13 is apparently just a “renamed” sample that had been collected in 2013 from a mine in Yunnan and previously catalogued as BtCoV/4991 [18]. Natural-origin supporters interpret this update as unremarkable - old bat virus sample dusted off once it turned out to be the closest match to a new one. No biggie. Proponents of the lab-leak theory, on the other hand, beg to differ. They view the later (Nov 2020) disclosure and renaming as part of a pattern of cover-up attempts from WIV that goes back (at the very least) to Sep 2019 with WIV taking their database online.

So, that’s the rough timeline. Let’s move on to the arguments for both sides.

The Natural-Origin Case (Zoonotic Spillover)

The zoonotic spillover hypothesis can seem sort of mundane, in the way most facts about infection disease emergence are mundane. In the SARS-CoV-2 case, we start with an virus in wild animals that gets moved around by the wildlife trade in China and Laos. Sometime in late 2019, the virus jumps to people at one or more locations where animals and crowds spend time in close proximity, the Huanan market being such a location. Stated that way, it seems like a common enough zoonosis scenario. But does the specific evidence we have for SARS-CoV-2 supports that traditional spillover narrative?

Let’s look at the key points in favor of the natural-origin scenario:

1. The earliest cases clustered around the Huanan market. Zoonosis proponents argue that, if you do an analysis of the earliest cases from December 2019, the cases are more tightly packed around the Huanan market than we would “otherwise predict” given Wuhan’s population distribution. A study by Worobey and colleagues makes that case in detail, reporting that the earliest known cases were geographically centered on the market. This includes cases with no “reported” market link [2]. The Worobey et.al. paper is a good one (along with the two rebuttal ones mentioned below) to read if you’re really interested in the early history of the outbreak.

Rebuttal / caveat. There was a widely circulated “4 to 6 km versus 16 km” figure which apparently came out of July 2022 press coverage of the paper, not the paper itself. The correct figures that should be used are the medians (use the erratum-corrected ones) - roughly 2.3 km for the two earliest lineage A cases and 8.3 km for the eleven lineage B cases, against an expected null (prediction) of about 16.1 km [2].

The clustering inference has been contested on statistical grounds. Stoyan and Chiu argue the whole “centroid marks the origin” assumption is unproven and that the underlying Monte Carlo test is flawed [6]. There was a separate critique that raised concerns about “proximity ascertainment bias” (meaning, early cases may have been found near the market because that’s where people looked!) [7].

Defenders of the clustering inference counter that the “unlinked” cases also centered on the market and that this is harder to dismiss as “ascertainment” [2].

Net result: there’s a residual methodological dispute here.

2. Two lineages = two introductions. Something that scientists determined fairly early on regarding SARS-CoV-2’s early genomic diversity is that there was a split into two genetic lineages, A and B. (Note: Lineage A is closer to ancestral bat coronaviruses like RaTG13, whereas Lineage B dominated early global spread. Both emerged during the initial outbreak in Wuhan.) Pekar and colleagues published a study arguing this is best explained by at least “two separate introductions into humans”, with both lineage progenitors circulating in animals before the jump to humans [3]. Now, under repeated animal-to-human spillover at an active wet market like Huanan, two introductions is not all that remarkable. It would be harder to explain that with a single lab-leak event (harder but not impossible, see rebuttal). Essentially, the point being made by the zoonosis side is about a genomic pattern distinction along with an interpretation of causality (namely the two introductions).

Rebuttal / caveat. The authors of the paper obtained this causal inference by “simulating” early epidemic trees. They asked - how often would a single introduction produce the observed A/B split with both lineages being present so early? The answer they found was that it rarely does.

Some things to note here. First, the original 2022 result was formally corrected. An October 2023 erratum from the authors fixed a coding error that had “overestimated the Bayes factors.” This dropped the level of support for the two separate introductions from “strong” to “moderate” [3]. Overall, the authors still favored multiple introductions.

Second, a critic (Michael Weissman - retired Urbana prof and one of my thoughtful mutuals on X) has argued that after a fuller correction, the balance actually tips toward a “single” introduction. (Michael’s Substack articles led me down a rabbit hole all by themselves.)

Third, the notion that two introductions is hard for a lab leak to explain is not really airtight. Two infected workers, or even one with a mixed sample, could (in theory) seed two lineages.

In 2025, when apparent intermediate (”T/T”) genomes surfaced in sequence analysis, a follow-up paper from some of the original authors concluded that the intermediates were probably “descendants of one lineage” rather than “true intermediates.” The authors saw that as confirming the existing evidence for separate introductions of lineages A and B into humans in late 2019. Their point essentially was - ruling out a true “intermediate” removes support for one lineage evolving from the other within humans. [32].

Net result: two introductions seems to be the better-supported reading. It fits repeated spillover more naturally than a single virus escape from the lab. But it could be also be regarded as a contested model-dependent inference based on sparse sampling as opposed to a demonstrated fact.

3. Wildlife DNA was found in the same spot as the virus. Zoonosis proponents note that environmental samples from one specific market stall that tested positive for SARS-CoV-2 also contained genetic material from wild mammals. This includes raccoon dogs, civets, and bamboo rats - all mammalian species capable of carrying the virus. In an analysis by Crits-Christoph et. al., wildlife DNA turned up in “all” of the SARS-CoV-2-positive samples from that stall (with viral and animal material together) [4].

Rebuttal / caveat. Technically, what this shows is co-location, not infection. So really, all the data establishes is that susceptible animals “were present” at the stalls that tested positive for SARS-CoV-2. It does not establish that any of those animals “were infected”. Also, Jesse Bloom reanalyzed the same data and found SARS-CoV-2 abundance correlated most strongly not with raccoon dogs but with various fish species. On the basis of that finding, Bloom argued that due to the co-mingling, we can’t tell if any animal was actually infected [8]. Note - this sort of thing is partly a limitation of metagenomic analysis in general.

Net results: even if, by sampling time, the key animals had largely been cleared, the data does put the right animals in the right place [4].

4. The susceptible animals were confirmed to be at the Huanan market. Live SARS-CoV-2-susceptible mammals, including raccoon dogs, were documented for sale at the Huanan market in late 2019. This can be ascertained from photographs as well as the environmental genetic record [2][4]. This fact addresses the early counterargument from lab-leak proponents that such animals weren’t sold there at all. They clearly were.

5. The pangolin receptor-binding domain. This is probably the key point against the “COVID virus looks engineered” intuition. (Even members of the infamous “proximal origins paper” group who are some of the strongest zoonosis advocates had some initial thoughts along the engineered-virus line. It was the pangolin virus similarities, among other details, that convinced them otherwise.) When the genetic sequence of SARS-CoV-2 was reviewed in the early days of the outbreak, scientists realized that certain receptor-binding residues in SARS-CoV-2 seemed unusual (as compared to the virus’ closest cousin, RaTG13, which differs at five of six corresponding residues). These residues determine how the virus’ spike grabs human ACE2, the prime cellular entry point for the virus. But then, some coronaviruses samples from Guangdong pangolins were found to match SARS-CoV-2 at all six key receptor-binding residues [1]. So SARS-CoV-2-like receptor features already existed in wild animal viruses. This demonstrates that these virus features have a natural analogue. Of course, this doesn’t trace SARS-CoV-2’s actual genetic route, which remains unknown.

Rebuttal / caveat. Not all that much to say, really. Some papers do refer to a five-residue set instead of six, and the counts end up looking contradictory. But, in any case, the general point (that these features exist naturally in the wild) seems robust.

6. Furin cleavage sites are not, by themselves, a fingerprint of engineering. A lot of the lab-leak discussion centers on the presence in SARS-CoV-2 of a “furin cleavage site.” This is a spike insert in the virus that makes it better at invading human cells. This is also the most famous of SARS-CoV-2 features that lab-leak proponents say clearly indicates lab modifications. But it turns out furin sites are widespread across other betacoronaviruses, just not sarbecoviruses. So a furin site feature as such isn’t “necessarily” a signature of lab-based tinkering [9]. Also, a furin-like S1/S2 site has even been reported in a bat coronavirus [10].

Rebuttal / caveat. This argument is weaker than it first sounds. The relevant comparison isn’t betacoronaviruses in general but sarbecoviruses specifically (SARS-CoV-2’s narrower subgroup). No sampled member of that subgroup carries a polybasic furin site. So “furin sites occur in nature” is true, but partly answers the wrong question. A narrower version of the natural-origin side’s claim is right (meaning, the feature isn’t intrinsically artificial, and natural mechanisms can generate it), but its absence in every known close relative is something that remains unaccounted for.

Net result: This point from the lab-leak side is a relevant one.

7. Feasibility is not evidence a lab modification actually happened. The natural-origin side argues that there is no evidence that WIV ever had an appropriate direct progenitor of the virus “in hand to be modified”. They do concede that that WIV had serious coronavirus expertise. In fact, the controversial DEFUSE project linked to WIV explicitly contemplated introducing furin cleavage sites into coronaviruses. And inserting furin sites into coronavirus spikes had been done experimentally before the pandemic. So, such an experiment was technically feasible for a capable lab like WIV. But, ultimately, feasibility isn’t “evidence it occurred.” No publicly identified backbone, construct, or experiment has actually ever been connected to SARS-CoV-2, so that’s a point in favor of a natural origin.

8. There are plausible natural routes to the odd CCG codons. One lab-leak argument concerns the sequence coding for the furin site (the so-called “CGG-CGG” business. I cover this in some more detail in the lab-leak section), which seems to point to lab modifications. The natural-origin counter: one hypothesis proposes the motif was picked up via recombination with human host mRNA [16]. And more directly, a similar furin motif in a “feline” coronavirus is encoded CGG-CGA, which is just one mutation away from CGG-CGG [33]. These are reasonably strong counters to the idea that the CGG-CGG sequence couldn’t have arisen naturally and so must be a lab signature.

9. The 2025 recombination-aware dating. In May 2025, a Cell study by Pekar and others took a new approach to an older controversy regarding the evolutionary history of SARS-CoV-2 [5]. Sarbecovirus genomes are “stitched” together by recombination, so different chunks have different evolutionary histories. If you treat the genome as one unit, the molecular-clock dating gets thrown off. So, the paper’s authors broke the genomes into “non-recombinant regions” and dated each one separately.

First, based on the revised dating, the closest bat-based virus ancestors of both SARS-CoV (2002) and SARS-CoV-2 are now inferred to have circulated only about one to six years before each virus emerged [5]. Not decades ago, as was originally thought. Thus, the lab-leak side’s “nearest known relatives of the virus diverged decades ago, that’s suspicious” framing turns out to be an artifact of ignoring recombination.

Second, the “geographic gap” issue. Those inferred ancestral viruses originally circulated more than a thousand kilometers from where the outbreaks started. How ever did they get to Wuhan? For SARS-CoV-2, the inferred ancestral range spans southwest China and northern Laos. Horseshoe bats disperse too slowly to cover that gap in a handful of years, so the paper’s authors argue the ancestor was carried the rest of the way by something faster than a bat - like wildlife trade. [5] Of course, it’s true that “faster than a bat” does not uniquely mean “wildlife trade.” But, what the wildlife trade has going for it is precedent (it’s how SARS-CoV-1 reached Guangzhou in 2002, for instance) and the fact that the wildlife trade did move susceptible animals into Wuhan’s markets.

Note: there is a separate unrelated question of whether the virus was shaped in a lab before it emerged. There’s a different study favoring the natural origin side that is relevant here. It sometimes gets confused with the 2025 recombination paper. This one was a 2026 Cell paper by Havens and colleagues. The authors tested for a change in the “selection regime” on the branch leading up to SARS-CoV-2. This is the kind of shift we’d expect from extensive laboratory passage or prolonged circulation in an intermediate host. The authors found no such shift. But they did detect exactly such a shift in the 1977 re-emergence of H1N1 (a virus widely thought but never confirmed to trace to a lab) as a positive control [29].

Rebuttal / caveat. The 2025 recombination paper is only a molecular-clock and phylogeographic reconstruction. The 2025 paper authors acknowledge sensitivity limitations in the paper itself. They note that further sampling, methodological improvements in disentangling complex recombination patterns, improved molecular clock calibration, etc. could allow for more detailed spatiotemporal inference. [5]

Human movement and even research-related collection are also faster than bat dispersal. The phylogeography by itself doesn’t single one method out. So the geographic gap is merely “compatible” with trade-mediated movement, and not uniquely so. As such, the overall claim is weaker than saying the paper “establishes” the wildlife route. Also, the paper doesn’t adjudicate against a lab pathway either, since research collection is also “faster” than a bat. Furthermore, no one has yet found a virus closely related to SARS-CoV-2 in any non-human mammal at the wildlife market or in its supply chain.

The Lab-Leak Case

Now let’s look at the other side. The lab-leak hypothesis has drawn a lot of lazy dismissal it doesn’t deserve (e.g., it’s a baseless conspiracy theory). But, it’s also had a lot of overheated advocacy (especially online) it can’t support.

The lab leak case is structured somewhat differently from the zoonotic spillover case. It doesn’t rest on direct positive evidence of a leak (there isn’t any). Instead, it’s based on a cluster of (circumstantial) facts. Taken together, its proponents argue, that cluster of facts is too much of a coincidence.

The conjunction argument. The strongest lab-leak proponents (e.g. Matt Ridley/Alina Chan) don’t stake their case on any single fact. They argue from the conjunction of multiple coincident ones. Let’s lay the skeleton of this out. A novel bat-origin sarbecovirus emerges in Wuhan, the one city that happens to host China’s leading bat-coronavirus lab. The virus emerges more than a thousand kilometers from its closest relatives’ wild range. There is no obvious progenitor virus that can be pointed to for a natural origin. It carries a furin cleavage site found in no sampled close relative. This emergence was just a year after an actual proposal (DEFUSE) to insert exactly such sites into exactly such viruses! That proposal named the WIV as a partner (even if the chimera-building part of work was slated mainly for the US side).

If you take all that together, it does seems fairly damning. Lab-leak proponents argue: WIV was doing gain-of-function research to explore how pathogen modification might result in increased risks of human infection and spread. Bad biosafety practices led to a lab leak of a modified virus that resulted in an outbreak somehow centered around the Huanan market.

Some points we should note…

First, a “conjunction” is only as improbable as its component facts are “independent.” But several of them aren’t. “Far from the bat range,” “furin site absent in close relatives,” and “no progenitor found” are related observations, not three unrelated coincidences. Sparse and geographically uneven sampling of wild sarbecoviruses feeds into more than one of them, as the 2025 recombination paper argues [5]. That’s not to say those facts are literally the same datum. But because a common cause “pushes” on several of them, we can’t simply multiply them as if they were independent. That would be a case of a cluster of correlated observations being treated as a set of separate improbabilities. On the other hand, this whole line of argument doesn’t dissolve the “whole” conjunction. Some of the facts really are independent of the sampling problem (e.g., the WIV’s location relative to the initial breakout, the DEFUSE proposal, the market case data, etc.)

Second, the “one city” coincidence is less improbable than it might appear. Wuhan is, after all, an 11-million-person megacity and a major hub of the central-China wildlife trade. This is exactly the kind of place a traded animal virus would plausibly surface. So what we really have is a a relative-probability question neither side can answer cleanly. Under zoonosis, how likely was a trade-driven outbreak to be “first detected” in Wuhan, rather than say, Guangzhou or Nanning? Under a lab origin, how much does a major sarbecovirus program at WIV that occasionally sourced samples from remote areas raise the odds for Wuhan specifically? These questions are not easily answered.

The WIV lab’s presence, of course, cannot be ignored. But the overall case may be weaker one than “the outbreak started right next to the lab” implies. Especially since the Huanan market, not the lab, is where the earliest cases cluster.

Let’s go over the individual pieces of the conjunction argument again.

1. The DEFUSE proposal. In March 2018, a grant proposal called DEFUSE proposed, among other things, introducing “human-specific cleavage sites” into SARS-related bat coronaviruses. That is, the proposal was for engineering the S1/S2 junction of the spike, the same region where SARS-CoV-2’s furin site sits. The proposal was submitted to DARPA under PREEMPT, involving EcoHealth Alliance, the WIV, and the University of North Carolina. DARPA declined to fund it, citing biosafety and gain-of-function concerns [14]. What makes this highly relevant (beyond guilt-by-association) is that it describes “in advance” an experimental plan targeting “the exact feature” that later made SARS-CoV-2 “stand out”. Whether the proposed work actually went ahead through some other non-DARPA funding stream is contested and unproven. But the fact is - the proposal existed, and it said what it said.

2. The furin cleavage site is absent in all known wild sarbecoviruses. No wild bat sarbecovirus sampled to date carries the polybasic furin site [1][17]. For a virus in this group to have that feature is “definitionally” unusual. Plus it’s the feature that makes SARS-CoV-2 stand out most from its relatives.

Rebuttal / caveat. This really is an argument from absence, which is only as strong as the completeness of the search. Sampling of wild sarbecoviruses is sparse, as we noted previously. The 2025 recombination analysis, for example, argues the true ancestors simply haven’t been sampled yet [5]. So “we haven’t found it in the wild” may be less problematic than it appears.

3. The CGG-CGG codon anomaly. Arginine can be coded by several three-letter codons. One particular codon, CGG, is relatively rare in SARS-CoV-2 and its relatives (~3 to 5% of arginine codons). The furin insert uses two in a row, CGG-CGG. CGG, moreover, is a codon that shows up a lot in human-optimized lab constructs. So - we have a rare codon, doubled, in the suspicious spot (furin cleavage), matching what labs tend to use in modifications [17].

Rebuttal / caveat. As Garry notes, codon rarity can’t confidently establish where such a doublet came from [9]. The disputed step in the whole debate is the treatment of the two adjacent codon positions as independent random draws. Whereas, as we discussed before, a single mutation from a CGG-CGA precursor (seen in a feline coronavirus) gets us to CGG-CGG [33]. So the specific “very low probability to happen naturally” cited by lab-leak advocates isn’t fully defensible under its own independence assumption. That said, the pair is indeed uncommon and the rebuttal does not settle how the coincidence arose.

Net result: The residual oddity survives. That is, even granting everything, tandem CGG-CGG is unusual for this virus. Something like “unusual and unexplained” is probably how we should consider this.

4. No identified intermediate host or progenitor virus. No animal virus that could plausibly be the direct ancestor of SARS-CoV-2 has ever been found. The closest known wild relatives, BANAL-52 from Laos (~96.8% identity) and RaTG13 from Yunnan (~96.1%), are still far too genetically distant to be direct progenitors [15][18]. RaTG13’s ~3.8% whole-genome difference represents (on a naive molecular clock) decades of divergence. This is a point often called the strongest evidence for a lab leak. But, a missing progenitor is roughly what we’d expect under “both” stories. Under zoonosis, it could be missing because wild-virus sampling is sparse and the ancestor may never have been caught. Under a lab origin, it could be missing because the relevant records of its existence are inaccessible due to stonewalling from the Chinese government. So a more appropriate framing might be - this is the single biggest “unresolved weakness in the zoonotic reconstruction.”

Rebuttal / caveat. Let’s add some relevant context here. The intermediate host for the original SARS virus took over a year to pin down. Wild-bat-virus sampling is limited, and the 2025 Cell analysis argues the relevant ancestors circulated recently but far from Wuhan and weren’t caught [5]. “We haven’t found it yet” and “there’s nothing to find because it came from a lab” are both consistent with the gap.

Net result: the gap keeps the question open rather than settling it.

5. The WIV was doing relevant coronavirus research. The WIV worked extensively on bat coronaviruses. A US House Select Subcommittee report (this was a congressional document produced under majority control, not an independent scientific review) built a circumstantial case around this. Among other things, the report pointed to possible coronavirus modifications and the September 2019 database takedown, right before the outbreak [19].

Rebuttal / caveat. The WIV’s stated reason for the database takedown was that it was in response to hacking attempts. The “gain-of-function” label on the specific NIH-funded work is (very controversially) disputed in terms of technical definitions. A deeper problem is that the takedown is (close to being) equally consistent with both stories. A lab with something to hide might pull the database, but so might an institution reacting to genuine hacking. Because both hypotheses predict it, it doesn’t effectively discriminate between the two options.

Net result: ambiguous.

6. China restricted the investigation. China limited access to samples, records, and independent investigators, and to date has not shared early-case sequences, detailed market-animal data, or WIV biosafety records with the WHO [20][21].

Rebuttal / caveat. If you really think about it, obstruction favors neither hypothesis. A government might restrict an investigation to hide a lab accident, or out of generic authoritarian reflex, national-pride defensiveness, or even unwillingness to be blamed for a natural outbreak on its soil. (Bad wildlife trade practices and all.) The “why would they hide it if they had nothing to hide?” doesn’t apply all that well to actions of the Chinese government in cases like this. This might just be standard operating practice for Chinese leadership. Afterall, they were accused of a cover-up during the 2002-2004 SARS epidemic.

Net result: ambiguous.

7. Geographic coincidence. The outbreak began in the very city hosting China’s leading bat-coronavirus institute, one campus of which sits about 12 km from the market. Lab-leak proponents find this hard to attribute this to mere chance.

Rebuttal / caveat. Two things. First, some WIV coronavirus work was documented at BSL-2 (a moderate biosafety level) - this fact is real. The stronger claim that the specific hypothesized gain-of-function type of work happened at BSL-2 is an inference. We don’t know one way or the other for sure.

Second, the 2025 recombination analysis offers a non-coincidental reason a bat virus might reach a distant city - the wildlife trade facilitates that type of thing routinely. A big city with a major wildlife market is the kind of place a traded animal virus might surface [5]. If we look at a map of western third of China, and consider only the largest cities (by population in the core urban areas) and proximity to the Guangdong/Laos corridor, and filter out only those cities with a strong retail presence for the wildlife trade, Wuhan pops up right away as one of the top cities on that list, if not the top city. As such, the coincidence is notable, but not “only” explicable by the lab-leak theory.

Net result: ambiguous.

Serial Passaging at WIV

There is a variant of the lab-leak theory that should be mentioned: a leak following serial passage of the virus. Serial passage involves repeatedly growing a field-collected virus in cell culture or humanized animals, letting selection adapt it over many generations. Serial passage is a routine technique used in labs. Interestingly, it leaves “no engineering signature,” meaning no restriction sites or synthetic “scars”, because the changes arise by selection rather than bioengineering. A virus shaped this way could look “natural” at the sequence level and still have been shaped in a lab.

The 2026 selection analysis by Havens et. al. took a close look at this and reported no signs of the prolonged adaptation serial passage would leave [29]. But it should be noted that, what that study tests for is a “change in selection regime” on the branch leading to SARS-CoV-2. That would amount to an intensification or relaxation relative to the virus’s bat relatives. The study found none (whereas 1977 H1N1, its lab-passage control, does show one). It should be noted that the study’s findings weigh most heavily against “extensive adaptive passage.” That’s the kind that leaves a sustained selective mark. It’s less of an argument against a brief handling episode that never ran long enough to shift the regime. (The authors do note that their test can miss weak signals from short lineages.)

So “no change-in-selection signal” was a solid study outcome, but the result is narrower than “no laboratory involvement of any kind.”

What the US Intelligence Community Said

The US intelligence picture has been persistently divided on the issue of a natural vs. lab origin for SARS-CoV-2.

In the 2023 declassified assessment, the National Intelligence Council and four agencies favored natural exposure. The Department of Energy and the FBI, on the other hand, favored a lab-associated incident. The CIA and one other couldn’t determine a either way. Almost all these judgments were reported at “low confidence” (the FBI’s, at moderate, in favor of a lab-leak scenario, was the outlier) [11].

In January 2025, the CIA released an assessment (ordered under Biden-era director Burns, declassified under Ratcliffe). The assessment judged, again at low confidence, that a lab-related origin is more likely than natural. It did call both scenarios plausible and noted that the judgement rested on no new intelligence [12]. Records released in 2026 (a January 2025 NIC memo obtained via a US Right to Know FOIA lawsuit) show officials still treating the question as unresolved and planning an independent expert panel. The memo did note that two agencies changed their confidence levels because of new reporting, details redacted [13].

On 18 June 2026, outgoing DNI Tulsi Gabbard released roughly 400 pages of declassified COVID-origins documents under an assertive banner (”Fauci Funded Wuhan Lab Research That Sparked COVID”). The headline item was a May 2020 Lawrence Livermore assessment that the “conditions” for an accidental release of a lab-modified coronavirus “were present” at the WIV in 2019 [30]. Note that this was not a new coordinated Intelligence Community assessment. It was a document release mandated by the 2026 Defense Authorization Act, framed by a departing official’s press statement. It was not a fresh confidence-rated analytic product agreed across agencies. Independent reviewers noted the released documents don’t establish the political claims made about them (one internal readout even undercuts them). “Conditions were present” should really be seen as a statement about capability, not a finding that a lab leak occurred. Notably, it doesn’t identify a precursor virus, a construct, or an incident.

So what does all of that suggest? The picture we get from the US IC suggests a genuine split, mostly low-confidence, perhaps shifting at the margins. In any event, it’s “not” the decisive tiebreaker partisans on either side present it as.

The WHO SAGO Assessment (June 2025)

The WHO’s Scientific Advisory Group for the Origins of Novel Pathogens (with 27 independent experts) published its multi-year assessment on 27 June 2025 [20][21]. The conclusion: the weight of available evidence points most strongly toward a natural, zoonotic origin. However, a laboratory-associated origin could “not be excluded”, largely because China did not provide the data needed to evaluate it (e.g., early-case sequences, market-animal records, WIV biosafety information).

The panel found no compelling" evidence of deliberate engineering. WHO Director-General Tedros framed the results this way: as things stand, all hypotheses must remain on the table, including zoonotic spillover and lab leak [20].

Unsatisfying perhaps, but also an honest position.

Where This Leaves Us

There is no decisive evidence for either hypothesis.

There is an asymmetry of evidence between the two hypotheses, but it isn’t a simple case of “direct beats circumstantial.” That’s not a real rule anyway - a direct observation can be non-discriminating, and a dated lab inventory with a near-match could be circumstantial yet devastating.

As I noted in the beginning of this essay, we do need to consider which observations are “more expected” under one hypothesis than the other. By that test, the zoonotic side has more positive, scenario-specific evidence - findings that a market spillover predicts and a lab origin mostly doesn’t. There is the geographic clustering of early cases, the two-lineage pattern, susceptible animals documented for sale, and a recombination-aware reconstruction compatible with wildlife-trade-mediated movement. There are also caveats attached to most of that: the clustering is a direct observation wrapped in a model-dependent inference, the two lineages are a corrected/contested reading, and the market animal DNA shows animal “presence, not infection”.

We should also note that many of those points draw on the same underlying source - the early-outbreak data collected, filtered, and released by Chinese authorities (case list, case locations, market samples, early sequences, etc.). It’s worth noting that, if we group the natural origin evidence by “data family” - early human-case ascertainment, early viral sequences, market environmental samples, wildlife-sale records, we see somewhat fewer truly separate pillars than the list of numbered points suggest.

The affirmative lab-leak case, once you set aside the things that are really shared gaps “rather than positive evidence”, rests on a smaller but cleaner set of points. There is the Wuhan research-program coincidence, DEFUSE and related research plans, documented biosafety concerns at WIV and relevant capabilities, and any other intelligence bears on pre-outbreak lab activity. If we set the clustering-statistics critiques [6][7] against these, it does seem that there are fewer affirmative, hypothesis-specific observations on the lab side. There is far more reliance on “context”. So, that’s a somewhat weaker evidentiary position than we have in the zoonotic scenario. But, on the other hand, “weaker” here means “less positive scenario-specific evidence,” not “circumstantial, therefore worth less.” The strongest thing the lab hypothesis has going for it (the sheer suggestiveness of the Wuhan coincidence plus the DEFUSE proposal) is exactly the kind of contextual evidence that a likelihood test treats seriously when the alternative doesn’t predict it as well.

Could new clues be discovered in the future? The reason this ongoing debate won’t resolve is as much due to “missing data” as to the strength of either case. “Missing data” isn’t symmetric between the two scenarios. What matters for each hypothesis is the evidence we’d “expect” to exist if it were true, and how recoverable that evidence would be.

What kind of new data might we expect for zoonotic spillover? Well, perhaps an infected animal, a close progenitor virus, a documented supplier-chain link to the market susceptible animals cleared before sampling, etc. For the lab-leak scenario? Mostly records from WIV - an inventory or genetic sequence record, an incident or biosafety report, staff serology, records of a matching experiment etc.

It should be noted that, since the zoonosis case is driven primarily by scientists who continue to publish papers on the topic, it has a higher change of producing new relevant insights that could affect the assessment.

There is also an asymmetry is in what’s “recoverable”. Under zoonosis, the decisive item (an infected animal) is the kind of thing that plausibly never existed in recoverable form, or existed briefly and was cleared. Under a lab-leak scenario, the decisive items (records, serology, sequences) plausibly “did” exist and could still exist, but sit behind a wall of non-cooperation.

The Bottomline

We don’t quite have a clean answer to the question of SAR-CoV-2 origins. We may never “ever” get a clean answer. The actions of the Chinese government may have negated the possibility of such a result.

Still, it’s a fascinating debate, and a largely scientific one. Most arguments, on both sides, are conducted in the language of phylogenetics, epidemiology, molecular biology, biosafety, etc.

The zoonotic hypothesis might be where the more developed positive evidence points, and where the WHO SAGO panel and several intelligence components lean. The lab-leak hypothesis is a credible alternative a serious person can hold, kept alive mostly by real gaps in the zoonosis hypothesis rather than positive proof in favor of a lab-leak.

This is a sensitive and politicized topic, and reasonable, well-informed people disagree about it. I’ve tried to represent both hypotheses (as best as I could) in their strongest honest form. I’ve also tried to flag where popular claims outrun the underlying data. If you think I’ve gotten a technical point wrong or been unfair to one side, I’d genuinely like to hear it.

References

[1] Andersen KG, Rambaut A, Lipkin WI, Holmes EC, Garry RF. “The proximal origin of SARS-CoV-2.” Nature Medicine 26:450–452 (2020). doi:10.1038/s41591-020-0820-9.

[2] Worobey M, et al. “The Huanan Seafood Wholesale Market in Wuhan was the early epicenter of the COVID-19 pandemic.” Science 377:951–959 (2022). doi:10.1126/science.abp8715. Erratum: Science, 15 March 2024, doi:10.1126/science.adp1133.

[3] Pekar JE, et al. “The molecular epidemiology of multiple zoonotic origins of SARS-CoV-2.” Science 377:960–966 (2022). doi:10.1126/science.abp8337.

[4] Crits-Christoph A, Levy JI, Pekar JE, et al. “Genetic tracing of market wildlife and viruses at the epicenter of the COVID-19 pandemic.” Cell 187(19):5468–5482 (2024). doi:10.1016/j.cell.2024.08.010.

[5] Pekar JE, Lytras S, Ghafari M, Magee AF, Parker E, Wang Y, Ji X, Havens JL, Katzourakis A, Vasylyeva TI, Suchard MA, Hughes AC, Hughes J, Rambaut A, Robertson DL, Dellicour S, Worobey M, Wertheim JO, Lemey P. “The recency and geographical origins of the bat viruses ancestral to SARS-CoV and SARS-CoV-2.” Cell 188(12):3167–3183.e18 (12 June 2025; Epub 7 May 2025). doi:10.1016/j.cell.2025.03.035. Preprint: bioRxiv 2023.07.12.548617, doi:10.1101/2023.07.12.548617.

[6] Stoyan D, Chiu SN. “Statistics did not prove that the Huanan Seafood Wholesale Market was the early epicentre of the COVID-19 pandemic.” Journal of the Royal Statistical Society Series A 187(3):710–719 (2024).

[7] Weissman MB. “Proximity ascertainment bias in early COVID case locations.” Journal of the Royal Statistical Society Series A: Statistics in Society 187(3):720–724 (2024). doi:10.1093/jrsssa/qnae021.

[8] Bloom JD. “Association between SARS-CoV-2 and metagenomic content of samples from the Huanan Seafood Market.” Virus Evolution 9(2):vead050 (2023). doi:10.1093/ve/vead050.

[9] Garry RF. “SARS-CoV-2 furin cleavage site was not engineered.” PNAS 119(40):e2211107119 (2022). doi:10.1073/pnas.2211107119. [See also the companion analysis by Chan YA, Zhan SH, “The Emergence of the Spike Furin Cleavage Site in SARS-CoV-2,” Mol. Biol. Evol. 39(1):msab327 (2022), doi:10.1093/molbev/msab327, on furin sites across coronaviruses and the feline-CoV CGG-CGA-to-CGG-CGG single-mutation point.]

[10] Zhu W, Huang Y, Gong J, et al. “A novel bat coronavirus with a polybasic furin-like cleavage site.” Virologica Sinica 38(3):344–350 (2023). doi:10.1016/j.virs.2023.04.009.

[11] Office of the Director of National Intelligence. “Updated Assessment on COVID-19 Origins” (declassified report, June 2023; based on information through August 2021). https://www.dni.gov/files/documents/FOIA/DF-2022-00031-Declassified_Assessment_on_COVID-19_Origins.pdf

[12] CIA assessment on COVID-19 origins, released 25 January 2025. Primary: the CIA’s own public statement, “CIA assesses with low confidence that a research-related origin of the COVID-19 pandemic is more likely than a natural origin based on the available body of reporting”; both scenarios called plausible, and the judgment explicitly not based on new intelligence. Secondary (context): CBS News and Associated Press reporting, 25–26 January 2025.

[13] National Intelligence Council, “Updated IC Assessments of COVID-19 Origins,” memorandum dated 17 January 2025 (14 pp.; released by ODNI in June 2026 in response to a US Right to Know FOIA lawsuit filed July 2025; substantial portions redacted).

[14] Primary source: EcoHealth Alliance, “DEFUSE: Defusing the Threat of Bat-borne Coronaviruses,” proposal to DARPA under the PREEMPT program (BAA HR00118S0017), March 2018; declined by DARPA. Leaked by DRASTIC (September 2021); additional drafts released via USRTK/USGS FOIA. Secondary source for provenance and context: The Intercept, “Leaked Grant Proposal Details High-Risk Coronavirus Research” (23 September 2021), https://theintercept.com/2021/09/23/coronavirus-research-grant-darpa/.

[15] Temmam S, et al. “Bat coronaviruses related to SARS-CoV-2 and infectious for human cells.” Nature 604:330–336 (2022). doi:10.1038/s41586-022-04532-4.

[16] Romeu AR. “Probable human origin of the SARS-CoV-2 polybasic furin cleavage motif.” BMC Genomic Data 24:71 (2023).

[17] Segreto R, Deigin Y. “The genetic structure of SARS-CoV-2 does not rule out a laboratory origin.” BioEssays 43(3):2000240 (2021). doi:10.1002/bies.202000240.

[18] Zhou P, et al. “A pneumonia outbreak associated with a new coronavirus of probable bat origin.” Nature 579:270–273 (2020). doi:10.1038/s41586-020-2012-7. Received 20 January, accepted 29 January, published 3 February 2020. Addendum: Nature 588:E6 (November 2020).

[19] US House Select Subcommittee on the Coronavirus Pandemic, final report and staff findings (2024). https://oversight.house.gov/report/final-report-select-subcommittee-on-the-coronavirus-pandemic/

[20] WHO. “WHO Scientific Advisory Group for the Origins of Novel Pathogens (SAGO) issues report on origins of COVID-19.” News release, 27 June 2025.

[21] WHO SAGO. “Independent assessment of the origins of SARS-CoV-2.” Full report to the WHO Director-General, 27 June 2025.

[22] Congressional Research Service. “COVID-19 and China: A Chronology of Events (December 2019–January 2020),” report R46354.

[23] South China Morning Post report (13 March 2020) citing unpublished Chinese government data, alleging a 17 November 2019 onset for a 55-year-old Hubei patient.

[24] Huang C, et al. “Clinical features of patients infected with 2019 novel coronavirus in Wuhan, China.” The Lancet 395:497–506 (2020). doi:10.1016/S0140-6736(20)30183-5.

[25] WHO. “WHO-convened Global Study of Origins of SARS-CoV-2: China Part” (joint WHO-China report, 2021).

[26] Worobey M. “Dissecting the early COVID-19 cases in Wuhan.” Science 374:1202–1204 (2021). doi:10.1126/science.abm4454.

[27] WHO. “Timeline: WHO’s COVID-19 response” / archived “WHO Timeline - COVID-19.”

[28] Contested-priority genome-sharing record: Zhang Y-Z / Holmes EC, virological.org post, 11 January 2020 (”Novel 2019 coronavirus genome”); genome obtained 5 January 2020; see also Science, “Dispute simmers over who first shared SARS-CoV-2’s genome” (2023) and The Lancet Microbe (2023) for the GISAID-vs-virological.org priority dispute.

[29] Havens JL, et al. “Dynamics of natural selection preceding human viral epidemics and pandemics.” Cell (2026). doi:10.1016/j.cell.2026.02.006.

[30] Office of the Director of National Intelligence, declassified COVID-19 origins document release, 18 June 2026 (DNI Tulsi Gabbard). https://www.odni.gov/index.php/newsroom/press-releases/press-releases-2026/4166-pr-11-26

[31] Pekar J, et al. “Timing the SARS-CoV-2 index case in Hubei province.” Science 372:412–417 (2021). doi:10.1126/science.abf8003.

[32] Pekar JE, et al. “Recently reported SARS-CoV-2 genomes suggested to be intermediate between the two early main lineages are instead likely derived.” Virus Evol 11(1):veaf008 (2025). doi:10.1093/ve/veaf008.

[33] Chan YA, Zhan SH. “The Emergence of the Spike Furin Cleavage Site in SARS-CoV-2.” Mol Biol Evol 39(1):msab327 (2022). doi:10.1093/molbev/msab327.

There’s a phrase the great physicist Frank Wilczek likes to use that I like: the “Core Theory.” That term is his, but what it refers to is not his pet idea but rather the working consensus of modern fundamental physics. This consensus is that a compact package of theories in modern physics describes everything ordinary matter and energy does. Well, essentially everything, as long as you stay away from a handful of exotic regimes.

The Core Theory should be familiar to those who’ve paid attention to modern physics. The first component of the theory is the Standard Model of particle physics. It’s an inventory of the fundamental constituents of matter and energy, as well as the interactions between them. We usually refer to those interactions as forces. There’s the electromagnetic force, the weak force, and the strong force, and the respective carriers of those forces (bosons). Then there’re all of the matter particles - fermions (quarks and leptons). And finally, we have the Higgs field that helps give massive particles some of their mass. The second component of the Core Theory is gravity, described by Einstein’s general relativity (GR). One point that I shall make in this essay, is that these are not really two separate stories, even though they are usually presented that way. They’re two halves of a description of the world we live in, that can be written in the same mathematical language of quantum field theory.

(One difference between this essay and other treatments of this material is that I am not going to describe gravity using classical general relativity, but what physicists call gravity as an “effective field theory.”)

This essay is a brief tour of what’s in the Core Theory, how far it reaches, where it breaks, where serious physicists push back, and how the higher-level world (chemistry, biology, a bacterium swimming in pond water, etc.) sits on top of it.

A few words about the word “consensus”. “Consensus” can paper over real disagreements, so I’d like to spell out the stance I take in this essay. Practically no professional physicist working in the area of fundamental physics disputes the following empirical claim: the Standard Model plus gravity accounts for every interaction of “ordinary” matter, to staggering precision. This is about as settled as anything in science. The pushback is usually on the framing. To keep this essay from becoming a short book, I’ll limit the discussion of the pushback to two camps. First, condensed-matter physicists who, in the tradition of Philip Anderson’s 1972 essay “More Is Different,” resist the idea that knowing the Core Theory means we “understand” the world at large. They grant, as most physicists do, that it’s all particles underneath. But they also insist that higher levels of organization genuinely have new organizing principles (the phenomenon of emergence). Second, many physicists are uneasy about including gravity in the package at all, given that we lack a “full” theory of quantum gravity. This is why I am using the effective-field-theory framing for GR in this piece. It lets gravity belong honestly in the quantum picture, as a low-energy approximation. So, to sum up, I guess we could say that the current consensus is strong on substance and somewhat looser on interpretation. But that is the normal condition of a mature science.

1. What’s Actually in the Theory (and How Well We’ve Tested It)

The Standard Model. The Standard Model accounts for three of the four forces that we know of. Electromagnetism, described by quantum electrodynamics (QED), governs light, chemistry, electronics, and every push and pull you feel that isn’t gravity. The strong force, described by quantum chromodynamics (QCD), binds quarks into protons and neutrons and then binds those into nuclei. The weak force governs radioactive beta-decay, and also sets the pace of the reactions that let the stars like the sun shine (the weak force “gates” how fast stars burn rather than supplying most of the energy, which is why the sun burns slowly and steadily). And, lastly, we have the Higgs field, switched on everywhere in space. The Higgs accounts for the mass of the three bosons that mediate the weak force, the W and Z bosons. It also accounts for a little bit of the mass of quarks and leptons. (As I discussed in a previous essay, for the protons and neutrons that make up the bulk of visible matter, this Higgs-given mass is only about 1%, with the strong force supplying the rest).

The empirical status of the Standard Model is excellent. QED is the most precisely tested theory in the history of science: the magnetic moment of the electron matches prediction to something like twelve digits. That’s the equivalent of measuring the distance from New York to Los Angeles and being off by less than the width of a human hair. The Standard Model predicted the W and Z bosons before they were found, predicted the top quark’s rough mass before it was found, and predicted the Higgs boson decades before it turned up at the Large Hadron Collider in 2012. There have been thousands of tests thrown at it, and it’s held up pretty well. It’s been poked at, stress-tested, and probed in collider after collider, and it keeps holding.

Gravity, as an effective field theory. In the usual framing, the gravity side of the Core Theory is supplied by Einstein’s greatest achievement - classical general relativity, which is a self-contained classical theory of curved spacetime. And like the Standard Model, GR’s empirical status is excellent. It predicted the bending of starlight, explained the precession of Mercury’s orbit, predicted the slowing of clocks in gravitational fields (the thing GPS quietly corrects for), predicted gravitational waves from colliding black holes (first observed directly in 2015), and the shadow of a black hole (imaged in 2019).

But we cannot treat GR as a “final” theory. GR as written is a “classical” theory, and the rest of the Core Theory is “quantum”. So a cleaner way to include gravity is as an “effective field theory” (EFT). An EFT is a quantum field theory that is “known” to be an approximation. Each EFT is valid below some energy scale, above which it no longer produces correct results.

We can think of EFTs like this. If we want to describe ocean waves, we don’t track every water molecule. We write down a theory of the wave’s height and length and let it ride, knowing full well that “down there” it’s really molecules, and that our wave equation will fail if you zoom in far enough. EFT does exactly that for gravity. It says: I’ll give you quantum gravity as a “series of corrections” to Einstein’s equations, each one smaller than the last at ordinary energies. GREFT, as it’s called, unravels near an enormous energy level called the Planck scale (around a quadrillion times the reach of our best particle collider). Below that scale, which includes “everything you will ever personally experience”, the theory makes perfectly good, testable, quantum-mechanical predictions about gravity. GREFT just declines to pretend it knows what happens at the highest energy levels.

How do EFTs work? What makes EFTs a legitimate scientific approach rather than some sort of dodge is a key piece of 20th-century physics called renormalization. The central discovery of that approach is just this: “the strength of a force is not a fixed number, it depends on how closely you look”. Let’s say we probe two electrons at everyday distances and measure the electromagnetic coupling. And then let’s say we probe them at (higher) collider energies, penetrating deeper into the cloud of vacuum activity surrounding each one. Turns out we end up measuring a noticeably larger coupling value. Physicists say the couplings “run” with energy scale. Usually, couplings run the same way - we get larger values at higher energies. Except for the strong force, which runs the other way, growing weaker at higher energies.

(Note: In modern physics, short distances go with high energies. Long distances go with low energies. Physicists often use those terms interchangeably - short distances also imply high energies.)

(Further note: In my opinion, the term “running of coupling constants” is one of the most terrible nomenclature choices for a physical phenomenon. Ugh.)

In any case, the running of coupling constants has a consequence that shapes everything. Physics effectively organizes itself into levels that are largely “independent" of each other. The details of whatever happens at unimaginably short distances “mostly wash out” by the time you get to larger atomic scales, leaving only “a handful” of numbers behind. A handful, as opposed to all of the specifics we’d need to fully characterize the behavior of the phenomena at much shorter length scales. This is why we can do chemistry without knowing about quarks. This is also why we can do “quantum gravity at ordinary energies” without knowing what replaces it at the Planck scale. Each layer is quasi-autonomous, sealed off from the noise beneath it. Effective field theories formalize this profound insight and are thus more than just convenient approximations.

People often say “we don’t have a quantum theory of gravity” or “gravity and quantum mechanics (or quantum field theory) are incompatible.” I have frequently pushed back against that framing over the years. That framing is true only in a particular sense. Sure, we don’t have a theory of quantum gravity that works “all the way down” at the Planck scale, where quantum spacetime presumably does something wild we can’t yet describe. (Savage qubit orgies, I imagine.) But in the entire energy range this essay is about, which covers aspects of reality ranging from stuff in our day-to-day lives to the sun’s core to a colliding pair of neutron stars, we absolutely do have a quantum theory of gravity. GREFT, the effective-field-theory version, treats the gravitational field as quantum, just like every other field in the Standard Model. It lets us compute quantum corrections to gravitational phenomena. It even predicts specific small effects that differ from regular gravity. Meaning, the leading quantum correction to Newton’s law between two masses is a calculated number. Now, those effects are far too small to measure, but the theory that produces them is legitimate, self-consistent, and predictive. So a much better framing than “gravity resists quantization” is “gravity quantizes just fine as an effective theory, only the deep, ultra-high-energy completion is missing.” For ordinary matter and energy, that missing piece at very high energies is like not knowing what’s inside a sealed box “way outside” the solar system. It’s real, but it’s also causally irrelevant to anything you’ll do today.

What does this mean? It means that the Core Theory is complete for a specific domain of applicability. And we can draw the boundaries of that domain with bright lines rather than relying on vague hand-waving.

2. What the Theory Is Really About

Before we talk about reach, it’s worth going over the “ontology” of the Core Theory. Ontology is a philosophical term that, in our context here, boils down to - what does the Core Theory say the world is “fundamentally made of”? This will take three short sections for me to go through. There’s a list of things that exist, the principle that organizes them, and finally, how it all assembles into our world.

Note: what the Core Theory pins down most securely isn’t a list of tiny objects, it’s “structure.” That’s essentially a web of symmetries and relations that can be written in terms of the fundamental constituents. So when I say the world “is made up of quantum fields,” treat that as the theory’s “most natural language”, not a claim that little continuous somethings (fields) are the “ultimate” furniture of reality. Some philosophers of physics (James Ladyman and the “structural realists”) go further and argue that the structure is what’s really real and the constituent objects are just a way of talking.

The world is made of quantum fields. The deepest insight of 20th-century modern physics, and the central ontological element in the Core Theory, is that the basic ingredients of reality are not particles. They’re quantum fields. A field, in this sense, is a physical quantity that has a value at every point in space and time. We already have an intuition for one: the magnetic field around a magnet has a strength and direction everywhere around it, filling the space. The Core Theory says - the whole world is basically like that.

The full inventory of fields is short and corresponds (roughly) to a field for each of the particles in the theory. We have six quark fields, each coming in three “colors,” which is eighteen. We have six lepton fields: the electron field, and those for the muon and tau, plus their three neutrinos. We have twelve force-carrier fields: the photon field, eight gluons, and the W+, W-, and Z bosons. And one Higgs field, which is a little unusual in that mix, in that it’s switched on with a nonzero value even in empty space.

The empirical status of these fields? Every one of those “has been detected”. Add gravity’s field, which is on a slightly different footing (we’ve measured gravitational waves, its classical ripples, but never an individual graviton, and quite possibly never will), and we complete the picture. That’s the component list for everything we have ever seen, touched, or been made of. Compact enough to write down on an index card.

Now, there’s a bit of an oddity buried in that list. Everything you’re made of uses just the first two quarks (up and down) and the electron. The rest are “heavier copies”: the muon is an electron that’s about 200 times heavier and otherwise identical, the tau heavier still, and the quarks likewise come in three escalating tiers. Three generations of the same pattern, differing only in mass, with the heavier ones decaying away almost immediately and playing “no role” in ordinary matter. We don’t quite know why the number of generations is three. It’s a research question that continues to be worked on.

Particles are quantized excitations of quantum fields. So where do particles come from? A particle is what we get when we add a discrete, minimal packet of energy to a quantum field, essentially a quantized “ripple in the field”. Poke the electron field hard enough in one spot and you excite one unit of it, and that unit “is” an electron. This is why all electrons are perfectly identical. They’re not tiny objects that happened to come out the same. Rather, they’re identical excitations of one and the same underlying quantum field, the way two small identical ripples on the same large pond are the same kind of thing. Electrons have no individual identity, no serial numbers, nothing distinguishing “this one” from “that one” beyond their states and relations. Swap two of them and the world is literally unchanged. This is the first real hint that the deepest layer is patterns and relations, not a population of distinct little things. A photon is a ripple in the electromagnetic field. A quark is a ripple in a quark field. And “quantum field theory” (QFT) is, at its core, the set of rules that these ripples come in whole-number units, one electron, two, three, never one-and-a-half, which is where the graininess of the quantum world comes from.

The rules underneath are quantum. I’ve been saying “quantum” this whole time without going into what that means. Some people think of QFT as a rival to quantum mechanics (QM). This is inaccurate. QFT “is” QM, applied to fields rather than particles, with special relativity folded in so the whole thing survives at high speeds.

Let’s actually start with that latter component - special relativity (SR). SR is Einstein’s second-greatest achievement, preceding his development of GR by around a decade. Its primary insight is that the speed of light is the same for everyone, no matter how they’re moving. And once we accept that, time and space stop being universal. Clocks run slow and lengths contract depending on our motion. Even “simultaneity” (the notion of “now” or “then”) becomes a matter of perspective. This raises the question: if motion, time, and distance are relative, how can we agree on anything? Well, in SR, what everyone “does” agree on is something called the (combined) “spacetime interval.” This is why we speak of spacetime as a single object rather than of space and time separately. SR also gives us E=mc², mass and energy being the same thing in different clothes. Why does SR matter here? Well, technically, it’s part of the Core Theory. Marrying SR to QM is also what forces the QFT picture on us in the first place. It’s what makes particle number changeable, so energy can turn into new particles and back. It’s why we need fields rather than a fixed roster of objects to describe reality. And, it’s what predicts antimatter. When Paul Dirac wrote down the first equation combining SR with QM for the electron in 1928, the mathematics insisted on extra solutions he couldn’t get rid of, describing something with the electron’s mass but the opposite charge. The positron, the antiparticle of the electron, was found in cosmic rays four years later. Nearly every particle turns out to have such an opposite partner (a few, like the photon, are their own antiparticle). When a particle meets its antiparticle, they annihilate into pure energy. So, without SR, we’d only have QM. With it, we get QFT.

Now let’s get into the basics of QM. Here are the three “quantum rules” that matter for everything that follows. First, what QM tracks isn’t a list of where things are and how fast they’re moving. What it tracks is a “quantum state”, described mathematically by wavefunctions, which live in an abstract mathematical space rather than in ordinary 3D. Quantum states (and their wavefunctions) hold everything there is to say about a system. Second, superposition: a state doesn’t have to be one thing or another, it can be a weighted blend of both. Moreover, those blends “interfere” like overlapping ripples, reinforcing here and canceling there. Third, uncertainty: at the quantum level, uncertainty is an essential characteristic of reality. Certain pairs of properties, position and momentum most famously, can’t both be “determined precisely” at once. That might sound like a limit on “what we can know.” But it’s more accurate to think about it as “what there is to know”.

Empty space isn’t empty. The uncertainty we talked about applies to quantum fields as well. The fields have to obey uncertainty even where nothing is in them, so they can never quite “hold still”. As a result, even when space is empty, the “vacuum” seethes with activity. Brief fluctuations flick in and out everywhere, all the time. And the Higgs field, as I mentioned earlier, is switched on throughout space with a nonzero value, so even the emptiest region you can imagine is soaked in it. Thus, empty space is less an absence than a (sort of) medium with properties. Those vacuum fluctuations aren’t just speculative. They have measurable consequences and are part of what the superbly predictive twelve-digit QED calculations compute. (Wilczek likes to call the vacuum “the Grid,” his own coinage, on the grounds that “empty space” badly undersells what’s there.)

Why don’t atoms collapse, and why don’t electrons spiral into the nucleus they are attracted to? The answer is uncertainty. Confining an electron to a tiny volume forces its (spread in) momentum to become wild, which “costs” energy. An atom’s size is sort of like a truce - between electromagnetic attraction pulling in and that quantum cost pushing out.

There’s a fourth rule worth adding here, because in addition to uncertainty, it helps explain why matter is solid. The Pauli exclusion principle says that two identical fermions (e.g., electrons) can’t occupy the same quantum state at once in any system. That’s why electrons in an atom stack into shells instead of all crowding into the lowest one. That is also a reason the periodic table has the shape it does and why chemistry is varied rather than monotonous. It’s also why you don’t sink through your chair. Pressing two objects together would force their electrons into shared states, and the exclusion principle refuses, pushing back hard. So, uncertainty gives atoms their size, and exclusion keeps them from passing through one another. The solidity of every object you have ever touched is not about stuff being packed tightly. Rather, there’s an underlying quantum bookkeeping mechanism that results in that solidity.

Finally, there’s quantum entanglement. If we let two quantum systems interact, they generally end up sharing a “single joint state” that can’t be split into “a state for this one” and “a state for that one.” Neither system has a state of its own anymore. The whole is genuinely not the sum of its parts. What exists is the relation between the two systems. Note that this is not a statement about our ignorance of the parts. The parts do not actually have separate states to be ignorant of. Nor is entanglement an exotic or fragile phenomenon. Rather, it’s the default. It happens whenever things touch, and it spreads.

3. Symmetry, the Organizing Principle

There’s one more layer underneath all of this. If you asked a working physicist what modern physics is “about”, a good number of them would say - “symmetry” (rather than particles or forces). A symmetry is just a change you can make that leaves something unchanged. For example, rotate a sphere and it looks the same. The physics version of this says the laws of physics themselves don’t care about certain changes. The symmetries in those laws mean that they work identically today and tomorrow, here and across the galaxy, whichever way you happen to be facing, etc.

That sounds almost too banal to be useful. But it turns out to be one of the deepest things we know about reality. Emmy Noether proved in 1918 that every “continuous symmetry of a physical law produces a conserved quantity.” For example, when the laws work the same irrespective of “when” something happens, that gives us “conservation of energy”. Not caring “where” something happens gives us conservation of momentum. Not caring in “which direction” something happens gives us conservation of angular momentum. So the great conservation laws aren’t “separate facts” about the world bolted on alongside the others. They’re just consequences of the laws’ indifference to aspects like where and when.

But this gets even better. It turns out that the forces we talked about earlier (electromagnetic, strong, weak) themselves are consequences of symmetries. For example, if we demand that the underlying theory stay unchanged under a certain (abstract, position-dependent) rotation of the electron field, electromagnetism falls out as a result. The photon “has” to exist, with exactly the properties it has, as a result of this symmetry. If we make the same demand with a different symmetry, we get the strong force and its eight gluons. Another gives us the weak force. This is what physicists mean by “gauge symmetry.” It’s also why the Standard Model is often written as a string of symmetry groups rather than a list of particles: SU(3) X SU(2) X U(1).

A symmetry the universe declines to respect. Everyone assumed for a long time that the laws of physics couldn’t tell “left from right.” Meaning a mirror-image universe would run by identical rules. This seemed as obvious as laws not caring which direction you face. In 1956, two theorists suggested nobody had actually checked this for the weak force. Turns out, the universe does care about this type of left vs. right symmetry. Radioactive cobalt nuclei, lined up in a magnetic field, spat electrons “preferentially in one direction” rather than symmetrically. The universe declines to fully respect this type of mirror symmetry.

The reason for this asymmetry is quite fascinating. It turns out that fermions (quarks, electrons, etc.) come in two “varieties”, called left-handed and right-handed, a kind of intrinsic twist physicists call “chirality”. Three of the forces - gravity, electromagnetism, and the strong force - are perfectly ambidextrous. They treat both varieties identically. The weak force does not. It couples “only” to the left-handed ones and ignores the right-handed ones entirely.

And this is why the Higgs has to exist. Because chirality exists, because left- and right-handed fermions “carry different weak charges”, we cannot" simply write down a mass for them in our equations. A mass term is precisely the sort of thing that “links” the left-handed version to the right-handed version. Doing that would be incompatible with the gauge symmetry that gives us the weak force in the first place. So in this naive version of our theory, every fermion is “required” to be massless. But wait - that’s obviously not the world we live in. We know for a fact that quarks and electrons have masses.

To solve this mismatch between our theories and the world as we know it, physicists added an explanatory “mechanism” in the theory to match reality. They postulated that there is a quantum field that fills all of space. Unlike other fields, this one has a nonzero value throughout the universe, and lets left and right link up “through” it. As a result, this field gives fundamental particles mass without breaking the symmetry outright. The postulated mechanism ended up being called, perhaps unfairly since it was an effort by many scientists, the Higgs mechanism. The field that was involved is called the Higgs field. (A Higgs boson is just an excitation of the Higgs field.)

The Higgs mechanism is an example of what physicists call “symmetry breaking.” In the first second since t=0 at the beginning of our universe, the Higgs field’s symmetry was spontaneously broken, resulting in a non-zero value for the field across space. This split the weak force from electromagnetism. The W and Z bosons became massive while the photon stayed massless. And a detection of the Higgs Boson in 2012 at the LHC confirmed these predictions made over 50 years prior. We should note - the Higgs isn’t just some optional extra “bolted on” to explain mass. Given chirality and gauge symmetry, something “like” the Higgs mechanism is necessary within the theory.

4. Putting It Together: Spacetime, Interactions, and Everyday Stuff

All righty. So far, we have fields, quantum rules, and symmetries. Three things remain before we can connect this picture to our world: where the fields “live”, how they push on each other, and how we get from that to the stuff of everyday experience.

Spacetime is the stage and gravity is its shape. Quantum fields need somewhere to “live”. That “stage” is spacetime, the 4D fabric of space and time. So, when we talk about a field having a value, that value is associated with a particular point in spacetime. I call spacetime the “stage,” but in GR spacetime itself is dynamical rather than a fixed backdrop. John Wheeler famously said - “spacetime tells matter how to move; matter tells spacetime how to curve.” That curvature is what we feel as gravity. So there’s no separate “gravitational force” pushing or pulling at anything. A planet orbiting the sun is just following the straightest available path through spacetime that has been been curved by the sun’s mass-energy. So, the “stage” image that suggests the notion of a container that exists independently of its contents? That is essentially the opposite of what GR teaches. A cleaner reading (the structuralist one again) is that spacetime isn’t really a container at all. More like a web of relations - specifically, the pattern of how events relate to one another.

Treating gravity as an effective theory changes the picture. This next bit might be the most abstract stretch in the essay, so bear with me. Everything I just said about GR so far is Einstein’s classical description of the theory. We have a smooth geometry and curvature as shape. The effective-field-theory version differs in subtle but important ways. We might be tempted to think of “GR as an EFT” as though it were only methodological caution. But it isn’t. It’s an ontological claim about “what’s there”. The metric in GREFT ceases to be a fixed geometric backdrop. It becomes a quantum field like others in the Standard Model. It has its own excitation (the graviton) and its own quantum state. So the geometry of spacetime itself can be uncertain or in superposition.

(Emergent spacetime? Below its cutoff energy, GR as an EFT isn’t some shabby approximation. It’s “correct”. But if spacetime is the right description only below a certain energy, then we might consider the possibility that spacetime is not fundamental at all. It may be emergent, relating to what’s underneath roughly in the same way that temperature relates to molecular motion. Temperature is entirely real at its own level, but it’s not part of the deeper vocabulary we use at the micro-level. Spacetime could be the same way. Real at the energy levels we encounter. But not quite “there” at the deepest levels. A good amount of current quantum gravity research leans this way. But also, it’s a research direction, not a finding. As such, nothing else in my essay depends on it. I mention it because it’s the natural destination of the EFT framing.)

Interactions are fields talking to each other. So with quantum fields, how does anything “happen”? Well, by quantum fields coupling. A ripple in one quantum field can create or absorb a ripple in another. When two electrons repel, what’s physically going on is that one electron (a ripple in the electron field) disturbs the electromagnetic field. That disturbance pushes on the other electron - which incidentally is also a ripple in the same original electron field. In the language of the QFT, we say the two electrons “exchange a photon.” Every force works this way, by exchanging a force-carrier particle. Electromagnetism uses photons, the strong force uses gluons, and the weak force uses W and Z particles. And gravity, in this language, uses gravitons (the quantized ripples of the spacetime field, which are perfectly real “in the effective theory” even though catching a single one is essentially impossible). So the four forces are really just four different types of “conversations between fields”. That’s how interaction works in the world.

(Note: I avoided saying “cause and effect,” since that phrase implies a direction. Kind of a “this makes that happen” type of thing. But the fundamental equations of the theory don’t contain that direction. They run the same forwards and backwards. Causation, like time’s arrow, is a higher-level pattern, not something intrinsic to the bottom layer of reality.)

Everyday stuff. Let’s pick an example of a macro object. But not too macro. Here’s how you get from the austere picture of the Core Theory to a bacterium.

Let’s start with a proton. It’s three quarks (ripples in the respective quark fields) bound together by a furious exchange of gluons (ripples in the gluon field). That’s the strong force in action. As I’ve written before, nearly all of the proton’s actual mass is the energy of that churning field inside it, not the quarks themselves. The same force also binds protons and neutrons into an atomic nucleus.

Now let’s add electrons. An electron (ripples in the electron field) is held to the nucleus by electromagnetism. This happens via photon exchange between the negatively charged electron and the positively charged nucleus. That electromagnetically bound system, governed by quantum mechanics, is what we call an atom. So the whole of the periodic table is electromagnetism arranging different numbers of electrons around different nuclei, with the exclusion principle dictating how they stack.

Atoms bond into molecules. This happens by atoms sharing and trading their outer electrons, which is again pure electromagnetism. Thus, chemistry is the electromagnetic force plus the quantum rules for how electrons are allowed to arrange themselves.

Molecules then stack into larger structures." We get proteins, membranes, crystals, cells, and here too the glue is electromagnetic. The organization just happens to involve ever more elaborate patterns (e.g., chemical bonds, forces that fold a protein, etc.).

(I’ll go over this build-up in some more detail in a later section.)

Notice the division of labor here. The strong force builds the nuclei and then “falls silent”, because its reach is too short to matter beyond the nucleus. The weak force only shows up for occasional radioactive decays and the reactions inside stars. Gravity is so feeble at these scales that it plays no role in chemistry at all. Gravity only shows up when we pile up enough stuff to create something planet-sized. Which means that essentially everything “from the atom on up”, all of chemistry, all of biology, every structure you can see and touch, is essentially the electromagnetic force acting on electrons, refereed by quantum mechanics.

5. Why, in the Solar System, This Is Basically the Whole Story

Here’s the part of the story I think should be more widely appreciated. At the energies and conditions found anywhere in the solar system, from the core of the sun to the chair you’re sitting in, the Core Theory accounts for “every interaction that can causally affect ordinary matter and energy”. Every one of them. There is no known phenomenon in your kitchen, your body, a distant asteroid, or the solar wind that requires anything outside this package to explain it.

When two things influence each other physically, one of the four forces is doing the work. A ball falls: gravity. Your hand doesn’t pass through the table: electromagnetism and the repulsion between electron clouds. A uranium atom decays: the weak force. The sun fuses hydrogen: the strong and weak forces together. There is no fifth channel or force that reaches ordinary matter at any level we’ve been able to detect. Physicists have looked, very hard, for a “fifth force” and for deviations in gravity at solar-system scales. Nada. Zilch. There, of course, remain unprobed windows at very short ranges or very weak couplings, but nothing that could reach out (or in) and nudge your coffee cup. Whatever else might be out there, it does not measurably touch the everyday world.

And yes, this holds even though the universe at large is dominated by dark matter and dark energy. Those two run the cosmos on galactic and cosmological scales. But their influence on anything inside the solar system is far too feeble to detect, let alone to matter. The completeness of the interactions claim is about “our” neighborhood.

This is why Wilczek is willing to say something that sounds arrogant but isn’t. That is - the physics underlying everyday life is “completely known”. Not “mostly.” The relevant particles, the relevant forces, and the relevant equations are all in hand. They’ve all been tested to punishing precision in exactly the energy range where we live.

6. The Gaps (Sorted by Whether They Reach Your Kitchen)

Now, “complete for a domain” is obviously not “100% complete.” Meaning, there are holes in the Core Theory when it comes to phenomena outside the domain of applicability. But I want to sort them into two buckets. Primarily because these two kinds of gaps get lumped together in a way that (badly) misleads people. Some gaps concern phenomena that “actually happen on Earth”. Others concern phenomena that only show up in high-energy particle colliders or in cosmic extremes, and which have no causal impact on daily life at all.

Gaps that touch things happening on Earth

Neutrino masses. We know neutrinos have mass, because they “oscillate.” Meanig they morph from one type (of neutrino) to another as they travel. The way the theory works, this would only be possible if they weigh something. And neutrinos stream through us by the trillions every second. These are mostly from the sun, so this is unambiguously happening here, now, in our bodies. But the Standard Model, as currently formulated, says that neutrinos should be massless. The reason for this traces back to the idea of chirality we discussed earlier. The theory’s inventory of fields/particles contains left-handed neutrinos but no right-handed ones. The Higgs mechanism that gives every other massive particle some percentage of its mass has nothing to link up to for neutrinos. So we have a real gap here rather than a loose end.

But, here’s why a gap involving particles that pass through your body by the trillions still “changes nothing” about daily life. Neutrinos “almost never interact” with anything. They carry no electric or color charge. Thus, they ignore electromagnetism and the strong force entirely. Only the weak force (barely) and gravity (negligibly) touch them at all. Roughly 65 billion solar neutrinos cross every sq. cm. of you each second, day and night (in fact, at night they’re coming straight up through the whole planet!). For a typical neutrino, the odds of interacting even once while crossing all 12,742 km of the Earth are about 1 in 10 billion. You will most likely live your whole life without a single neutrino ever “touching you”. In fact, this is actually a bit of a pain for running tests on neutrinos. Catching them at all takes heroic instrumentation. Ray Davis famously ran a 600-tonne tank of cleaning fluid for two decades to collect a few hundred events. Even modern neutrino detectors see roughly one interaction per ton every several hours. So while the origin of neutrino mass is a genuine hole in the theory, the neutrinos themselves have almost “no causal impact” on everyday chemistry, biology, or geology.

Gaps that only matter in particle colliders or cosmic extremes

Dark matter and dark energy. Together, these constitute the largest gaps in our mass-energy accounting of the cosmos. Roughly 85% of all matter in the universe is some unidentified substance the Core Theory doesn’t contain. Because this type of matter does not interact with electromagnetism, we call it “dark matter”. The remaining 15% is regular matter. On top of that, we know that the expansion of the universe is accelerating. This is best understood as driven by a “dark energy” intrinsic to spacetime that we understand even less about. Both phenomena are real, as far as we can tell. Both dominate the universe at the largest scales. But here’s the thing - neither has any relevance to solar-system-scale phenomena. The dark matter density near the Sun is minuscule, something like a few hundredths of a solar mass spread across an entire cubic parsec. So the total enclosed by Neptune’s orbit is a rounding error on a rounding error. People have looked for dark matter in the solar system, using high-precision planetary tracking. The results? We can’t “even detect” the expected amount. To see the effects of dark matter, we have to look at galactic or cosmic scales. So, whatever dark matter is, it barely interacts except through gravity. There’s no impact on chemistry, no heating, nothing local. Dark energy is even more remote still. It’s a property of cosmic-scale spacetime that only makes itself felt over billions of light-years. So these two large mysteries turn out to be irrelevant to everything within a few light-hours distance from you.

(Note: I’m treating dark matter and dark energy as gaps in the Core Theory’s “inventory”. I’m also leaving the broader cosmological story (inflation, expansion history, the standard cosmological model, etc.) alone entirely. That’s deliberate. The Core Theory is a claim about what ordinary matter is made of and how it interacts. Cosmology asks what the universe contains and how it got this way. They’re related, not the same. But cosmology deserves its own essay rather than a corner of this one. Hint, hint.)

Quantum gravity at the Planck scale. Remember, the effective-field-theory framing tells us gravity’s quantum description falls apart near the Planck energy. But that energy is “so far beyond” anything reachable, so far beyond even the LHC, so far beyond any conceivable machine, that it’s causally irrelevant to “everything” except the interior of black holes and the first sliver of an instant after the Big Bang. There’s no denying that we do need to figure it out. It’s a gap in our fundamental understanding. But it’s also a gap that has “zero bearing” on any phenomenon you will ever encounter.

(Incidentally, this is why theories involving quantum gravity (e.g., string theory) are so difficult to test. The energy levels involved are just way out of the reach of humanity.)

The matter-antimatter asymmetry, the hierarchy problem, and the various collider-scale anomalies. There are several puzzles about why the Core Theory’s parameters take the values they do. There are also tiny discrepancies at the frontier of measurement. They concern the deep structure of the theory rather than its coverage. None of these puzzles or discrepancies reaches down into chemistry or biology, or disturbs anything in the world of ordinary matter.

So the one gap in the Core Theory that even touches solar-system-scale stuff at all involves particles (neutrinos) that pass through you almost entirely untouched. And the gaps that would rewrite our deepest understanding of reality, including dark matter and dark energy, all live at scales the solar system never realizes. (So there’s a justifiable reason I made that completeness claim in Section 5.)

A third kind of gap: the theory isn’t beautiful enough

There’s a category of gaps I’ve not addressed so far. Leaving it out would likely give you an incomplete picture of how physicists view the Core Theory today. The gaps we discussed above are areas the Core Theory “doesn’t cover”. This third kind of gap is about things it covers but “doesn’t explain”.

Let’s start with the “free parameters” of the theory. The Core Theory has somewhere around twenty numbers that we cannot yet derive within the body of the theory. These include the masses of particles, the strengths of couplings, and the angles that describe how quarks mix. We have a solid theoretical understanding of how these parameters are used within the theory. But the theory itself doesn’t “predict” their exact values. Rather, we measure them, write them down, and hand them back to the theory as inputs. You could argue that a framework that requires twenty arbitrary numbers to get going is not what anyone would call “elegant”. As a result, physicists have been dissatisfied with this state of affairs for fifty years.

The three generations of fermions (quarks and leptons) belong on this list too. We don’t know why the pattern repeats three times, or why the masses climb so steeply from one generation to the next.

Then there’s the “assembly” of the whole theory. The Standard Model isn’t governed by just one symmetry. We have three. The electric charges of the quarks and leptons come in a peculiar pattern of thirds that looks arbitrary. Until you notice it’s exactly what you’d get if all three symmetries were “pieces of one larger symmetry” that broke apart early on. There’s actually a second hint pointing to a more unified symmetry. Remember when we talked about couplings “running” with energy scales? Well, if we extrapolate the three Standard Model couplings up to enormous energies, they “very nearly converge” on a single value. It’s as if these three forces were originally one force that split apart as the universe cooled. (Note that use of the qualifier “nearly” - this observation is tantalizing rather than conclusive.) That near convergence is the primary driver behind the idea known as “grand unification”. Interestingly enough, grand unification comes with a testable and rather startling prediction. If it were true, the proton, which we ordinarily treat as eternal, should very occasionally “decay”. Now, we have searched extensively for a proton decay event for decades. Nobody has seen it happen. Moreover, the searches have killed the “simplest” grand unification models without outright killing the idea.

And then there’s the strong-CP problem. A certain parameter in the strong force could, in principle, take any value, and there’s no obvious reason it shouldn’t be sizable. However, the measurement indicates it sits very, very close to zero (measured to at least 10 decimal places!). Keep in mind - this isn’t a contradiction or flaw in the theory per se. It’s just very unlikely on its own. Sort of the way it’s unlikely we could shuffle a deck of cards and end up with it being perfectly ordered. The most elegant fix for the strange “zero-ness” of this parameter introduces the existence of a new, very light particle called the “axion”. The axion would explain the zero. And interestingly, in a bit of luck almost too good to be true, the axion also happens to be an excellent dark matter candidate. Axions have not yet been detected, though people are looking hard.

(Frank Wilczek named this hypothetical particle after a brand of detergent called Axion because it seemed to “clean up the problem.”)

None of these gaps is a flaw in the Core Theory per se. Every one of these blemishes is consistent with all the data. But together, they look like clues suggesting there are some adjustments to be made. A theory that has twenty free numbers, three symmetries bolted together, and one parameter sitting at zero due to no obvious reason is a theory that’s telling us, that perhaps, there may be something underneath it.

7. Everything Above Is Emergent (and Reducible, in Principle)

So if the Core Theory governs all the interactions of ordinary matter, where do chemistry, materials, cells, and organisms come from? We briefly touched on this during the discussion about molecules and larger structures. But there is a bigger organizational principle at hand - and that is “emergence”. Specifically, what philosophers call “weak" emergence, which I want to distinguish carefully from the woo-woo kind.

“Strong” emergence (the woo-woo kind) would mean that higher-level phenomena involve genuinely new “causal" powers, forces or influences “not” present in the underlying physics and not derivable from it “even in principle”. To date, scientists have seen zero evidence for that anywhere in the ordinary world. But the tidy weak/strong split hides a genuine third possibility that philosophers of physics take seriously.

Let’s consider a common physics phenomenon you may have heard about in high school - called “phase transition”. Examples of phase transitions include water freezing, a metal becoming a magnet, etc. These are “symmetry-breaking events” of the same general kind as the Higgs mechanism from Section 3. This is a nice illustration that the concept isn’t confined to particle physics. The sharp, qualitative change represented by a phase transition seems to only appear cleanly in the idealized limit of infinitely many particles, a “singular limit” where the higher-level description can’t be smoothly recovered from the lower one by just zooming out. (You might need to read that last line a couple of times if you are not familiar with the specifics of phase transitions.) Robert Batterman and others argue such cases are neither spooky new forces nor boring hard-to-compute reductions. They argue that the higher-level story really is ineliminable, yet no new causal power has been added. So it’s right that no “new force” appears up there, but “fully reducible in principle” then becomes a little too smooth as a stance.

Weak emergence is the real and ubiquitous thing. It means the higher-level behavior is “entirely a consequence” of the underlying physics, and follows from it with no new ingredients. But the higher-level behavior is also so complicated to derive from scratch that, in practice, we describe it with its own higher-level concepts and laws. Temperature is the classic example. There’s no “temperature force” in the Core Theory. Temperature is just the average jiggling energy of a swarm of particles. It’s completely reducible to the motion of molecules. Also, it would be insane to compute that by tracking every molecule. So we use thermodynamics, a higher-level theory that emerges from, but never contradicts, the physics below it. (Incidentally, pressure as a macro variable works the same way that temperature does.) Other examples of weak emergence include wetness, superconductivity, the fractional quantum Hall effect, etc. We also see this in certain species - e.g., flocking behavior in birds and ant colony foraging. And we see it in complex socioeconomic phenomena like market prices.

Chemistry is electromagnetism plus quantum mechanics applied to electrons and nuclei. Nobody doubts this in principle, even though solving the equations of the Core Theory exactly for a large molecule is brutally hard. Similarly, biology is chemistry with a few billion years of selection layered on top. Each level is “reducible in principle” to the Core Theory, but “irreducible in practice” to it. There is no contradiction between those two statements. The “in practice” barrier is about computational difficulty and the usefulness of higher-level concepts, not about new physics sneaking in at higher levels.

I should flag that even this “layer cake” image - a fundamental floor at the bottom, chemistry and biology as upper stories, etc., is a specific philosophical framing choice, and not the only one. A well-developed rival (Ladyman and Ross, building on Daniel Dennett’s “real patterns”) resists calling the bottom level uniquely “fundamental” at all. In their philosophical framing, a biological pattern like natural selection is just as “real” as an electron field. Because, in their view, “real” means a robust, projectible pattern we’d miss if we only tracked the particles.

In fact, the effective-field-theory framing of GR leads us in this direction. The lesson there is that the notion of “fundamental” is scale-relative. Each theory is autonomous at its own energy. Each theory is also agnostic about what sits beneath. This autonomy is the result of the renormalization we discussed in Section 1. That is - the details of very short distances “wash out” and leave each level largely “sealed off from the noise below” it.

So, while I’ll keep the layer cake image because it’s intuitive, a reader so inclined could legitimately accept every empirical claim about the Core Theory in this essay while rejecting the metaphysics of “levels.” The physics doesn’t “force” that approach.

I raised similar points when talking about quantum measurement in a previous essay. So, the reductionist claim (”it’s all Core Theory underneath”) and the anti-reductionist practice (“but we’ll obviously use biology to talk about cells, not quantum fields”) are both correct. They are not in conflict.

This also closes the loop on Philip Anderson’s point I mentioned in the opening of this essay. When he wrote “More Is Different,” his target was a specific type of arrogance. The kind that suggests that because particle physicists hold the keys to the fundamental laws of physics, they “alone” study anything deep and everyone else fills in details. Anderson’s criticism is apt. New organizing principles (superconductivity, the rigidity of a solid, the logic of a cell, etc.) really do appear at higher levels. They’re genuine discoveries, not mere “footnotes to the Standard Model”. But always remember, we are just talking about the technical language we use to describe various phenomena. None of this is a claim that some new force or field sneaks in above the atom that is not captured in the Core Theory. Anderson is quite accepting of the reductionist substrate. His “dissent” is a healthy correction to a sociological issue, a cultural overreach. That viewpoint sits comfortably alongside the claim that the underlying physical laws, say in biology, are known.

8. A Bacterium, All the Way Down

Let me make this concrete with an example. “Biology reduces to physics in principle” is worth walking through. Take a single E. coli bacterium swimming in a drop of pond water. Where’s the Core Theory, and where’s the emergence?

Start at the bottom. Every atom in that bacterium is a nucleus (quarks bound by the strong force) surrounded by electrons (bound by electromagnetism). Don’t forget that 99% of the mass of that atom comes from the energy of the strong-force field inside the protons and neutrons in the nucleus rather than the bare quarks themselves. Every chemical bond in the bacterium’s cell holding its DNA, proteins, and cell membrane together is electromagnetism acting on electrons in the atoms, governed by quantum mechanics. When the bacterium metabolizes a sugar molecule, that’s electrons rearranging in chemical reactions. Which is to say, it’s “entirely” electromagnetic, entirely inside the Core Theory. Gravity is present too, of course, but at this scale it’s laughably weak and irrelevant. The electromagnetic forces holding the cell together are something like a trillion trillion trillion times stronger. The weak and strong forces are doing their jobs “inside the atomic nuclei” but aren’t otherwise driving the biology.

So, there is nothing in the “substance” of the bacterium, nothing in the “forces” it exploits, that lives outside the Core Theory. Not one interaction. If we had an impossibly large computer and the bacterium’s exact quantum state, the Core Theory equations would, in principle, predict the whole thing. (Note: this is not one quantum state per particle, since by now everything in there is thoroughly entangled, but rather a single state for the whole organism, as we discussed earlier.) It could predict the swimming, the eating, the dividing, etc. We should note the practical fact that this computation is likely forever out of reach. The point is about what the laws “contain”, not about whether we could ever actually run them. (This is why Sean Carroll likes using Laplace’s demon as a thought experiment in discussions related to emergence and reducibility.)

Emergence “stacks” up. The bacterium’s flagellum spins like a corkscrew. It’s driven by a molecular motor that runs on a gradient of protons across the membrane. That motor is “just” electromagnetism and quantum chemistry. But the “concept” of a rotary motor, of a proton gradient storing usable energy, of “swimming toward food,” are higher-level descriptions that you’d never recover by staring at wavefunctions or scattering amplitudes. They’re emergent, in the weak sense. That is, they are fully caused by the physics below, but hopeless to compute from it. And, they are enormously useful as their own layer of explanation. So a statement like “this bacterium is swimming up a sugar gradient to find food” is a “true” statement.

That sort of layering - physics at the bottom, chemistry above it, molecular biology above that, behavior at the top, etc., - each level emerging cleanly from the one below, is representative of the relationship between the Core Theory and the living world.

9. Where the Theory Gets Extended Next

So what actually pushes the Core Theory forward - to solve those gaps we talked about? Well, the near-term action is in “Beyond the Standard Model” (BSM) physics. I want to go over some promising research areas and also what type of experimental setup is used to pursue each one.

Neutrino masses and oscillations. Personally speaking, this is the most concrete crack in the current theory, and the one most likely to yield next. We need to know the answers to questions like: how do neutrinos get their mass (some version of the “seesaw” mechanism is the leading guess), are neutrinos their own antiparticles, and exactly how do the three types of neutrinos “mix”? That middle question re antiparticles might seem trivial, but it’s critical. If a neutrino is its own antiparticle (what we call a “Majorana” particle), the seesaw mechanism becomes the natural story. The way to find out specifics then is to hunt for an ultra-rare nuclear process called “neutrinoless double-beta decay”, which can only happen if that’s the case. How do we pursue this: we use giant underground detectors, shielded by rock from cosmic noise. Using these setups, physicists watch for neutrinos fired across hundreds of miles or for a rare decay event. Long-baseline neutrino experiments (to study mixing) and neutrinoless double-beta decay searches (for the Majorana question) are the main approaches here. This is underground, patient, deep-shielded physics.

Flavor anomalies and rare decays (FCNC and friends). Within the theory, we know that quarks change “flavor” via the weak force, mediated by W bosons. This process is a charged current interaction. But there are hints (from time to time in experiments) of a quark changing flavor (e.g., from top to charm) while interacting with a neutral particle (like a Z or Higgs boson or a photon). This happens through processes that the Standard Model should allow only “very rarely”. Physicists call these flavor-changing neutral currents (FCNCs). FCNCs are “nearly” forbidden in the Standard Model for (a variety of) technical reasons. But we have a good sense of the rate at which these rare events might happen. So, if a rare decay happens at a rate even slightly off from prediction, that could be a fingerprint of “new physics”. How do we pursue this: we use high-precision “flavor factories” and dedicated detectors at the LHC. These specialize in counting decays with amazing accuracy. So, this pursuit is a precision game, not a brute-force one.

The Higgs sector. The Higgs is one of the newest pieces of the Core Theory. As such, it’s also probably the least thoroughly mapped. After all, we only discovered the Higgs boson at the LHC in 2012. Key questions associated with Higgs-sector research are as follows. Does the Higgs couple to other particles exactly as the theory predicts? Does the Higgs interact with “itself” the way the theory says? Is it truly a single, simple particle, or just the first member of a family? How do we pursue this: Higgs sector research is the strongest argument for a next-generation collider. A proposed “Higgs factory” (an electron-positron collider tuned to churn out Higgs bosons in a clean environment) would measure its properties far more precisely than the LHC can. And a “muon collider”, a technically challenging but exciting idea, could reach much higher energies in a compact machine, potentially probing the Higgs self-interaction directly.

Axions and the strong-CP problem. We briefly touched on axions earlier - they are one of the most attractive open hypotheses in physics. With axions, we have a hypothetical new particle that would explain why that strong-force parameter we talked about previously sits implausibly at zero and could also act as a source of dark matter, solving two problems at once. That sort of double payoff, even if only conjectured, is rare enough in the field to be suspicious (in a good way!). How do we pursue this: well, not colliders. Axion hunting uses resonant microwave cavities inside powerful magnets, tuned slowly across frequency ranges, listening for an axion converting into a faint photon. We also use astrophysical searches for the imprint that axions would leave on stars and neutron stars. So mostly cheap, patient physics, but with a significant potential return. (Cheap relative to particle colliders, that is.)

Unification and proton decay. This might be the longest shot really, but it could also be the most transformative if we sort it out. As mentioned previously, if the three Standard Model symmetries are fragments of one larger symmetry, the proton should decay (even if extremely rarely). How do we chase this: we use enormous underground detectors. We can often use the same facilities that are also used for neutrino detection work. This involves watching thousands of tons of water or liquid argon for a single proton to decay. So far, decades of watching have found nothing and pushed the proton’s lifetime past 10^34 years. But a positive detection of a proton decaying would be among the most consequential results in the history of physics, and the next generation of detectors will push even deeper.

The rest of the frontier. Dark matter gets chased on three fronts at once. We have underground detectors waiting for a dark particle to bump a nucleus. We also have particle colliders trying to “make” it. And we have space telescopes watching for its astrophysical fingerprints. So the search also involves quite a few space experiments - precision satellites, cosmic-ray observatories, and gravitational-wave detectors, probing many aspects of the theory no ground machine can reach.

If I had to bet on where the next genuine “extension” of the Core Theory comes from in the near term, I’d put neutrinos first, flavor anomalies second, and the Higgs sector third. The Higgs sector search has the highest ceiling since it needs a new collider. Axions are the dark horse. Comparatively cheap to look for, and the payoff would be enormous.

The Bottomline

The Core Theory is one of the great underappreciated facts about our world. In the entire domain of ordinary experience, the whole solar system, every chemical reaction, every living cell, we already possess the complete set of laws governing how matter and energy interact, and we’ve tested those laws to a precision that borders on the ridiculous.

To pull the whole thing together:

• The Core Theory is the Standard Model plus gravity. In my presentation of this, I treat gravity as an effective field theory, which is a working “quantum theory of gravity” for every energy scale we’ll ever meet, agnostic only about phenomena at the Planck scale.

• At the bottom, the Core Theory is all about quantum fields and spacetime, along with symmetries and the relations they enforce. The forces of nature fall out as what’s required to keep those symmetries intact. Above the level of the atom, essentially all of chemistry and biology is essentially just one of those four forces (electromagnetism) endlessly elaborated, refereed by quantum mechanics.

• In the solar system, this Core Theory package accounts for every interaction that can touch ordinary matter.

• We do have gaps in the theory, but the neutrino-related ones are the lone “Earth-touching ones” and they have virtually no impact on everyday phenomena on Earth. The deep mysteries (Planck-level quantum gravity, dark matter, dark energy, etc.) all sit at scales the solar system never reaches.

• A third kind of gap in the theory is aesthetic rather than empirical: twenty-odd free parameters we cannot derive values for, three symmetries bolted together, one parameter sitting implausibly at zero. But none of that contradicts the data or undermines the theory itself.

• Everything higher up is weakly emergent: fully caused by the physics below, hopeless to compute from it, and best studied in its own right.

Above the foundation provided by the Core Theory, everything else, chemistry, life, minds, weather, a bacterium chasing sugar, emerges. Not by magic, and not by adding new forces. Rather, by the ordinary miracle of simple laws organizing themselves into systems of staggering complexity. No new force appears at the higher levels, even if (as the physics of phase transitions hints) the reduction back down to lower levels isn’t always as clean as a slogan would make it.

So the next time someone tells you physics is nowhere near understanding the world, you can nod politely, and then quietly note that for the entire world they’ll ever touch, taste, or live inside, we cracked the fundamental explanatory physics a while ago. The mysteries that are left are real, but they’ve been chased out to the edges - the very small, the very energetic, the very far away.

This accomplishment, the development and packaging of the Core Theory, surely stands out as one of the greatest human achievements of all time.

Thanks for reading! Subscribe for free to receive new posts and support my work.

For some decades now, physicists have been working on uncovering the details surrounding a very interesting fact: if you want to know how much “stuff” you can cram into a region of space, the answer is not set by the volume of that region. It’s set by the “area of its boundary”.

If this clashes with your intuitions, join the rest of us who have wrestled with the implications of this notion over the years. That’s not how anything else works. If we want to know much water fits in a swimming pool, we’d measure the pool, not the tiles lining it. If I ask how many books fit in a library, you count shelf space, not the area covered by wallpaper. Volume scales as length cubed. Surface area scales as length squared. Every intuition we have says the “interior” is where the capacity lives.

But when we ask how much “information” can fit inside a region of space in a universe that has gravity, the volume of the region turns out to be the wrong number. The real answer appears to be tied to the “area of the surface” enclosing that region.

This is the holographic principle. Research into this concept began as an accounting puzzle about black holes. Oever time, the concept mutated into one of the most-cited ideas in modern theoretical physics, and has become the lens through which many in the field now views spacetime itself. The field has produced interesting calculational victories in places nobody expected. It has also produced a slow-burning crisis of conscience, because the version of holography we understand best appears to describe a universe that is not ours, and may never describe ours.

1. It Starts With Black Holes (As These Things Usually Do)

In the early 1970s, Jacob Bekenstein was worrying about a thermodynamic “embarrassment”. The second law of thermodynamics says entropy never decreases.

(Entropy is, roughly, a measure of how many microscopic arrangements are “compatible” with what you see from the outside. A shuffled deck has more entropy than a sorted one, because there are vastly more ways to be shuffled than there are to be sorted. It’s a count of “hidden possibilities”.)

Now take a hot cup of coffee, with all its entropy, and drop it into a black hole. The coffee is gone. Its entropy seems to have vanished as well. The outside universe just got “tidier”, for free. The second law seems to have been violated, and it’s been violated by a process as strenuous as letting go of a mug.

Bekenstein’s move was to say: the entropy isn’t gone. In fact, the black hole “has” entropy, and when you “feed” it coffee cups and unwilling grad students holding those cups, its entropy goes up by at least as much as you fed it. The entropy ledger ultimately balances after all.

But then, exactly how much entropy does a black hole actually have? Bekenstein’s 1972 answer is the starting point for the tour of physics we take in this essay: A black hole’s entropy is proportional to the area of its event horizon. Not the volume of the black hole, but its area.

At the time this was proposed, Stephen Hawking thought this was nonsense. Any thing thing with a temperature has to “radiate”. And a black hole, by definition, is the thing that nothing escapes from. So it can’t radiate, so it can’t have a temperature, and so it can’t have entropy. In Hawking’s view, Bekenstein had made a category error, and that was that.

Then in 1974, Hawking redid the calculation with more rigor - applying quantum field theory to the curved spacetime around a horizon - and discovered that black holes radiate after all. They have a temperature, they glow faintly, they lose mass, and given enough time they evaporate away entirely.

So Hawking set out to disprove Bekenstein’s idea and instead ended up elaborating the details that made it work. Hawking also supplied the exact coefficient Bekenstein had been missing. A productive way to lose an argument, if you ask me.

The coefficient turned out to be one quarter. Specifically, the entropy of a black hole is one quarter of the horizon area, measured in units of the Planck area, which is about 10^-70 square meters - an absurdly, comically small tile. If we tile the horizon with those Planck bits, count the tiles, we’d have counted the black hole’s hidden information.

So the “information content” of a black hole scales with its “boundary”, not its “interior”. And a black hole is plausibly the most entropic thing that can occupy a given region at all - a claim I’ll come back to and complicate shortly, but let’s hold onto it (loosely) for now.

Hawking’s result also raised a question that has driven the field ever since. If the black hole eventually evaporates away to nothing, then where did all the information that fell in go? This is the black hole information paradox, and holography is, in a sense, part of the field’s collective answer to it.

2. From Curiosity to Principle

Gerard ‘t Hooft in 1993, and then Leonard Susskind in 1994, took the black hole result and asked the question: what if this “information capacity scales with area instead of volume” isn’t a special fact about black holes? What if it’s a fact about “space” itself?

The logic works like this. Suppose we try to pack more entropy into a region than a black hole of that exact size would have. If we keep adding stuff, at some point the region collapses into a black hole - which puts us right back at the black-hole value. So the black hole of a certain size represents the “maximum entropy budget” of an equivalent region of space. Which, considering what we discussed in the previous section about the Beckenstein bound, would mean: the maximum information in any region of space is set by the area of its boundary, in Planck units.

(Note: this argument is suggestive rather than airtight. The original version has known loopholes and disputed counterexamples in field theory, and pinning down what “the entropy inside a region” even “means” quantum-mechanically turns out to be part of the difficulty.)

It was Lenny Susskind who gave this idea the name that stuck. A hologram is a two-dimensional film that encodes a three-dimensional image. The “holographic principle” says the universe works the same way - all the physics happening inside some volume can be fully encoded on the surface that bounds it.

Note the strength of the claim here: we’re not saying “the boundary contains a useful summary of the interior.” Rather, the claim is “the boundary contains “all of it”, with nothing whatsoever left over.”

3. The Bousso Bound: Getting the Statement Right

The naive bound we just talked about has a problem. The problem is that “a region of space at a given moment” is not well defined in relativity. There is no universal “now”, and different observers slice spacetime into moments differently. Worse, in a closed universe, you can find surfaces whose interior contains “arbitrarily large entropy” while the surface itself stays small, because the very notion of “interior” is ambiguous once space “wraps around”. So in settings like cosmology, the naive bound simply fails.

Raphael Bousso, in 1999, fixed it by changing what you count. (One of my favorite physics papers.) Instead of counting the entropy on a spatial slice “inside” the surface, we count the entropy crossing a particular “light-sheet“.

Here’s how we calculate this. Start with our surface. Now let’s fire light rays inward from every point on the surface, and follow them. The rays will sweep out a “sheet” through spacetime. Crucially, we also only follow them as long as they’re “converging” - meaning, as long as the cross-sectional area is shrinking. The moment they’d start to spread out, you stop.

The Bousso bound: the entropy passing through that light-sheet is “at "most one quarter” the area of the surface you started from.

The result of using the Bousso bound instead of the naive one? Unlike the naive version, it “works”. It holds in cosmology, in collapsing stars, in expanding universes, in every case where the simpler formulations fall apart. And the bound doesn’t care how we slice spacetime, which was the original sin of the naive version.

Why is this technical detail important? It’s because the Bousso bound suggests that holography is not a quirk of black holes, and not a feature of one particular theory. It looks like a structural constraint on “any” theory that has both gravity and quantum mechanics in it. If the bound is right, then whatever the “final theory” turns out to be, it “has” to be holographic. Because the number of degrees of freedom it can support in a region is already fixed, and it’s constrained by an area.

That’s a startling thing to know about the final theory “before” we know the actual final theory.

(Technical note: the bound is very well supported and has survived every case anyone has thrown at it, but it is not a proven theorem in full generality. )

4. Gauge/Gravity Duality: Where the Principle Became a Machine

For about five years, holography was a beautiful principle with no concrete implementation. Nobody could write down an example.

Then, in late 1997, Juan Maldacena wrote a paper that has become one of the most-cited in the history of physics and sparked a small revolution in theoretical physics. He gave an explicit example: the AdS/CFT duality (Anti-de Sitter Space/Conformal Field Theory duality)

The claim (de-jargonized): a specific “theory of gravity” living in a five-dimensional curved spacetime is exactly, precisely, in every measurable respect, the same theory as a specific “theory of particle physics with no gravity at all” living on the four-dimensional boundary of that spacetime.

So, we have two separate theories that look nothing alike. One has gravity, curved spacetime, black holes, and strings. The other is a quantum field theory of the same broad family as the ones we use to describe quarks and gluons - living in flat space, with no gravity anywhere in it whatsoever. And it turns out they are “the same theory”. Every question you can ask in one has a “translation” into the other, with matching answers. Weird, huh? (It “is” weird and sounds weird to most physicists who come across it for the first time.)

The theory with gravity lives in anti-de Sitter space (AdS), a spacetime with constant negative curvature. This spacetime is called the “bulk.” The theory that lives on the boundary of that spacetime is a conformal field theory (CFT). Hence the standard name for this duality: AdS/CFT, or more generally, gauge/gravity duality.

(Technical note: the best-tested version pairs type IIB string theory on five-dimensional AdS with a gauge theory on the boundary called N=4 super Yang-Mills. This gauge theory is emphatically not the theory of the real world, a point I will return to with some force.)

In any case, Ads/CFT has become the best known example of the holographic principle in physics.

The word “duality” incidentally is one of the most-misunderstood words in the subject. A duality is not an analogy. It is not a resemblance, or a useful metaphor, or an approximate mapping that works in some limit. A duality is the claim that two descriptions which look completely different are “descriptions of the same thing”. Like a person’s name in English and the same person’s name in Mandarin point to the same referent (the person) despite being written down in different symbols.

Why This Was Immediately, Obviously Useful

The reason AdS/CFT exploded in the theoretical physics field is not primarily philosophical. It is that the dictionary (the map between the two theories) is “strong/weak”.

Our best known theories of matter and energy in the universe are based on quantum field theories (QFTs). QFTs have parameters called coupling constants, which are numbers indicating how strongly particles in the theory interact. When couplings are small, the particles mostly ignore each other and you can calculate the interaction behavior by treating the interactions as small corrections. This is called perturbation theory, and it is the workhorse of the QFT arena. It is why quantum electrodynamics, where the coupling is conveniently tiny, can predict the electron’s magnetic moment to something like ten significant figures, one of the most precise agreements between theory and experiment anywhere in science.

When the coupling is “large”, perturbation theory stops working. The corrections are no longer small. The “series” involved do not converge. Sadly, an enormous amount of interesting physics lives in exactly that regime. For example, the confinement of quarks inside protons, high-temperature superconductivity, nuclear matter at extreme densities, etc.

We do have one serious non-perturbative tool: lattice gauge theory, which chops spacetime into a grid and simulates the thing numerically. Lattice QCD works, and it computes hadron masses to percent-level accuracy, which is a genuine triumph. But it has some hard limitations. It works in “Euclidean” time, a mathematical trick that turns the problem into something a computer can sample, and that trick fails badly under certain conditions (real-time dynamics and for matter at finite density). So questions like “how does this plasma actually evolve, second by second” and “what happens inside a neutron star” remain largely out of reach.

This leaves a large and important class of strongly coupled problems where we have historically been stuck. And that’s where AdS/CFT comes in to assist.

In the AdS/CFT dictionary, when the field theory on the boundary is “strongly” coupled, the corresponding gravity theory in the bulk is “weakly” curved and easy to work with. The regime where one description becomes intractable is precisely where the other becomes simple.

So we take our very-hard-to-calculate strongly-coupled field theory problem, translate it through the dictionary, and it becomes a problem about a black hole in a curved five-dimensional spacetime. Which we can solve, with regular classical general relativity! Then we translate the answer back.

Also, because the gravity calculation is done in ordinary Lorentzian time, holography can reach exactly the real-time, dynamical questions where lattice methods run into problems. Meaning we can tackle how a plasma flows, how fast it thermalizes, how it dissipates, etc. This why the most famous physical result of the whole program (which I’ll discuss in a bit) is about a “flowing fluid”.

This is also why AdS/CFT is not some fringe topic. It is why condensed matter physicists, who do not care about quantum gravity or string theory, started reading string theory papers.

5. Celestial Holography: The Attempt to Get Out of the Box

Anti-de Sitter space has a “boundary” that is a key part of the architecture. Think of AdS like a box. Light rays sent outward reach the conformal boundary in finite time, so we impose boundary conditions there by hand, and the usual choice is reflecting. So there is an “edge to AdS space”, and the dual field theory “lives” on the edge. Convenient.

But our universe does "not have that kind of wall. Our universe is mostly flat. And the entire scaffolding of AdS/CFT, built on having a nice timelike boundary to hang the field theory on, does not (obviously) transfer.

It is natural to describe the flat-space program (for our universe) as an “escape” from AdS, a workaround to make it work for our universe. Someone like Andrew Strominger, who has done more than anyone to build a version that works with our universe, would think of this the other way around. In his view, AdS is the artificial case! It is a box with reflecting walls that nobody ordered. Flat space is where scattering actually happens, where colliders and gravitational-wave detectors actually live. The field only spent twenty-five years doing quantum gravity in a box because that is where there was early low-hanging fruit. Now that the program has matured, the flat-space program is the upgraded attempt to go look for models that work in our universe.

Holography in flat spacetime. In flat spacetime, particles can be thought of as coming in from infinitely far away and flying off to infinitely far away. Each incoming or outgoing massless particle heads in some direction, and we can label that direction by a point on a sphere. Stand outside on a clear night and point at where a photon came from: that is a point on the “celestial sphere”.

Celestial holography says: let’s say we rearrange scattering theory, the thing we use to do measurements in particle colliders, so that instead of labeling particles “by momentum” we label them “by where they hit that sphere”. If we do that, the four-dimensional scattering amplitudes start transforming exactly the way correlation functions of a two-dimensional conformal field theory do. Hmm, that sounds familiar. This starts to look a lot like a duality between how particles scatter in our universe to a field theory on the celestial sphere (at “null infinity”).

(This next part involves going into some history.)

BMS symmetry groups In the 1960s, a group of physicists and mathematicians formulated something called the BMS symmetry group as part of their investigations into the flow of energy due to propagating gravitational waves. This result was (mostly) set aside by the particle physics community for fifty years (of course, the relativists never stopped studying it). The BMS group is the actual asymptotic symmetry group of flat spacetime, and it is infinite-dimensional, which came as a shock since everyone had expected that to be a more common group in relativity - the Poincaré group.

The infrared triangle Then a few years ago, Andy Strominger figured out that three separate discoveries in the history of theoretical physics turned out to be “the same fact in different languages”: a formula for emitting very low-energy gravitons (soft theorems), a hidden infinite family of symmetries (the BMS ones) of spacetime at infinity, and the permanent shift a passing gravitational wave leaves in nearby objects (memory).

(Technical detail: What Strominger did was to show that the soft theorems (statements about emitting very low-energy gravitons, known since Weinberg in 1965) are exactly the Ward identities of those BMS symmetries. So the soft theorems follow “by derivation” from the symmetries. Add the memory effect, a permanent displacement left behind by a passing gravitational wave, and you get what Strominger calls the infrared triangle.)

Asymptotic symmetry The structure behind all of that turned out to be bigger than anyone expected. In 2021 Strominger showed that the celestial symmetries organize into a w-infinity algebra, an infinite tower of soft graviton symmetries, implying gravity has vastly more symmetry than suspected. This massive symmetry is known as “asymptotic symmetry,” and it acts on the edges of spacetime, specifically the "celestial sphere" at infinity. And, this sort of symmetry makes certain types of difficult problems “solvable”. Whole families of 2D conformal field theories were cracked using symmetry and bootstrap methods, and for many of the most important ones nobody has a useful Lagrangian even now. (I say that because one of the usual objections to this approach is that that nobody has written down the celestial theory’s Lagrangian.) In any case, all of this serves to augment the toolset that physicists can deploy to solve complex problems in field theory.

The catch (so far). There is no celestial CFT. Well, at least not in the sense of an independent object you could hand someone and say “go forth and compute with this.” The dictionary has been written almost entirely from the bulk side: we take what we already know about four-dimensional scattering and observe that it “looks like” CFT data “once transformed”. The original ambition was to run it the other way. Meaning we define a boundary theory and use it to learn something new about the bulk. That has not happened yet.

And even the tentative dual that exists today is a “strange” CFT. Its conformal weights (used to describe how an object changes when you stretch or shrink its space) run over a continuum and are generically show up as complex numbers. Moreover, its correlators are distributional (think spiky), and it is unclear how far these objects differ from ordinary unitary CFTs or whether they obey anything like the usual axioms. The celestial sphere also sits two dimensions (2D) below the bulk (4D) rather than one, a structural oddity nobody has fully explained. Meanwhile a rival program, Carrollian holography, attacks the same problem by keeping retarded time (used in Carrollian CFTs as a time coordinate) and putting a three-dimensional theory on null infinity. It also has the advantage that you can write explicit Carrollian theories down. Which framework is right, or whether they are two views of one thing, is open and actively contested.

Empirical possibilities The memory effect may be detectable by next-generation gravitational-wave observatories, and in papers on this topic, we see this billed as an upcoming “test” of celestial holography. Sort of, but not really. Memory is actually a prediction of ordinary classical general relativity, derived long before anyone said the word celestial. A more defensible claim is that memory, soft theorems, and asymptotic symmetries are three faces of a single structure, we can only observe one of those faces, and observing it confirms “the structure is there”. So the claim really is weaker than “we can test celestial holography,” yet stronger than “this is untestable in principle.”

6. The Successes

It would be easy to dismiss all of this as beautiful mathematics. That would be wrong.

Solving strongly-coupled field theories. This is currently the main use of AdS/CFT. AdS/CFT gives non-perturbative answers to questions about strongly-coupled gauge theories that were previously unanswerable. In the appropriate limit (boundary theories with many colors, very strong coupling) those answers are exact. For a field where “we cannot compute anything at strong coupling” had been a standing complaint since the 1970s, this is a big deal! Whether that limit is close enough to the real world is something I’ll discuss in Section 8.

The viscosity “bound.” Using the gravity dual, Kovtun, Son, and Starinets computed the ratio of shear viscosity to entropy density for a strongly coupled plasma and found a remarkably small universal number, which they conjectured was a lower bound on how perfect any fluid can be. (A fun paper to read.) Then heavy-ion collisions at RHIC and the LHC produced a quark-gluon plasma that behaved as an almost perfect fluid, with a viscosity in the right ballpark and far below what perturbative QCD had originally predicted.

I do need to say that this (apparent) bound is not actually a bound: it has since been violated in holographic models with anisotropy or higher-derivative corrections. And the experimental number comes from fitting hydrodynamic simulations to collision data, with real model dependence. But.. holography says strongly coupled plasmas should be nearly perfect fluids with a specific small viscosity. Perturbative QCD said otherwise. It’s turns out that holography is right in this regard - so this is a genuine result.

Black hole entropy, counted. Strominger and Vafa in 1996 counted the microscopic quantum states of a particular black hole directly in string theory, and the count reproduced the Bekenstein-Hawking area formula exactly, factor of one quarter included. (This was also an interesting result for string theory.) This was the first time anyone derived black hole entropy from a microscopic theory rather than postulating it. The result has its limits. The black hole in question was an extremal, supersymmetric one in five dimensions. Supersymmetry is what protects the state count as we dial the string coupling from the regime where we can “count” states to the regime where the object is actually a “black hole”. For a garden-variety astrophysical black hole, no such protection exists, and the procedure does not work.

Entanglement entropy and the geometry of spacetime. Ryu and Takayanagi in 2006 found that if we take a region on the boundary and ask: how “entangled” it is with the rest of the boundary. Turns out - entanglement on the boundary equals “area” in the bulk. Well, that’s the Bekenstein-Hawking formula again, but now it has nothing to do with horizons at all. It’s about “arbitrary” regions, and it’s about the quantum information structure of the boundary theory. (It’s a fascinating result really that’s changed the tenor of the field.)

What could that mean? Well, spacetime geometry might just be entanglement, seen from the other side of the dictionary. Mark Van Raamsdonk wrote a paper that makes this vivid. If you take the boundary theory and “dial down the entanglement” between two halves of it, the corresponding bulk spacetime “pulls apart”. If you reduce the entanglement to zero, the bulk “splits” into two disconnected pieces. The thing holding spacetime together in this model appears to be quantum entanglement!

The Page curve. Around 2019, holographic methods finally produced a calculation of how the entropy of Hawking radiation evolves as a black hole evaporates, and it came out right. The entropy rises, then falls, tracing the curve that unitarity demands (the Page curve) rather than rising forever (as Hawking’s original calculation implied). The machinery involves something called “islands,” which are regions inside the black hole that end up counted as part of the Hawking radiation’s description.

Note that what this delivers is the correct “entropy curve”, not a “mechanism” per se. Nobody can tell us how to decode the infallen encyclopedia (if we tossed an encyclopedia into a black hole) out of the radiation. All we can say is that the bookkeeping is now consistent with the encyclopedia being “in there somewhere”. But at least, thanks to holography, the black hole information paradox is in dramatically better shape than it was in 2018.

7. The Connection to Information Theory

Let’s step back a bit and try to identify what all these results have in common. Bekenstein’s entropy is an information count. The Bousso bound is a bound on information. Ryu-Takayanagi is a statement about entanglement, the central concept in quantum information theory. The island calculations are about which information is where. The holographic principle, at its bottom, is not really a claim about geometry, but rather about “how much can be known” (and where the knowing “lives”). So what’s common here across these concepts is “information.”

We’ve talked about the boundary “encoding” the bulk. But “how”? If we want to know about some tiny region deep in the middle of the bulk, which part of the boundary do we look at? The answer, from Almheiri, Dong, and Harlow in 2014, is that the bulk is encoded in the boundary the same way a quantum error-correcting code encodes its data.

Quantum error correction, incidentally, is one of the key technologies that makes quantum computers possible. You take one logical qubit’s worth of information and smear it “redundantly” across many physical qubits, such that no single physical qubit holds the information but the collection does. If we damage any one of them, or several, we can still recover the logical qubit intact, due to the redundancy.

Well, it appears that this is how the bulk relates to the boundary. Information about a point deep in the bulk is not stored at any particular boundary location. It is smeared redundantly, and you can erase chunks of the boundary and still reconstruct the deep bulk. Regions near the boundary are encoded locally. The deeper we go, the more redundantly they are stored.

This suggests that the reason spacetime “has” a well-defined geometry (meaning the reason a point in the middle of a room is a robust, stable, agreed-upon thing) may be that the “universe is running an error-correcting code”. If so, locality, the fact that things here don’t instantly affect things there, may not be a primitive feature of reality. It could be an approximate property of a redundant encoding.

8. The Uncomfortable Part

Let’s look at some of the biggest challenges that the field faces.

The universe is the wrong shape. AdS/CFT works in anti-de Sitter space, which has a “negative” cosmological constant. Our universe has been observed to have a “positive” cosmological constant. So, our universe is really asymptotically de Sitter, or something like it.

Moreover, the whole structure of AdS/CFT depends on the existence of the boundary. AdS happens to have a nice timelike boundary at infinity where we can put the dual field theory. But, de Sitter space does not. It has a cosmological horizon instead. More specifically, different observers in de Sitter have different horizons, and there is no obvious place to stand and write down the dual theory. Every attempt at a de Sitter duality/correspondence runs into some structural weirdness or other - the proposed duals are non-unitary, or Euclidean, or have complex central charges, or live at some abstract future infinity no observer can access.

So - we have this beautiful, precise, holographic duality (AdS/CFT), but it describes a universe that is “not the one we live in”.

The dual field theories are the wrong theories. N=4 super Yang-Mills, the boundary theory in the most common AdS/CFT duality example, is not how our universe’s field theory works. N=4 SYM is supersymmetric and exactly conformal (it has no massive particles and thus no mass gap, and since interaction strength does not change across scales, does not exhibit confinement). Real gauge theories like the ones we have in our universe break both those properties, and confinement in particular is one of the things we most wanted explained. So, at best, we have built “QCD-like” holographic duals, and those models are useful, but they are still just “models.”

The “limits” it needs are not the limits we live in. The useful version of the AdS/CFT duality requires a huge number of colors and very strong coupling. Real QCD has three colors and is not, at the energies we care about, in the extreme strong-coupling regime. Moreover, the gravity side of the duality is only simple when the field theory is very strongly coupled. Moving away from that limit means adding stringy corrections, at which point the gravity side stops being easy.

It is a conjecture. There is no “proof” of AdS/CFT. There is an overwhelming pile of evidence that suggests that the duality exists and its exact. But after nearly thirty years we still cannot derive it from first principles or explain precisely “why” it is true. We have a dictionary that works, and no theory of “why” the dictionary exists.

The possibility that we never get there

It is entirely possible that the holographic program produces a complete, rigorous, beautiful theory of quantum gravity for a universe we do not live in, and never manages to say anything empirically checkable about ours.

Anti-de Sitter space is like a box, and the AdS/CFT correspondence depends on that box-like nature. A box gives us a boundary, a boundary gives us somewhere to define the dual theory, a well-posed notion of scattering, a clean separation of inside from outside, etc. Almost every conceptual tool in the toolkit is downstream of having that type of “wall”.

But, cosmology in our universe has no wall. In cosmology, the observer is “inside” the system. We can’t stand outside the universe and measure its scattering amplitudes. In fact, there’s a real possibility that de Sitter holography is hard not because aren’t clever enough, but because the structure holography needs “does not exist in a universe like ours”.

Perhaps AdS/CFT is a technique that requires an exterior, and cosmology, by definition, has no exterior. If so, AdS/CFT is only a very sophisticated solved model. Kind of like an Ising model of quantum gravity: exactly solvable, very instructive, but it describes nothing we can point a telescope at.

Defending the theory

Let’s try and defend holography from some of these criticisms.

We may well have bad premises behind some of our criticisms. We started out by treating AdS/CFT as the paradigm of what holography “is”, noticed that the paradigm does not fit our universe, and concluded that holography might not reach our universe. But that inference only works if AdS/CFT “is the model” of the thing. What if it’s not? What if it’s only the model of the “easy case”?

We have evidence of the existence of black holes through telescopes and LIGO. The information paradox is not just some sort of puzzle exclusively about anti-de Sitter space. Meanwhile the infrared structure of gravity, the asymptotic symmetries and soft theorems and memory effect (the infrared triangle we talked about earlier), is flat-space physics. That’s derived with no cosmological constant of either sign, and some of it is observable.

So we are left with something like this: the optimist says we have mistaken the limits of one implementation for the limits of the principle, and the principle is doing fine. The pessimist says every implementation of Ads/CFT have requires an “asymptotic structure” (the boundary) to hang the dual theory on, and our universe has no comfortable asymptotic structure.

9. Current Frontiers

What’s the current state of research in the field today?

De Sitter holography. There are several programs trying to making holography work in de Sitter space, none clearly “winning”. Some put the dual on a stretched horizon. Some try holography at finite cutoff, moving the boundary in from infinity to a finite distance, which might let you handle spacetimes without nice asymptotic boundaries. Some put the theory on the observer’s cosmological horizon.

Defining the flat-space dual. The celestial holography program has a symmetry algebra and no theory. The Carrollian program has explicit theories and a less developed dictionary. A genuine, independently defined dual for asymptotically flat space would be the biggest event in the subject since Maldecena’s 1997 paper.

Making the islands make sense. The Page curve calculation works, but it computes the radiation’s entropy using a formula that says “include this region inside the black hole, which is spacelike-separated from the radiation.” Why is that even legitimate? The formalism gives the right answer but nobody is satisfied they understand why.

Complexity and the interior. What quantity on the boundary corresponds to the “volume behind a black hole horizon”, which keeps growing long after the black hole has thermalized? Susskind’s answer is computational complexity. That is, the growth of the interior corresponds to the growing difficulty of the quantum computation the boundary is performing. If that’s on the right track, it would link spacetime to computer science and information theory in a way that would have seemed very strange twenty years ago.

Wormholes and the ensemble problem. In the low-dimensional cases where we “can” calculate carefully, the gravity theory appears dual not to a single boundary theory but to an “ensemble” of them, sort of like an average over random theories. If that is general, it is philosophically disturbing: gravity would correspond not to a definite microscopic theory but to a statistical average over them. This continues to be an area of research.

10. What Does It Mean? (Philosophically)

(This section engages in philosophical musings, so feel free to skip it if that’s not your cup of tea.)

Suppose AdS/CFT is exactly right. Two descriptions, perfectly equivalent, indistinguishable by any possible measurement. One has five spacetime dimensions and gravity. One has four and no gravity.

Which “one” is the world? There is now a serious philosophical literature on this. My primary reference here is Sebastian De Haro and Jeremy Butterfield’s The Philosophy and Physics of Duality, the book-length treatment of exactly this question.

Option 1: The boundary is fundamental, the bulk is emergent

This seems to be the majority intuition, especially among people who come at the problem from a quantum information perspective. The field theory on the boundary is the real thing. Gravity, spacetime, the extra dimension, black holes, etc. in the bulk: all “derived.” Sort of the way temperature is derived from molecular motion. (We call this emergence.) Nobody thinks temperature is fundamental. It is a real, robustly measurable, enormously useful summary of underlying microphysics. On this view spacetime is like temperature. Real in exactly that way, which is to say emergent and non-fundamental.

Why is this appealing? Well, it takes gravity, the recalcitrant force that has resisted quantization for eighty years, and says: we never had to quantize it. It was never fundamental. We were trying to find the quantum of temperature. Hence some of the newer lingo we’ve been hearing: emergent gravity, emergent spacetime, the bulk emerges from the boundary, etc.

But there are problems with this framing.

The duality in Ads/CFT is meant to be an isomorphism. Every state, every quantity, every dynamical fact on one side has exactly one counterpart on the other. Nothing is lost in translation, and the map runs both ways.

But emergence requires a map that loses information. This is not incidental to emergence, it “is” emergence. Temperature “emerges” from molecular motion precisely because a huge number of distinct microstates give the same temperature. The map is many-to-one. That throwing-away “is” what makes the higher level a genuinely “new level” rather than a rewrite.

A map cannot simultaneously preserve everything and throw things away. If the duality is exact, nothing was lost going from boundary to bulk, and so nothing “emerged”. The bulk no more emerges from the boundary than the boundary emerges from the bulk. They are really the same theory in different guises.

Either the two sides carry the same information, in which case we have a duality and no emergence, or one side is a lossy summary of the other, in which case you have emergence and no duality. We do not get both.

Option 2: Neither is fundamental, the duality is

If two descriptions are exactly equivalent and no measurement distinguishes them, then “which is real” might well be a “malformed” question (like asking whether the electron is really a wave or really a particle). What is real is the “structure” both descriptions instantiate. This is a version of structural realism: the world is a pattern of relations, and theories are notations for that pattern.

But it also seems weak. It is tempting to imagine the shared core of a duality as a thin abstract skeleton. But when you actually compute it, the common core of AdS/CFT is “rich”. It has states. It has quantities with definite values, masses, spins, energies, angular momenta, etc. It inherits a conformal symmetry from both sides. It’s not some ghostly pattern of relations. It’s a physical theory so we would really be a straightforward realist about it.

For example, let’s take electric-magnetic duality, which is “rigorously established”, not some conjecture . Under it, “purely electric” and “purely magnetic” are not “intrinsic” properties of a state. Does that mean the ontology is “empty”? No. Compare special relativity, where there is no such thing as an intrinsically electric field, because a boost turns it magnetic. Nobody concluded from relativity that electromagnetism is a pattern with no substance. They concluded that there is an “electromagnetic field”, a perfectly real entity, richer than either of its frame-dependent shadows, and that “electric” and “magnetic” are “relational” facts about how you are looking at it.

So in this framing, the common core is not “thinner” than the duals. It is often “thicker”, because it inherits from both.

So it is Option 1 or Option 2?

The De Haro-Butterfield book does have an interesting criterion that let’s us address this.

When are we “entitled” to say two duals describe the same world? Only when the theory is “unextendable”: when its domain of application is already “everything”, with nothing outside it that could break the tie.

For example, Kramers-Wannier duality maps a hot Ising lattice onto a cold one. So are hot and cold the same thing? Obviously not. Well, why not? Because we measure temperature with a thermometer, and the thermometer is “not” part of the Ising model. It sits “outside”. If we extend the description to include it, then the tie breaks, and the two duals “come apart”. The duality was exact “within its domain”, and the domain was “not everything”.

Similarly, whether a gauge field counts as electric or magnetic is settled by its coupling to an '“external current”. So the thing that “breaks” the symmetry always lives “outside” the theory.

So: duals genuinely describe the same world “only” if the theory is a cosmology. For a duality, we’d need a theory of everything, containing its own measuring apparatus, with no outside to appeal to. Anything less, and there is always a possible extension lurking that could pry the descriptions apart and reveal they were never saying the same thing.

The grand ontological reading of AdS/CFT, dimensionality is representational, spacetime is not fundamental, all of it, is “licensed only if the theory is unextendable”. And N=4 super Yang-Mills paired with strings on AdS-five is not a theory of everything. It is not even a theory of anything actual. It has no thermometers in it.

So, the philosophical payoff and the empirical payoff turn out to be both gated by “the same condition”. To be sure the ontological lesson actually holds, you need the final theory. But if you had the final theory, you would not be reasoning from toy models!

Problems using Ads/CFT as the benchmark

The emergence-versus-isomorphism argument rests on a premise: that a duality is an “exact isomorphism”. But where did that premise come from? From AdS/CFT, the paradigm case, which happens to have a clean invertible dictionary.

But suppose the flat-space advocates are right that AdS is the artificial case, and the real relation between a gravitational bulk and its holographic description, in a universe without a convenient boundary, is “not” a clean isomorphism. The hints are already there: the celestial dual sits at codimension two, its correlators are distributional, and in the low-dimensional cases where you can check, gravity looks dual to an “ensemble” of theories rather than a single one, which is very much not an isomorphism. If that is how it goes, then we have used AdS-shaped philosophy to police the language of a program trying to escape AdS.

Personally, I do not think this rescues the emergence slogan. If the relation is not an isomorphism, then it is not a “duality” either, and we need new vocabulary rather than the old vocabulary of emergence. But it does show that “given an exact duality”, emergence talk is confused. It does not show that we have an exact duality in the case we actually care about.

So where does that leave us?

The most philosophically important suggestion to come out of theoretical physics in fifty years is that space may not be fundamental. Not “space is pixellated at the Planck length.” Something stranger: space may be what a certain kind of quantum information structure “looks like” when you plot it. Distance is a measure of entanglement, locality is an error-correction property, and gravity is the shape of the encoding, etc.

What we have is an exact isomorphism between a gravitational theory and a non-gravitational one in a universe with a boundary, plus some strong suspicions that the lesson generalizes to less idealized situations like our own universe. What we do “not” have is license to say the bulk emerges from the boundary in any ordinary sense, because an isomorphism is not an emergence. Nor do we have the unextendability that would promote “the duals describe the same world” from a well-motivated interpretation into a settled fact.

The Bottomline

To summarize what we’ve covered:

• Black hole entropy scales with the “area” of the horizon, not the volume inside it. That’s the seed of everything else in this essay.

• The Bousso bound generalizes this properly, and suggests holography isn’t a quirk of black holes at all - it looks like a structural constraint on “any” theory with both gravity and quantum mechanics in it.

• AdS/CFT turned the principle into an “exact example”. A theory of gravity in five dimensions is exactly the same theory as a particle physics theory in four, with no gravity in it. Strongly coupled problems that were hopeless became black-hole problems that were tractable.

• It has actually delivered. Near-perfect fluids in heavy-ion collisions. A genuine microscopic count of black hole entropy (for a very special black hole). The Page curve. And a set of results tying entanglement to geometry so tightly that they may be “the same thing”.

• The bulk sits inside the boundary the way a quantum error-correcting code sits inside its physical qubits. Locality, the fact that “here” doesn’t instantly affect “there”, may be a redundancy property rather than a fundamental one.

• But the best-understood version of all this lives in a “box” (Ads), and our universe doesn’t have walls. The flat-space program is trying to get out of the box, and has produced real theorems but no dual theory.

• A key philosophical result, “spacetime isn’t fundamental”, seems to be gated by a condition (unextendability) that only a theory of “everything” could ever satisfy.

The folks who work on celestial holography argue that we should not confuse the limits of one implementation for the limits of the idea, and that the flat-space program is already producing theorems about the infrared structure of the universe we actually live in. It’s worth noticing that the pessimistic reading is most attractive from the armchair, and the optimistic one is held by the people actually doing the calculations.

The residue survives either way. Whatever the final theory of everything turns out to be, it will have to satisfy the Bousso bound. Which means its information capacity is set by area, not volume. Which means it’ll be holographic in “some” sense. Which means the interior of a region is not an independent thing with its own separate stock of facts.

These are very interesting and possibly unsettling facts about our universe.

Background

Most physicists aren’t all that keen on debating philosophical issues related to quantum foundations, i.e., clarifying what certain concepts in quantum mechanics (QM) actually mean.

(Some physics departments actually go so far as to discourage papers and grant proposals focused on quantum foundations.)

There are also physicists who deny that a problem really exists. But what the “shut up and calculate” crowd, as this group is often referred to, needs to come to terms with is - at the heart of QM, a key pillar of modern physics, are some thorny unresolved theoretical issues.

For example, let’s look at how quantum systems evolve in the textbook formulation of QM (often called the “Copenhagen” interpretation.) There are two incompatible rules as to how the dynamics of a quantum system evolve over time.

Rule one is embodied by the Schrödinger equation. The rule is essentially smooth, deterministic, and reversible. We hand it a quantum state, and it’ll tell us exactly what that state becomes at every future moment. This rule is, as far as we can tell, “exactly true”. We have successfully tested it to great precision.

Rule two is the measurement postulate, or as some refer to it, the so-called “collapse” postulate. When you measure a quantum system, its smoothly evolving superposition (see Rule one) abruptly jumps to a “definite outcome”, with probabilities given by the Born rule. (Explainer: each possible outcome carries a number called its “amplitude.” If we square that number, we get the probability of actually seeing that outcome. That squaring is the Born rule.) This rule is also, as far as we can tell, exactly true. Every experiment confirms it.

The trouble is - these two rules cannot “both” be fundamental. The first says superpositions never collapse. The second says they do, precisely when we “look”. Worse, rule two quietly smuggles in undefined words. What counts as a “measurement”? What counts as an “observer”? A Geiger counter? A cat? A grad student? The math of textbook quantum mechanics genuinely “does not say”. It just tells you to draw a line somewhere between the quantum system and the classical apparatus, perform the collapse there, and not ask too many questions.

This is the “measurement problem”, and it has been sitting at the foundations of QM for roughly a century. If you look closer, it actually breaks down into a few distinct puzzles that people tend to mash together:

1. The preferred basis problem. A “basis” is just the set of distinct alternatives you choose to describe a state in terms of “alive-or-dead”, say, or even using some bizarre “alive-plus-dead” and “alive-minus-dead” options. The unsettling part is that the math treats all these choices as “equally valid”. That is, a single superposition can be written infinitely many ways depending on “which basis you pick”. So “when we look”, why do we always see the “familiar” alternatives - definite “positions” - rather than the weird combinations (e.g., here-and-there, alive-plus-dead, etc.)? What “singles out” the classical-looking description?

2. The definite outcomes problem. Why do we observe “one” result rather than a smeared-out superposition of all of them?

   1. Think of this problem as following on from the preferred basis one. Choosing a basis is like choosing a tree from the forest. Choosing a single outcome is like choosing a branch of that tree. Even if you solve the preferred basis problem re tree selection, you still have the definite outcomes selection re branch selection.

      1. Also, please note that I’m speaking figuratively. No one is doing any “actual choosing”. Nothing in the unitary dynamics does any selecting at all. Rather, we might say the branches all coexist and the question is why we find ourselves on one specific one.

3. The probabilities problem. Where does the Born rule come from? Why does squaring the amplitude to get the probability work, and not some other recipe?

4. The objectivity problem. This one is (sort of) underrated and often not listed as a problem. Why do “different observers always agree” on the outcome? You and I can both check whether the cat is alive without destroying the result, but we always concur (meaning the cat is either alive or dead for both of us). In quantum mechanics, measurement generally disturbs the thing that is measured, so this “universal agreement” is genuinely strange. But it’s also the very thing that makes the classical world feel “solid” and “out there”.

The orthodox response (the Copenhagen tradition, in the textbook form nearly every working physicist actually operates by) is to keep the Schrödinger equation as the dynamics, and then treat measurement, with its definite classical outcomes, as the place where the theory makes contact with the world. This “works” (amazingly) well as a recipe, as it turns out. Its embarrassment is that “measurement” and “the classical apparatus” are left as undefined primitives. John Bell, in his famous broadside “Against Measurement,” wrote: a fundamental theory of reality shouldn’t have the word “measurement” in its axioms, as though the universe came pre-divided into systems and apparatus. What “physical process”, he demanded, actually constitutes a measurement?

My thesis in this essay is that Wojciech Zurek (and others), building on the theory of quantum decoherence, finally gives that question a physical answer, and that this lets us keep the orthodox framework while replacing most of its hand-wavy ambiguity with calculation. Note that this isn’t a “new theory”. Better to think of it more like a “conservative upgrade” to the existing theory we already use in modern physics.

But before I make this case, let’s look at some alternative approaches to addressing the issue. Over the years, there have been several serious attempts to sort out the measurement problem, pursued by serious physicists (and philosophers of physics).

1. The Field of Contenders

Objective collapse theories (GRW, CSL, Penrose). Folks who subscribe to some sort of “objective collapse” approach say that wave function collapse is “physically real” and then go on to write down new dynamics that describe how it happens. The Ghirardi-Rimini-Weber model and its continuous cousin (Continuous Spontaneous Localization) add a tiny random “hit” to the Schrödinger equation that “localizes wavefunctions spontaneously”. The hit rate is set so that a single particle almost never collapses, but a macroscopic object (with its ~10^23 particles) collapses essentially instantly. Roger Penrose has a related idea where gravity triggers the collapse. In any case, within this category of methods, the collapse is built into the equations, and no observers are required. The cost? Well, we end up with a “modified” version of QM. Proponents have had to introduce new physical constants and new dynamics, leading to predictions that differ (ever so slightly) from standard QM. Which is great, actually, because that means it’s testable. The catch? Experiments (underground searches for the tiny, predicted heating, interferometry with ever-larger molecules) keep “not seeing” the effect, effectively squeezing the allowed parameter space tighter and tighter. So, this group of approaches is not dead but is definitely on the back foot.

Pilot-wave theory (de Broglie-Bohm). Also known as Bohmian Mechanics. Here you keep the wavefunction exactly as is, (meaning it evolves by the Schrödinger equation), but you add “actual particles” with “actual definite positions” at “all times”, guided by a “pilot wave”. Nothing ever collapses. The particle was always somewhere - we just didn’t know where. This solves the definite-outcomes problem (positions are always definite after all) and reproduces the Born rule given a “quantum equilibrium” assumption about initial conditions. However, this approach comes with a steep price tag. The guiding equation is blatantly, irreducibly “nonlocal” in a way that sits awkwardly with Einsteinian relativity. Moreover, extending pilot-wave theory to quantum field theory (the basis of the current Standard Model of particle physics) is notoriously thorny. There’s also now a whole second layer of “ontology” (a philosophical term for “what the theory says actually exists”) on top of the wavefunction, which now does double duty as a real physical field pushing those particles around.

Many-worlds (Everett). Upfront, this is actually the most austere approach of all. Basically, we take the Schrödinger equation, declare it the “whole story”, and toss the collapse postulate away. There’s no second rule. When we measure a superposition, we don’t collapse it. Rather, we “join” it. The apparatus, then us, then the lab, then the air molecules all get swept into one giant entangled superposition, and what looks like “collapse” is just us finding ourselves in “one branch” of it. “Every” possible outcome happens, each in its own branch.

There are some good aspects to this approach. In fact, the approach I’m going to advocate shares some machinery with the Everettians. Many-worlds is famously minimal in its modifications to the core QM theory. There are no extra particles, no extra guiding equations, no collapse dynamics, no new constants. It just takes the Schrödinger equation we already use and refuses to add anything to it. Note that it’s this “refusal-to-modify instinct” that I’m sympathetic to. Also, I’d like to point out that modern Everett advocates (David Wallace, Sean Carroll, etc.) lean heavily on decoherence to make their approach work. So, this isn’t a “rival” camp to what I’m proposing below, so much as a sort of neighbor.

The bigger problem is the ontological consequences on the backend of the Everrett approach. To get that lean formalism, (full-blown) many-worlds asks you to believe in an unfathomable, ever-branching profusion of (equally real?) parallel universes, all of which are forever inaccessible to us. So, the formalism itself is austere. But the reality it describes is probably the most extravagant in the history of physics. There are also some technical issues here. The Everett approach, on its own, doesn’t (explicitly) tell us “what” a branch is. The Schrödinger equation gives us one enormous evolving wavefunction (the so-called “wavefunction of the universe”). It doesn’t come pre-labeled with saying “world A here, world B there.” Nor does it (explicitly) hand us the Born probabilities. If “all” outcomes happen, in what sense is one outcome “more probable”?

Hence, my own preference, which is more conservative than any of these approaches listed above. I would prefer an approach that does not add particle locations or guiding equations (Bohm), or add new collapse dynamics (GRW), or commit to a literal multiverse (Everett). I want to keep the textbook QM framework that already works and simply “fill in the gaps” with real, established physics.

2. Why You Haven’t Heard Much About the Zurek approach

Before getting into the core material of this essay, a quick word on a genuine puzzle. If this approach is as promising as I’m claiming, why does it get a fraction of the popular attention lavished on many-worlds and pilot waves? You can buy a shelf of trade books on the multiverse. There’s no equivalent for the Zurek program. Why is that?

I think there are a few reasons - these are based on my observations of the field over the past couple of decades. Note that none of these reasons are “because it’s wrong.”

The Zurek approach isn’t a “flashy” interpretation, so it makes for a worse story. Many-worlds gives us infinite parallel universes. Bohm gives us secret guiding waves. These are “quotable”. Influencers on social media can wax poetic about (or heavily criticize) those approaches. The Zurek approach, on the other hand, gives us... a careful account of why the classical world looks classical. The selling point is mostly “I’ll tell you less about ultimate reality, but I’ll back it all up with calculations.” Kind of a hard pitch for a popular-science story.

A good portion of the Zurek program got "absorbed into mainstream physics so thoroughly it stopped feeling like a position. Decoherence is now a standard part of the QM toolkit. It shows up in quantum computing error analysis, in experimental work, in textbooks, etc. The very success that makes the approach credible also makes it feel like “engineering” rather than “philosophy.” Besides, the QM foundations-and-philosophy media enterprise runs on unresolved drama (or at least it seems that way to me). Many-worlds and Bohm stay newsworthy precisely “because” they remain so contested.

The genuinely bold parts of the Zurek program are either too technical or too new. Envariance (deriving the Born rule from symmetry) is mired in a technical circularity debate that’s very hard to popularize. Quantum Darwinism is recent and was only experimentally confirmed in the lab in 2025. There’s been no breakout popular-science moment or book, and no charismatic popularizer carrying the banner the way Sean Carroll or David Deutsch does for Everett or Tim Maudlin for Bohmian Mechanics.

And advocates of the Zurek program have positioned it as interpretation-neutral infrastructure, which is intellectually honest but… politically weak. A program that says “this is just what quantum mechanics does, no matter your metaphysics” doesn’t recruit any evangelists to the cause. The interpretations that get all the press are the ones with partisans who show up to fight the good fight.

So one could argue that the obscurity is sociological, not evidential. Anyhow, with that out of the way, here’s the actual physics.

3. What Zurek Is Actually Doing

Here’s what makes Zurek’s program a “conservative” fix to orthodox QM rather than a radical one.

Bohm adds particles and guiding equations. GRW adds collapse dynamics. Everett adds many worlds. The Zurek approach is - you don’t need to add anything. You just need to take seriously a piece of physics that was sitting in plain sight the whole time, namely “the environment”.

Every real quantum system is relentlessly, continuously entangling with its surroundings. That is, the system is getting its state tangled up with the state of everything it touches. You can no longer fully describe the one without the other. A dust grain in sunlight is struck by something like 10^11 photons a second. A “measuring device” is thus just a big system soaked in an even bigger environment. How does textbook QM deal with this? Well, it sort of performs a sleight of hand in the original formulation - it treats the measured system as “isolated”, then collapses it by (axiomatic) fiat. But, in reality, nothing is usually isolated! The environment is “always in the room”. Zurek’s approach, pursued since the early 1980s and built on the broader theory of decoherence developed since the 1970s (Zeh, Joos, and others), centers around: what happens if you stop ignoring the environment?

The answer, it turns out, is the sort of physical mechanism John Bell was asking for when he wanted a real answer to the measurement problem. The classical apparatus in orthodox QM that Bohr had to “assume” and the definite-outcome world the Copenhagen framework “presupposes” - both of these emerge from ordinary “unitary” QM once you apply it to system-plus-environment (unitary is just the adjective for the smooth, no-collapse Schrödinger evolution of Rule one I mentioned at the beginning of the essay).

Let discuss the core concepts involved.

4. The Core Concepts

Decoherence and einselection: vindicating the preferred basis

Let’s start with the preferred basis puzzle. Why "position and not some bizarre superposition basis?

The answer is that the environment is constantly “monitoring” certain observables and not others. The interaction between a system and its surroundings has a particular form. It usually depends on position, because forces depend on where things are. That interaction picks out a special set of states - namely, the ones that “don’t get smeared by being monitored”.

Most superpositions, when they entangle with 10^11 incoming photons, instantly leak their “which-state” information “into” those photons (which-state = identity or exact configuration of a quantum system). The delicate “phase relationships” that make quantum superposition a superposition get scattered to the winds. But a few special states are robust. If the system is in one of those special states, the environment scatters off it without disturbing it much, and the state “survives”. Zurek calls these survivors “pointer states” (as in, the states a measuring pointer can actually rest in).

Zurek named the process that selects them einselection (environment-induced superselection). The environment, simply by interacting, acts like a sort of sieve. Fragile superpositions are destroyed almost instantly. Robust pointer states persist. The operational tool for “finding the survivors” is what Zurek terms the “predictability sieve.” You basically run every candidate state forward while coupled to the environment, measure how fast each one degrades (that is, how much entropy it generates), and rank them. The ones that stay predictable (that resist the environmental churn) are your pointer states. For realistic systems, these come out to be the roughly-localized, classical-looking states. So the position basis “wins” not because we put it in by hand, but because that’s what the system-environment interaction selects.

This means that the classical/quantum cut that Bohr drew pragmatically (the dividing line he placed by intuition between apparatus and system) turns out to be a “real physical feature” you can derive. This dissolves the preferred basis problem! And it also answers Bell’s “what counts as classical?” question. Critically, the quantumness itself decays exponentially fast.

(Bit of technical background: physicists track a quantum state as a grid of numbers. Down the diagonal sit the ordinary probabilities: 70% chance of “here,” 30% chance of “there,” the kind of thing a classical coin could produce. The off-diagonal entries are the extra ingredient. They record how the possibilities relate to each other, and they’re what makes a real superposition behave differently from a classical “we just don’t know yet.” When those off-diagonal numbers fade to zero, the quantum state becomes indistinguishable from a shuffled deck, and that fading is what we call decoherence.)

Thus, decoherence is the process of those off-diagonal numbers “shrinking to zero” leaving behind a grid that looks exactly like plain classical odds. And it happens really fast. For a macroscopic object, decoherence times are something like 10^-20 seconds or even faster (see Table 1 on page 11 here). This is vastly quicker than any other timescale in the problem. This is why we never see a superposed cat (dead “and” alive). The superposition isn’t forbidden per se. Rather, it’s just annihilated faster than anything could ever notice it.

Quantum Darwinism: explaining objectivity

Decoherence tells us why certain states survive. But it doesn’t explain why “many independent observers all agree” on the single outcome that emerged.

This is Zurek’s second big idea. Why do we treat the position of the moon as an objective fact? Because the information about it is “redundant”. The moon scatters sunlight in every direction, so its position is imprinted on countless independent packets of photons streaming through space. Any one of us can intercept “some” of those photons and read off the same answer, without anyone’s reading disturbing the moon or interfering with anyone else’s. Thus, the information isn’t held in one fragile place and is broadcast in a zillion copies.

Zurek calls this quantum Darwinism: the environment doesn’t (just) destroy superpositions, it “selects and proliferates” records of the pointer states. The “fittest” states (that is, the einselected ones) are precisely those that can spawn many redundant copies of themselves throughout the environment. Thus, a fact is “objective” exactly when it’s recorded so many times, in so many independently accessible fragments, that everyone who checks gets the same answer.

This essentially reframes the role the environment plays in the framework. In old-school decoherence, the environment is a “sink”, a place where quantum coherence goes to die (as we discussed in the prior section). In quantum Darwinism, it is also a “communication channel.” A witness, so to speak. The environment is the medium through which the classical world “advertises itself” to all of us at once.

The Born rule: envariance

Now, what about those probabilities governed by the Born rule? Here Zurek offers his most contested contribution, envariance (environment-assisted invariance). The idea is - instead of postulating the Born rule, let’s derive it (purely) from the “symmetries of entangled states.” Roughly speaking, when a system is maximally entangled with its environment, certain symmetry operations on the system can be “perfectly undone” by operations on the environment alone. (Read that last line twice if necessary). Zurek argues this forces "equal-amplitude branches to carry equal probability”, and then bootstraps the general case for envariance from there.

I will acknowledge that this is the part of the Zurek program where experts genuinely disagree. Critics of the envariance solution to Born probabilities argue that the derivation implicitly “assumes some notion of probability” to get started, rendering it circular. Now, defenders do argue that it doesn’t. I think it’s best to say that this part of the theory should be seen primarily as suggestive. But let’s also note that “even if envariance is not right”, the other pillars (decoherence, einselection, quantum Darwinism) stand on their own.

Recent developments: Riedel and the branch problem

Now let’s talk about some of the most recent developments related to this approach.

Decoherence tells us the classical world emerges as a set of stable, redundantly-recorded, non-interfering components - the so-called “branches.” But there’s a deep question that physicist C. Jess Riedel asks, most recently in a 2025 piece pointedly titled “Wavefunction branches demand a definition!”

(Technical note: a “branch” here is a term in a decomposition of the global state vector relative to a chosen factorization of Hilbert space into system, apparatus, and environment.)

Riedel’s challenge goes like this: the decoherence story usually starts by “assuming” a split of the world into “system” and “environment.” It also assumes we know which variables are the macroscopic, classical-looking ones. But imagine you’re handed the raw wavefunction of every atom, with no labels. Meaning, nobody tells you which atoms are “the coffee cup” and which are “the air.” Could you write a formal procedure, an actual algorithm, that takes the bare wavefunction and spits out the branches as an output? That is, we are not going to rely on just human intuition about what’s “macroscopic.” We won’t use some sort of pre-chosen system-environment split. We want a procedurally calculated “state in, branch structure out”.

Per Riedel’s challenge, if we can actually do that, we will have something remarkable. We replace the vague, observer-dependent notion of “measurement” with a precise, universal, derivable definition of what the distinct outcomes of any quantum process actually are. And this would work for for the photodiodes in our labs, the whole universe, the aftermath of cosmic inflation, etc. Riedel’s point is that the somewhat abstract problem then becomes a concrete, well-posed math problem.

Now here’s the fun part - there are people making real attempts to actually do this. Recent proposals try to define a branch by how “hard” it is to undo. (Riedel reviews the work of Taylor & McCulloch and Weingarten). The intuition is simple (and elegant). Two parts of the wavefunction count as separate branches when they’re easy to “tell apart” but very hard to “stitch back together” into a superposition. That asymmetry - easy to distinguish, near-impossible to recombine - is then the fingerprint of an irreversible, classical “fact”. (This kind of works the same way it’s easy to scramble an egg and effectively impossible to unscramble it.) One proposal even ties the stability of branches to the general tendency of this kind of “un-mixing difficulty” to keep growing over time, linking the appearance of collapse directly to physical irreversibility.

These attempts aren’t finished, of course. Riedel is candid about the gaps. For example, one proposal offers unique branches but may force them too sharply (Weingarten). Another is better motivated but doesn’t yet provide a unique answer (Taylor & McCulloch). Neither has a clean relativistic version. In any case, progress continues on this particular thread of the program.

So, where does this leave us? Well, the orthodox QM framework pretty much left “measurement” as a black box. Decoherence pried open that box and found an underlying explanatory mechanism. Now, the frontier work is making that mechanism more precise, such that it could be stated as an algorithm.

5. The Empirical Evidence for the Zurek Approach

At the end of the day, we need evidence to back up theories. And this is where the Zurek approach distinguishes itself from most of the other interpretations. Large chunks of it have been experimentally confirmed. Meaning - these chunks are now established physics, not just a stance.

Decoherence has been extensively verified. Since the 1990s, experiments have observed decoherence happening “in slow motion” by engineering systems in which it’s slow enough to time. Serge Haroche’s group (who won the Nobel Prize in 2012) famously watched superpositions of light decohere, photon by photon, in a cavity, measuring the decay of the off-diagonal terms and confirming that larger superpositions decohere faster. Exactly as the theory says! Matter-wave interferometry with large molecules (fullerenes etc.) shows interference fringes vanishing as we allow the molecules to interact more with their environment (emit thermal photons, collide with gas). Decoherence has become textbook, lab-confirmed, engineering-relevant physics. The entire field of quantum error correction exists because decoherence is real and quantitatively well understood. The 2025 Nobel Prize in Physics went to Clarke, Devoret, and Martinis for showing that a superconducting circuit large enough to see can tunnel and can occupy discrete energy levels, behaving as a single quantum object. The same platform became the superconducting qubit, where coherence and its loss are now measured as a matter of routine engineering.

Quantum Darwinism has now been directly observed. This is another recent development. The redundancy-and-proliferation story got its first experimental support around 2019, in photonic systems and nitrogen-vacancy centers in diamond. These showed the characteristic signature: information about a system saturating once an observer captures a small fragment of its environment, with extra fragments just repeating what you already know. That plateau is the smoking gun of redundant encoding. (Something to note: several of these experiments are best described as “quantum simulators” of the effect, i.e. engineered systems that realize the Darwinism dynamics rather than a naturally decohering object caught in the act, a distinction worth noting.)

The most comprehensive demonstration to date came in 2025, when Zhu, Salice, Touil and collaborators used a superconducting quantum processor to probe the effect. Where the earlier experiments had only detected information-theoretic signatures in narrow settings, this team was able to map the branching structure of the global state directly. They were able to watch the mutual information saturate as predicted and confirm the geometric picture that underpins the framework. The system is still an engineered one, of course, rather than an everyday object decohering in the wild, but it is the closest anyone has come to seeing the whole mechanism at once.

The implications are fascinating - as long as observers eavesdrop on a suitably large fragment, they always agree on their conclusion if they’re in the same branch, illustrating how classical reality emerges from a structured quantum universe!

Now, this does not prove that Zurek’s view is the “uniquely” correct one. In fact, I’m not sure that experiment could really “settle” a question of interpretation. What it does prove is that the “mechanisms” the Zurek program is built from (einselection, redundant proliferation, branch structure, etc.) are “real, physical, and measurable”. This makes the program stand out when we look at the alternatives. Bohmian particles? Never observed (and by construction unobservable?). GRW collapses? Searched for, never seen. Everett’s other worlds? Inaccessible in principle. But decoherence and quantum Darwinism? We have empirical data that confirms important aspects of the framework. Since the goal here was a “conservative fix”, the Zurek framework does that by (primarily) leaning on mainstream confirmed physics rather than introducing new (speculative) metaphysics.

6. The Strongest Objections

Let’s also look at the some of the main criticisms of the Zurek program. I’m going to group these into two buckets - what the Everettians would say (Sean Carroll, David Deutsch, David Wallace, etc.) and what realists like Tim Maudlin might say.

The Everettian objection: “You’re just using the Everett framework and refusing to admit it.” (I’m paraphrasing here, but I think that’s a fair summary of one of their main points.) A committed Everettian would say this whole “conservative upgrade” framing is a dodge. Look at what we’ve to actually commit to in the Zurek framework. We have “the Schrödinger equation is complete, it always evolves unitarily, nothing collapses”. Well, those are the defining characteristics of the Everett program! If the global wavefunction never collapses and the other branches genuinely contain observers seeing definite outcomes (which Zurek and decoherence says they do), then those branches are '“as real” as the many-worlds ones. Declining to call them “real”, from the Everettians’ perspective, shouldn’t be seen as “conservative”, but rather a refusal to fully outline the (Zurek framework) ontology while helping oneself to all of Everett’s machinery.

My response: Ouch. Well, the Everettians are right that the Zurek program maintains a good chunk of the Everett program’s upfront framing (no collapse, the wavefunction as the complete story, etc.). I think they are also right that this puts the Zurek program in the “same broad family” as Everett. But I think the difference comes down to what the Zurek program is “obligated” to assert. The Everett folks think taking the formalism seriously “forces” us to grant “full reality to every branch”. I think that’s an extra metaphysical step, not a logical entailment. The physics (decoherence, einselection, redundancy) is identical whether or not you make that step, because no experiment can reach the other branches to confirm or deny their “reality.” So we take the conservative path: assert what can be empirically verified, stay quiet about what can’t. But yes, I’ll concede the main framing point: this Zurek view is indeed sort of “Everett-adjacent”.

The realist’s (e.g. Maudlin) objection: “Decoherence does not produce definite outcomes, and you’ve admitted as much.” Maudlin has actually said something exactly that like that. And we’d have to concede that he’s right. But, let’s take a look at the difference between two situations that look identical on paper. First - let’s look at a genuine “mixture” scenario. That is - the cat really is either alive or dead. Well, one of them for sure, we just don’t yet know which (kind of like a coin already flipped but still covered by your hand). Second, let’s take a look at “superposition” scenario. That is - the cat is in the strange both-at-once state that has no classical counterpart. What decoherence does is make a superposition “look”, on paper, exactly like an ordinary “we just don’t know which” mixture, by driving those off-diagonal numbers to zero. Now the rebuttal to this would likely be that “looking identical to” a real mixture is not the same as “being one”. The realists would say - the cat is still “formally” in the both-at-once state. Decoherence has merely guaranteed that the two possibilities can no longer interfere with each other, not that one of them has actually become the case. Maudlin would likely invoke his famous argument here (which I like quite a bit) that the measurement problem just “is” the inconsistency of three claims: 1) the wavefunction is the complete story, 2) it always evolves smoothly with no collapse (unitarily), and 3) measurements have single determinate outcomes. Maudlin says you always have to deny one. So any sort of “conservative upgrade” that touches none of those still has the contradiction.

My response: Got to concede the central point to Maudlin here. Decoherence does “not”, by itself, deliver a single determinate outcome. But let me state the actual claim precisely - decoherence and quantum Darwinism don’t dissolve the measurement problem, but they “maximally narrow” it. Of the four puzzles I opened with, three (preferred basis, objectivity, and, more tentatively, probabilities) get genuine physical answers from the Zurek program. What’s left is the bare determinacy question, why “this” particular outcome is experienced rather than the superposition, and we have to concede that this remnant is basically Maudlin’s “deny one of the three.” But I will also assert that the Zurek program shrinks the trilemma down to a single, sharply isolated, “empirically inert” question, while explaining everything around it with confirmed physics. Maudlin will then say probably “inert is not solved,” and he’s right.
Riedel concedes something along those lines too: even a full branch definition “would hardly eliminate all the mystery of the measurement problem.” But the point stands that the Zurek program has made more progress than any other approach.

(A note about claims being “empirically inert”: Imagine a claim that a tiny, invisible gremlin is hiding in your room. If this gremlin makes no noise, leaves no footprints, cannot be felt, and interacts with absolutely nothing, it is “empirically inert”. Because you can never run a test to prove the gremlin is not there, the idea is untestable.)

Maudlin’s likely follow-up: “Objectivity is answering a question nobody’s measurement problem was asking.” He’d also likely say: quantum Darwinism is a nice result about “information structure”, but it’s orthogonal (irrelevant) to the real problem. Observers agreeing with each other presupposes there are outcomes for them to agree on, and that’s the very thing in dispute.

My response: I think the criticism undersells what’s been accomplished. Sure, objectivity presupposes definiteness rather than producing it. But the objectivity problem is a “legitimate explanatory target” in its own right. I’d also like to point out that Bohmian Mechanics and GRW would also have to explain why records across the world agree, and the answer isn’t clear at all! Quantum Darwinism is a genuine, now experimentally confirmed contribution to “that” question. (So this was more like a successful side-quest.)

The realists’ parting shot: “The fact that you still can’t define a ‘branch’ after sixty years is a symptom, not a frontier.” The realists would read the Riedel review material very differently. If the approach can’t even non-arbitrarily say “what the outcomes” are without an after-the-fact algorithm to dig them out of the universal wavefunction, that, they would say, is evidence the framework tries to carve structure that isn’t fundamentally there. Whereas some alternatives (a precise microphysics, Bohmian corpuscles, GRW mass-density, etc.) would tell you what exists up front, no extraction procedure required.

My response: If you already believe the fundamental furniture of the world should be specified up front, then yes, needing an algorithm to “locate” the branches looks like an admission of failure. But if you think macroscopic structure is “emergent” (like temperature, fluid flow, etc.), then needing a principled procedure to extract that structure from the microphysics is completely normal. We do that for every other higher-level science. Nobody calls thermodynamics ill-founded for doing that. So whether the branch-definition problem reads as “symptom” or “promising frontier” depends on a prior commitment about emergence that the physics alone won’t settle. I’m going to go ahead and say that this is a promising frontier. Because, the tools now being used to tackle the problem (quantum circuit complexity, approximate error correction, etc.) genuinely did not exist when the measurement problem was originally posed. So “we have a precisely stated problem and powerful new tools aimed at it” is a very different situation from “we have an incoherent muddle.”

7. Open Issues: Sorting Philosophy from Physics

Okay. Let’s look at some genuinely open questions that the framework is still addressing. I’ll split this up into what I think is merely philosophical vs. actually physics.

The philosophical residue (where physics arguably can’t help)

After einselection picks the basis, decoherence kills the off-diagonals, quantum Darwinism makes the outcome objective, and envariance hands you the probability weights, one thing remains: the global state is still, formally, a superposition of all the branches. The residue, as we discussed in the prior section, is the bare question of why one outcome is 'experienced as actual.

Now you might ask - why I am being dismissive and filing this under under philosophy rather than physics? Well, the physics is “identical” no matter how we answer it. Every branch contains the same einselected, redundantly-recorded, Born-weighted structure whether or not we decide the other branches are “real.” No experiment can distinguish “the other branches exist” from “they don’t,” because any experiment we run happens “within” a branch and sees the same thing either way. Embrace the branches (The Everett path) and the question dissolves into “why am I me?”. Reject the branches and we (have to) posit a brute fact about which branch is actual. Either way, more decoherence calculations will not settle it. Gonna call one this philosophy and leave it to taste.

The real physics that’s still open

Now the part where there’s unfinished physics.

1. The branch-definition problem. As discussed, we don’t yet have a rigorous, universal, observer-free definition of a branch derivable from the bare wavefunction. We have promising complexity-based proposals with real gaps (the Riedel summary I discussed previously outlines these). None are yet simultaneously unique, well-motivated, and relativistic. How it might get resolved: we should pay attention to the quantum-complexity approaches. If someone shows the “easy to distinguish, hard to interfere” criterion gives a unique decomposition that (a) forms a proper tree in time, (b) recovers the known decoherence of ordinary hydrodynamic variables, and (c) survives a relativistic limit, that would essentially '“complete” the program, converting “measurement” from a primitive into a theorem about wavefunction structure.

(Re hydrodynamic variables, what I mean is - a natural test for the program’s ambitions is whether it can account for the classicality of hydrodynamic variables. Gell-Mann and Hartle argue that the classical variables of a many-body system are the local densities of conserved quantities, energy, momentum, particle number, integrated over small volumes. Zurek’s framework locates classicality in coupling to an environment instead. Both approaches work. Both are plausible. Whether they are two faces of one mechanism or two distinct routes remains open, and settling it would tell us how general einselection really is.)

2. Pointer states in the hard regimes. Einselection works beautifully when the system-environment coupling dominates. But there are regimes where “which states get selected” becomes subtle and timescale-dependent. Examples of these scenarios include when the system’s own internal dynamics compete with environmental monitoring, or when it starts in a mixed state. Recent work, notably by Sebastian Deffner’s group, has been turning the loose “pointer states usually exist” into rigorous, checkable conditions, by providing precise algebraic criteria for exactly when objectivity emerges for a broad class of two-body interactions. How it might get resolved: this looks like a tractable classification program, that might be plausibly wrapped up (in the sense of “we know the necessary and sufficient conditions”) within a few years.

3. Making objectivity fully rigorous. Zurek’s original quantum Darwinism used a quantity called quantum mutual information as its measure of “how much the environment knows” about the system. (That’s just a QM analog of Shannon mutual information.) J. K. Korbicz and collaborators pointed out, constructively, that mutual information is actually “too weak” to guarantee genuine non-disturbing objectivity on its own. They developed a sharper notion called “Spectrum Broadcast Structure” (SBS) that demands the global state literally factor into perfectly distinguishable records across observers. Effectively, when a state has SBS form, objectivity becomes a “theorem”. This would be a strong self-correction within the research program. Status: nearly resolved on the theory side; the remaining work is showing SBS emerges generically rather than in hand-picked models.

The Bottomline

So where does this leave us?

What Zurek offers (augmented by Riedel, Korbicz, Deffner, and a growing community) is, I think, the most attractive resolution on the table to address some of the issues within QM we laid out at the start. It’s attractive precisely because of “how little it asks us to add” to conventional QM. If we want to keep the textbook QM framework that every physicist already uses, whose primary embarrassment was leaving “measurement” undefined, and patch that hole with physics we’ve confirmed in the lab - then the Zurek approach is the best way forward today.

Decoherence is real and measured. Einselection is calculable. Quantum Darwinism was directly observed in 2025, with multiple observers provably driven to agreement.

Final scorecard: the preferred-basis problem is “solved” (einselection). The objectivity problem is “solved”, and now experimentally demonstrated (quantum Darwinism, sharpened by SBS). The probability problem is “plausibly addressed but contested” (envariance). The definite-outcomes problem splits cleanly into a derivable structural part (branches, with real open physics) and an empirically inert metaphysical residue (which-branch-is-mine, where people are free to choose their poison).

Now, the Zurek framework is not a complete theory (yet). But this is just the incompleteness of competing, irreconcilable philosophical approaches. And more importantly, it’s just the ordinary incompleteness of an active research program with well-posed problems and scientists making measurable progress on them.

When the dust finally settles on the measurement problem, I’d bet that the resolution looks “less” like a brand-new collapse mechanism or a hidden layer of particles, and “more” like the orthodox QM framework augmented by the Zurek program.

So, Eureka is too strong. But Zureka!? I think we’re allowed that one.

Thanks for reading! Subscribe for free to receive new posts and support my work.

In those essays, I was primarily making a “philosophical” case. In this essay, I try something different. If the physicalist story is right, then there should be an actual history - a real, datable, fossil-and-genome-grounded sequence of events by which chemistry bootstrapped itself into “something that has a point of view”. This is an attempt to summarize some of the key points in that evolutionary history.

I’m going to borrow a narrative device from one of my favorite popular science books - Steven Weinberg’s The First Three Minutes, where he narrates the history of the (very) early universe as a sequence of timestamped frames (”it is now one hundredth of a second after the beginning...”). Weinberg’s clock ran in fractions of a second. Evolutionary history runs in hundreds of millions of years. I’m going to take as the zero point the origin of life, somewhere around 3.7 to 4 billion years ago. Everything below is measured forward from there.

(So the timeframes in the section headings “count up” from the origin while the calendar dates in the text “count down” toward today. So “T ≈ 3.48 billion years” means approx. 220 million years ago. The two should always reconcile. If they don’t, I’ve fumbled some very basic arithmetic. Welp.)

Before we jump into the essay proper, I’d like to point out some background material. There’s a decent-sized shelf of recent books by serious scientists trying to do exactly what I’m attempting here. Which is - give consciousness an evolutionary natural history rather introduce some sort of speculative metaphysics. Here are some of my favorites.

Todd Feinberg and Jon Mallatt’s The Ancient Origins of Consciousness (and its follow-up, Consciousness Demystified) builds the most explicit fossil-grounded timeline. Antonio Damasio’s The Feeling of What Happens roots the whole thing in the body and homeostasis. Nicholas Humphrey’s Sentience: The Invention of Consciousness argues (somewhat provocatively?) that the lights came on “very late”. And Daniel Dennett’s From Bacteria to Bach and Back rounds out some of the philosophical gaps in the gradualism - discussing how capacities accrued by degrees with no magic threshold. I rely on all of these books (and more) to varying degrees in this essay. I should note that in some cases, these authors disagree with each other about “when on this journey the lights come on”, but this is to be expected in a developing scientific subfield.

I also want to note that the some of the branches of this tree are contested. The fossils “are what they are”, and in some periods, the record can be sparse. Sometimes, molecular clocks are less precise than we’d like them to. Also, reasonable people argue about the order of events. (I’ll try and flag some of these disputes.) But… the “shape” of the story, which is one depicting increasing integration, increasing modeling, increasing behavioral flexibility, etc. - all of that tracking the elaboration of biophysical machinery - this is not in serious doubt.

Okay. So let’s start the clock.

T = 0: A Chemical System That Cares

It is now roughly 3.7 billion years ago. There are no neurons or nerves. There isn’t even multicellularity. There are just cells. We have the first self-reproducing, metabolizing chemical systems, traceable in an unbroken line of cell divisions to every living thing today.

Let’s look at one key feature here. Even at this stage, cells are not “passive bags" of chemistry”. They have “boundaries” they maintain against entropy. They have an “inside” they keeps distinct from an “outside.” They contain a set of processes that have to keep running or they die. And, by this point, we have the bare minimum necessary to give the words “stimulus” and “response” some meaning. A rock does not “respond” to its environment in any sense that requires the rock to persist. A cell does. This asymmetry, that there is something the system is “for” is all about cells “staying” alive.

I want to point out this we don’t need some inner (magic) light to make this happen. A cell maintaining homeostasis is essentially doing systematic thermodynamic management. Moreover, it’s the “kind” of thermodynamic setup that has a built-in evaluative structure. There is effectively a “good for me / bad for me” evaluative process going on in the cell. This structure is the raw material that natural selection will spend the next 3.5 billion years sculpting into something that writes essays about itself on Substack.

T ≈ 1 Billion Years: First Decisions

It is now about a billion years after the origin of life. Unicellular life has evolved something we can sort of call “behavior.” That is - these single-celled organisms have the ability to “sense a gradient” in the environment and “bias its movement” accordingly.

The canonical example (and probably one of the most closely studied behavior in all of biology?) is bacterial chemotaxis. An E. coli cell swims by alternating two modes. It engages in smooth “runs” and reorienting “tumbles.” Of course, the cell is too small to measure a concentration at two points in space simultaneously. So it does something cooler. It measures concentration “over time” as it swims, and compares the present to the recent past. If things are getting better, it suppresses tumbling and keeps going. If not, it tumbles and tries a new direction.

To compare present to past, the cell needs to “store” the past. That is, it needs a molecular memory, implemented in the methylation state of its receptors. And the timescale of that memory turns out to be evolutionarily “tuned.” Meaning, it’s optimized to make the “most useful” comparison given how fast the cell swims and how gradients are distributed in its world.

This is good progress already. The very first behavior in our story already contains (in miniature form) three ingredients that every nervous system will eventually scale up in one way or another - sensing, memory, and action selection.

We can even watch this capacity evolve in simulations. If you “grow” populations of “virtual bacteria” with mutable signaling pathways and let selection run, we end up with chemotactic response dynamics reliably “emerging”. And.. the “type” of dynamics that evolves depends on the statistics of the environment. Meaning, sparse, fluctuating worlds select for more complex adaptive responses.

Now, there’re some scientists who want to go further and call this level of “behavior” cognition (“basal cognition”) or even sentience (this is the “senomic” crowd). I am skeptical of these arguments. Calling chemotaxis “cognition” is fine as long as it’s understood as a claim about “function” (sort of like information-guided action selection), and not some sort of claim about “cognitive experience”. There is no reason whatsoever to think an E. coli cell has a point of view. What it has is essentially an optimized feedback loop. So we should recognize the long-term functional continuity here (meaning this is the lineage that will eventually lead to minds) while refusing any conflation with cognitive experience. (So, no, there is nothing it is like to be a bacterium tumbling up a serine gradient).

In From Bacteria to Bach and Back, Dennett emphasizes this point, insisting on making the distinction between “competence” and “comprehension”. The bacterium is fabulously “competent.” We see that it can tackle a real navigation problem. But it does not have even a shred of “comprehension.” The bacterium has no idea it’s doing anything. Dennett’s slogan on this is nice and pithy - evolution is a process that produces “competence without comprehension.” And as we’ll see, comprehension itself, when it finally shows up billions of years later, is just a particularly “elaborate stack of competences”.

Once we bear that distinction in hand, I think the temptation to see a teensy little glimmer of mind in the chemotaxis pathway evaporates. We’re just looking at competence. Comprehension (and experience etc.) are things that get built much later, as one of many competences piled up and wired together.

T ≈ 2 Billion Years: The Eukaryotic Upgrade

It is now roughly two billion years after the origin of life. Eukaryotes show up somewhere in the window of about 1.6 to 2.1 billion years ago (Note that the exact date is contested by specialists in the field). Many of you already know this, but the way it comes about is quite amazing. Essentially, something “extraordinary and contingent” happened. One prokaryote ends up living inside another, and instead of being digested, it hangs around for the long term. (Stories about guests who overstayed their welcome and would not ever leave come to mind.)

The engulfed cell becomes the mitochondrion. The result is the eukaryotic cell. (This happens at least once again later on when plant eukaryotic cells add chloroplasts.) Bigger, more compartmentalized, and a much bigger energy hog! Cognition, even in its most primitive forms, is “metabolically expensive”. Sensing, signaling, memory, and (later) neurons all cost energy. You don’t get to build a brain (which in humans burns about a fifth of the body’s energy but is only roughly 2% of its mass) without first solving the energy problem. The mitochondrion is where that gets solved. Almost everything cognitively interesting that follows from here on is partly a story about how to manage an effective and efficient energy budget.

T ≈ 3 Billion Years: Many Cells, One Body

It is now about three billion years after the origin of life - let’s call it around 700 to 800 million years ago in calendar time. Cells have started staying together after division, specializing, and dividing labor. Multicellularity has arrived. Note that this was not a singular event like the case of the endosymbiotic one that led to eukaryotes. Multicellularity evolved independently many times over - in plants, fungi, several lineages of algae and, of course, animals. The branch we care about is animal multicellularity, which traces to a colonial flagellate ancestor (these are our closest living single-celled relatives and look like the collar cells of sponges). But now, with the animal body, comes a new problem that will turn out to be “the problem” for the rest of our narrative. That problem is “coordination”.

A single cell senses and acts as one unit. A multicellular animal has separate cells in different places facing different bits of the environment, and they need to act as a “single agent”, not a mob. Take this scenario - if the cells on the multicellular organism’s left detects food and the cells on the right detect a predator, “the organism” need to do “one coherent” thing, not tear itself in half. How did evolution sort this out? Well, it now starts selecting for machinery that “binds distributed sensing into unified action”. I mentioned integration problems in my hard problem essay as the functional core of consciousness. Well, here it is - showing up, for the first time in this timeline, and serving as a concrete engineering constraint on a blob of cells. Of course, this is hundreds of millions of years before anything we’d call a mind.

Antonio Damasio points out that there’s a second thing the multicellular body adds, which is that the body now has an “interior” to regulate. Basically, a teeming internal set of cells that has to be kept within the narrow envelope of conditions “compatible with staying alive”. Damasio’s argument is that this “homeostatic imperative” is the true “root” of mind. Long before there’s “perception of the world”, there’s “monitoring of the self.” That is, the organism ceaselessly tracks its own internal state, and the valenced “this is good / this is bad for my continued existence” signal that tracking generates. In Damasio’s view, “feeling” was originally the felt “registration of the body’s own condition.” The outward-looking mind is built on top of that inward-looking foundation. Now, whether or not you buy the strong version of this claim (that you can’t have any experience without this visceral substrate), the structural point is a good one and will matter later. The nervous system that’s about to evolve over time does NOT arise to “contemplate the world”. It arises to “keep a body alive” in the world. Cognition originated to satisfy the requirements of homeostasis.

T ≈ 3.1 Billion Years: The Invention of the Neuron

It is now somewhere around 600 to 650 million years ago. The solution to the coordination problem arrives, and we could say that this is the single most important event in this entire history: the evolution of neurons.

A neuron is a cell specialized for fast, targeted, long-distance signaling. It can carry information from where it’s sensed to where it’s needed. And it does this quickly. And it delivers information to a specific destination rather than broadcasting it. Once you have neurons and synapses, you can build “neural circuits.” .

We do have an unresolved question here: did neurons evolve “just once”, or more than once? The traditional answer to this used to be a sort of tidy ladder: sponges (no nervous system) → cnidarians (nerve nets) → everything else, with essentially one origin. The genomic era complicated that quite a bit. Comb jellies (ctenophores) come out in a number of analyses as the earliest-branching animal lineage, earlier even than sponges. There’s a 2023 chromosome-level synteny study that’s the most recent heavyweight review on that side, though the sponge-sister idea still has serious support. In any case, the ctenophore nervous system is genuinely weird. It’s built partly from a fused syncytial nerve net and using a different molecular toolkit than the rest of us. This has kept alive a serious hypothesis: that neurons evolved “independently” in ctenophores and in the lineage leading to cnidarians-plus-bilaterians. This would then be a case of convergent evolution of the most consequential cell type in the animal kingdom. The competing story is the older one: a single origin in the last common ancestor of animals, with later loss in sponges and placozoans. The debate is still ongoing.

I find the convergence hypothesis fascinating for a reason that’s directly relevant to my broader argument. If neurons really did evolve twice, that really does suggest that nervous systems are an engineering solution that physics and selection will reach for whenever the “coordination problem” gets hard enough. We already know this happens at the level of organs. For example, eyes have evolved independently dozens of times. Thus, the camera eye of an octopus and the camera eye of a human are a textbook case of convergence. (Note: the optics did indeed converge independently. But the genetic switch that says “build an eye here“ (the Pax6 regulatory control gene) and photoreceptor proteins themselves are conserved across all of these lineages. So even the convergence is convergence “built from a shared toolkit!”) These types of structures are essentially “attractors” in design space. They evolve because they work. (This is why I don’t lose sleep over the panpsychist’s “no consciousness from non-consciousness” intuition. Function emerges from arrangement all the time in biology.)

One more thing sponges teach us (by counterexample). Sponges have no neurons. But, they do have cells that express a large number of “presynaptic” genes (molecular parts list of a synapse). These cells make contact with other cells that express the “postsynaptic” genes. So evolution didn’t invent the neuron from scratch. Existing secretion and signaling machinery that was already lying around was repurposed (the core secretory apparatus apparently goes all the way back to yeast). This is just how evolution always works - it involves tinkering rather than engineering. (This is worth remembering whenever someone insists that some biological capacity is too special to have a gradual, mechanistic history.)

T ≈ 3.15 Billion Years: Bodies That Move With Purpose

It is now the late Ediacaran, around 555 to 580 million years ago. And we start seeing animals in the fossil record.

The earliest large animals, from around 580 million years ago, are “sessile”. That is, they sit on the seafloor with a branching, frond-like form unlike anything we see today. They don’t go anywhere. (Slackers.) But by about 555 million years ago we get the first trace fossils of animals that “moved” - the dickinsoniids, flat oval creatures that left successive resting marks. This is essentially an animal repositioning itself to feed.

(I should note, and this took me a while to figure out, that some of the specifics here are not settled. Dickinsonia‘s symmetry is a glide reflection rather than true bilaterality. In any case, the main claim that we are looking at here is is quite modest. That is, by the late Ediacaran, “something” was moving itself around under its own control.)

Self-generated movement is a bigger deal for the whole cognitive story than it sounds. The moment an animal can move itself, it faces a problem that sessile things never face. It has to tell the difference between sensory changes “caused by itself” (by moving) and sensory changes “caused by the world”. The world looking different because “you turned” versus because “a predator moved” are very different facts. The solution every later nervous system lands on is called the reafference principle. This was worked out by von Holst and Mittelstaedt in the 1950s. The idea is to use a copy of each motor command and use it to “predict” (and cancel out) the expected sensory consequences of your own movement. (Read that last line a couple of times!)

Now, we’re not saying that an Ediacaran flatworm had the circuitry for tracking reafference in any developed form. That’s coming down the line. But once animals being to move, that’s when this starts becoming worth investing in as a capability. This can also be seen as the origin point of what Anil Seth (and others who subscribe to the predictive-processing model which I discussed in the hard problem essay) talk about. The brain acts as a prediction engine, continuously modeling the causes of its sensory signals. Of course, timeline-wise, we don’t have brains yet. It’s just that the evolutionary pressures that eventually demands that sort of capability have come into play. All because something on the Ediacaran seafloor needed to know whether “it moved or the world did”.

Comparative genetics tells us the machinery for this by this point was already substantial. When we look at the deep similarities in how nervous systems develop across bilaterians, what we see is that the last common ancestor of bilaterians (living somewhere in this time window) already had a centralized nervous system with an anterior concentration of sense organs and interneurons. (Molecular clocks put bilaterian origins at 600–700 Mya). Taking a look at what’s getting realized here, we have a “front end”, with some sensors and related processing clustered “up there” near where the animal was going. Well - that’s “a head'“! The evolutionary “decision” to put the brain near the sensors at the leading edge of a moving body is one that shows up fairly early (and never reconsidered).

(Note: whether the bilaterian central nervous system itself has a single origin is (like the neuron origin question), actively debated. The evidence seems to favors a single origin, but a hemichordate-based alternative argues for independent origins in protostomes and chordates. This pattern of “is this feature convergent or conserved?” is a recurring one. Often, the honest answer from the field is usually - “we’re still working it out.”)

T ≈ 3.16 Billion Years: The Cambrian, and the Birth of Seeing

It is now about 540 to 520 million years ago. The Cambrian explosion is underway. The Cambrian explosion is probably the most famous evolutionary period (in the popular consciousness) besides the Mesozoic periods featuring dinosaurs. While there is some continuity of complexity from the Ediacaran, what we get here is the abrupt appearance of most major animal body plans. And (one could argue that) the single most important cognitive innovation of this period is vision. And by that, I mean real, image-forming, spatial vision.

The Cambrian apex predator Anomalocaris (a type of radiodont) had compound eyes up to three centimeters long, each containing more than 16,000 lenses. These are among the most acute eyes that have ever existed, sharper than most living arthropods. And, as you might expect, its prey were keeping up too. The early arthropod Cindarella (with an ‘a’ not an ‘e’ like the fairytale) had stalked compound eyes with over 2,000 ommatidia, good enough to spot a predator’s approach.

We should take a moment here recognize that vision really changed everything cognitively. A chemical gradient only tells you “food is somewhere in that general direction.” An actual internal image tells you “there is a specific object, at a specific location, of a specific shape, and it is getting closer.” This type of image-forming vision, which we take for granted as a feature of most animals, presents the world to the organism as a structured field of distinct things with spatial relationships. Processing that flood of sensory information requires some serious neural hardware. We can see this in the fossil record. The task of wiring tens of thousands of optic axons back to the brain was really demanding. And it apparently drove the evolution of myelin-like insulating sheaths to keep the signals fast and clean. The “need to see” created a data-processing problem, and the brain grew to solve it.

The standard “Light Switch” hypothesis (championed by Andrew Parker) holds that the evolution of eyes was a “catalyst” for the Cambrian explosion itself. Once you have one lineage that can see well enough to hunt visually, everyone else faces ferocious selection pressure. Both predators and prey were locked in an arms race - you need to see better, or be eaten. Hide better, move faster, react quicker.

I think this is the single best illustration of the claim in this essay that “cognition is an adaptation”. Brains didn’t get more sophisticated because the universe was unfolding some latent mental potential (or by tapping into some sort of proto-consciousness). They got more sophisticated because there were “other animals trying to eat you.” And animals with better internal models of the world ate more and got eaten less.

Here, I’d like to reference one of my favorite books on the evolution of consciousness. Feinberg and Mallatt, in The Ancient Origins of Consciousness, argue that the Cambrian’s complex brains brought with them the first “sensory consciousness”. Or should I say - the most basic version of “something it is like to be” - the unified inner world of subjective experience. F&M describe this as happening roughly 520 to 560 million years ago. Their reasoning is structured as follows. Lets take a look at the neurobiological features that we’re confident accompany consciousness in animals, namely elaborate hierarchical sensory processing, isomorphic neural maps of the body and the world, a certain depth of integration, etc. Then let’s look for when those features first appear in the fossil and comparative record. The answer appears to be that image-forming vision and the brains to process it are the key evolutionary signature. And, equally importantly, it happened more than once - independently in the vertebrates, the arthropods, and the cephalopods. So we are running into the evolutionary convergence phenomenon again. Except now it’s not just applied not to eyes or neurons but to “experience itself”. If F&M are right, sentience is just one more attractor in design space. It’s just something the type of thing that natural selection reaches for when the problem is hard enough. This is also about as anti-panpsychist a conclusion as one could draw. Thus, octopuses are conscious NOT because consciousness was “lurking in its quarks and electrons” (or in some mysterious field), but because evolution by natural selection resulted in the neural machinery to achieve important functional characteristics.

I should note that this “early evolution” (timeframe-wise) of consciousness in the Cambrian period is not something everyone agrees on, even if all they agree on the evolutionary origins of consciousness. For example, Nicholas Humphrey, in Sentience, looks at the very same evolutionary tree and reaches nearly a different conclusion. Humphrey feels that “phenomenal consciousness” (the felt, qualia-laden kind) is a “late and narrow” innovation, requiring recursive feedback loops in the brain that monitor the organism’s own responses. More importantly, he thinks it’s present only in warm-blooded mammals and birds. Meaning the Cambrian octopus and the bee, for all of their competence, are “unfeeling zombies.” He even ties the threshold to warm-bloodedness itself - which would push the real birth of feeling forward by hundreds of millions of years.

So who’s right about when this happened? F&M’s “Cambrian + everywhere the wiring is rich enough,” or Humphrey’s “much later + only in warm blood”? This being an evolving field, the best thing to say is that we don’t know. But, for the purposes of this essay. we don’t need to! The argument I’m making is about the types of “physical mechanisms” that are involved in achieving this result (image-forming sensory hierarchies, recursive self-monitoring loops, thermoregulated fast feedback etc.). Thus, it’s an argument about “thresholds in neural architecture”. Or to put it in even simpler terms, the argument is that there is no need to reach for any non-physical ingredients. The dispute between the F&M and Humphrey accounts is essentially over “engineering specifics.” That is, they disagree about which arrangement of neurons is sufficient for the “lights to come on.” But the entire discussion is nevertheless conducted in the vocabulary of evolution, comparative neuroanatomy, and the fossil record. As I wrote in a prior essay, that is exactly what a hard problem of neuroscience, as opposed to a hard problem of metaphysics, looks like.

T ≈ 3.2–3.36 Billion Years: Vertebrate Brains and the Move Onto Land

It is now around 500 to 340 million years ago. Vertebrates show up in the fossil record. They appear to have a basic three-part brain plan (hindbrain, midbrain, forebrain) similar to us. Also, towards the end of this period, around 340 million years ago, some amphibians evolve the amniotic egg, with a membrane that lets it be laid on land without drying out. This “frees” the early amniotes from needing to be by water. As a result, they split into two great radiations: the synapsids (leading eventually to mammals) and the sauropsids (leading to reptiles and birds).

I’m going to move quickly through this stretch of time. Because, the story is essentially just more of the same. We see progressively more elaborate nervous systems supporting progressively more flexible, integrative behavior. I described this type of gradient across the (living) animal kingdom in my panpsychism essay. Fish, then amphibians, then the amniotes, each show up as being more behaviorally flexible than the last. The reptilian forebrain has developed a dorsal cortex with three layers. It’s three now, but it’s about to become six.

(Note: the precise homology between the reptilian dorsal cortex, the avian pallium, and the mammalian neocortex is a long-running dispute in comparative neuroanatomy.)

T ≈ 3.48 Billion Years: The Six-Layered Sheet

It is now about 220 million years ago, in the synapsid line, and now we have one of the genuine landmark events in the history of cognition. We see the evolution of the six-layered neocortex. This is a structure that exists exclusively in mammals.

Essentially, the three-layered cortex of the amniote ancestor gets elaborated through changes in how neural progenitor cells divide and migrate during development. The result is a six-layered sheet.

The specifics of how exactly this setup works is still a work in progress. One influential idea is that the neocortex amounts to a single canonical local circuit. That is, it’s a sort of a “cortical column,” repeated across the sheet. Each “copy” runs roughly the same computation on different inputs. The strong version of that idea (meaning the column is a real, discrete, uniform module) is contested. Lots of neuroscientists apparently think “the column” is more just structural bookkeeping than biology. In any case, my argument in this essay only needs the weak version, which isn’t in dispute. That argument is - the neocortex is built from a “repeated local motif.” Essentially, the mammalian brains has acquired more cortical computing power (mostly) by tiling more of the same, rather than inventing new machinery each time.

This is an important architectural characteristic. This idea of “a generic, repeated computational module, tiled across a sheet, scalable by just adding more tiles.” means you can get more cognitive power largely by just growing the sheet. Evolution essentially found a design that scales. (This has some interesting analogues in the contemporary discussion around AI and GPUs/compute scaling.)

Let’s go back and take a look at Nicholas Humphrey’s claim that “phenomenal” consciousness (the feels like part) is late and warm-blooded. Mammals are by definition the warm-blooded synapsids. (Birds independently arrived at endothermy on the other branch). In Humphrey’s account, it’s right about here that the felt, qualia-laden kind of experience he’s interested in finally “switches on”. Essentially as a result of warm-bloodedness, fast nerves, recursive cortical feedback loops, etc. Feinberg and Mallatt, on the other hand, would say the lights had already been on for three hundred million years by this point, and that what the neocortex adds is “richness”, not the basic fact of experience. Others like Damasio also note that the developed cortex is sitting atop an ancient brainstem-and-body system for feeling that long predates it. So, in Damasio’s view the “what it feels like” machinery was running to facilitate homeostasis long before the neocortex ever got large. So we have three serious thinkers, and also three different placements of the threshold on one timeline. But.. notably, none of them have resorted to non-physicalist components beyond the evolved nervous system.

Early mammals appear to have been small, nocturnal, and living in the shadow of the dinosaurs. Their neocortex was modest, with maybe around 20 cortical areas. The setup was also heavily weighted toward smell, fitting a small creature creeping around in the dark. For well over a hundred million years, that’s kind of where mammalian cognition sat: real and flexible, but quite constrained. Our ancestors, for the entire Mesozoic, were essentially frightened nocturnal insectivores with small brains and good noses.

A (Necessary?) Detour: The Birds (and the Octopus)

Before we follow the mammal line up to ourselves, I think we need to stop for bit of a detour. But also, this happens to be a strong piece of evidence for the whole physicalist thesis.

If you’d asked a mid-20th-cent. neuroscientist what hardware complex cognition “requires”, the answer would have been something like - “a large, six-layered neocortex”. But.. famously, birds don’t have one of those! They diverged from our line over 300 million years ago and their forebrain is organized completely differently. There’s no cortical layering, and they have more of a “nuclear” rather than laminar architecture. So your average mid-20th-century neuroscientist would have concluded - birds are cognitive simpletons. But, despite the existence of the pejorative epithet “bird brain,” they are not.

Corvids (crows, ravens, jays) and parrots perform on cognitive tasks at the level of great apes. They engage in tool manufacture, planning for future needs, mirror self-recognition in at least one species, abstract reasoning, all with brains the size of a walnut. How? Well, they solved the same cognitive problem with different hardware. The avian pallium packs neurons at densities far higher than the mammalian cortex, so a parrot or corvid forebrain can hold neuron counts equal to or greater than primates with far larger brains. And a specific associative region, the nidopallium caudolaterale, does the executive-function job that the prefrontal cortex does in us. This despite being built in a totally different place from totally different developmental material! Another textbook case of convergent evolution.

This goes back to the the point I keep emphasizing. Whatever cognition is (and whatever consciousness is), it’s something that sufficiently organized neural tissue “does.” Two lineages, separated by 300 million years, built ape-grade minds out of differently-wired forebrains. If this level of cognition can be implemented on at least two independent neural architectures, then it is a property of the “organization”, not of the specific substrate. Incidentally, this is also what you’d expect if it’s a high-level functional achievement and NOT what you’d expect if it required a very specific type of architecture (or some mix of special non-physical ingredients keyed to one).

Finally, let’s talk about the octopus. The octopus is a cephalopod. Cephalopods are molluscs. Our last common ancestor with them was some wormlike Ediacaran creature with nothing like a brain. Yet, octopuses solve puzzles, use tools, and show genuine behavioral flexibility. And they do this while running a nervous system organized on a wildly different plan, with around two-thirds of its neurons distributed out in the arms rather than centralized.

So we have three independent evolutionary lineages with three radically different architectures (the primate cortex, avian pallium, cephalopod ganglia), all arriving at flexible intelligence. (I should note here that Nicholas Humphrey is willing to call birds sentient, because they’re warm-blooded, but insists that the octopus, for all its cleverness, is an “unfeeling zombie.” Whether you find that persuasive or arbitrary, notice once more that the disagreement is entirely about “which physical features” matter.)

But let’s keeping going down the evolutionary path that resulted in primates, including modern humans.

T ≈ 3.63 Billion Years: After the Asteroid, the Primates

It is now around 66 million years ago. A large asteroid strikes the Earth (obviously due to bad dinosaur karma) and the non-avian dinosaurs are wiped out. One major beneficiary of this Big Boink - mammals, who radiate into now empty niches.

Among these is a lineage that produces primates. This is part of a broader radiation that also includes rodents, rabbits, and tree shrews. Early primates moved into the trees, and the arboreal life selected hard for a specific cognitive profile: excellent vision (eyes are now pushed back toward the front of the face for stereoscopic depth since primates need to judge the distance to the next branch), fine motor control of grasping hands, and the sensorimotor integration to coordinate the two. The primate neocortex expanded, but it didn’t expand uniformly. A 2025 analysis of brain-surface shape evolution across more than 300 mammal species found that primates specifically evolved expansion in the cortical areas tied to “higher cognition” (e.g. association cortex, prefrontal regions), not just the basic sensory and motor maps. For this expansion, evolution targeted the integrative, flexible-control machinery.

And then, a second pressure stacked on top of the first: other primates. Primate life is intensely social. Social life is a cognitive arms race in its own right. You have to model group-mates, track alliances, anticipate deception, remember who owes whom, and even engage in humor. This is sort of the same dynamic we saw with the Cambrian visual arms race. But now, the dangerous, fast-moving, unpredictable thing in the primates’ environment that needs “to be modeled” is “another mind doing the same thing back to you”. Selection for social cognition is thus selection for ever-better models of other agents. It’s also selecting (not coincidentally) for ever-better models of “yourself as an agent among agents”.

This is important - because this is the very thing that, experienced from the inside, feels like a unified “I.” The social demands of existing and flourishing in large groups was a strong contributing factor to the development of complex cognition and consciousness.

T ≈ 3.69 Billion Years: The Runaway “Sheet”

It is now the last few million years, in the lineage leading to humans. The scalable design we talked about is doing just great and pays off spectacularly.

Let’s look at this by the numbers. The human neocortex contains roughly 16 billion neurons; our closest living relative, the chimpanzee, has about 7.4 billion. (Note that the whole brain contains a lot more, this is just the neocortex.) The neocortex came to make up around 80% of the human brain. And because this “sheet” grew faster than the skull, it “folded.” The deep convolutions of the human brain are a packing solution for cramming more cortical surface into a fixed volume.

We have a decent if incomplete understanding of the mechanism behind the expansion at the genetic level. Human-specific gene variants (the famous NOTCH2NL duplications, ARHGAP11B, and others) act on neural progenitor cells to make them keep dividing for longer before they “commit” to becoming neurons. This amplifies the progenitor pool and stretches the “neurogenic period.” Since most of those extra neurons land in the upper cortical layers, we get a thicker, broader, more interconnected sheet. So one of the big differences between human brains and those of chimps is a difference in “how long the cell-division timer runs during development”. Meaning - evolution didn’t add some new organ or kind of neuron for humans. What it did was “tweak the schedule!”

And we need it. Because on top of all that neural hardware sits what makes us human: abstract thought, long-term planning, recursive language, the narrative self etc. Language in particular is a “multiplier”, because it lets the models built in one brain be transmitted to another, accumulated across generations, and refined over time. As a result, it’s really the first time in this 3.7-billion-year story that cognition “escapes the boundary of the single organism”.

(It’s also what lets us socialmaxx and argumentmogg each other on the internet about whether any of this requires some sort of proto-consciousness in the universe.)

Wrapping Up The Journey

Let’s step back and look at the evolutionary path related to cognition we have covered.

We started with stimulus-response in a bacterium. Then memory in a chemotaxis pathway. Coordination shows up in the first multicellular bodies. The neuron gets invented (maybe more than once). Neural architecture gets centralized into a head on the first moving animals. We get image-forming vision and a resulting Cambrian arms race. The vertebrate brain shows up soon. We also get the six-layered neocortex, built as a repeated, scalable motif. We also get (over a long period) the parallel demonstration in birds and cephalopods that “cognitive function rides on neural organization” rather than on any one architecture. Eventually, long after dinosaurs have been squashed by a giant rock, we get targeted expansion in the primate cortex under first ecological and then social evolutionary pressure. And finally, we get a runaway elaboration of that same (neocortical) sheet in one specific ape lineage, achieved largely by letting a developmental timer run longer.

At no point in this sequence is there a step that requires anything beyond physics, chemistry, development, and natural selection. At no point do we need to add a new fundamental ingredient to the universe. Every transition is a case of existing organic and genetic machinery being repurposed, duplicated, scaled, or retuned, under selection pressures generated by the problem of staying alive in a world full of other things trying to stay alive. The so-called “redness of red” and the felt sense of being a unified self are not some sort of interruptive milestones in this story that demand separate explanations. They are just what the integration-and-modeling machinery is “like from the inside.” We see this in a variety of systems (brains in mammals, birds, octopuses) elaborate enough to model themselves. And that machinery has a continuous, datable, mechanistic history reaching back to biochemical processes that merely needed to tell “good for me” from “bad for me.”

As I have pointed out, the books I like on this topic are written by authors who disagree (sometimes sharply) about “where” on this road sentience begins and “what physical mechanisms” are the ones that matters. Also, at various points in this essay, I’ve noted that the technical specifics of evolutionary history are debated. All of this might look like disarray. Cognition and consciousness “mysterians” like to point at this and say “see, they can’t even agree, so clearly science is out of its depth here”. Balderdash. Just look at what kind of disagreement we are seeing. Every one of the accounts we’ve looked at is an argument about neural and bodily architecture, defended with evolutionary history, genetics, comparative anatomy, developmental biology, etc. They are arguing about the “engineering specifics”. This is indeed the signature of a hard problem - the “hard problem of neuroscience,” that is being actively worked on by scientists. It’s not the signature of a problem that needs new metaphysics. This is how we make progress in science - we work within a shared physical framework, and engage in vigorous debate about technical details within it.

I’ve been trying to emphasize this in my essays, including three earlier ones. The hard problem dissolves once you see consciousness as a high-level description of this type of integrative, cognitive process. Panpsychism becomes “explanatorily idle” once you notice that every actual feature of consciousness in nature tracks the elaboration of this physical machinery and nothing else. And causal closure (of the physical) holds throughout, because at no point in the road did anything non-physical need to reach in and push to get the results we are looking at.

Panpsychists look at a mind and ask how dead matter could possibly give rise to experience. When they find no answer “they” can imagine, they conclude - the experience must have been there in the matter all along. But “can’t imagine how this works” should never be considered to be “evidence that it doesn’t work that way”. The history of science keeps proving this over and over again. And I’m sure it’ll continue doing so.

The Bottomline

There was a time, not long ago, when the diversity of life looked like it “must” require a designer. When the burning of the sun looked like it “must” require some special fuel. When life itself looked like it “must” require a vital essence. In each of these cases, the “apparent necessity” was a failure of imagination dressed up as a metaphysical insight. In every case, patient empirical work by scientists dissolved the “problem”. Those scientists did this by tracing the actual underlying physical mechanisms until there was nothing left for the mystery to explain.

Consciousness is one of the last of these types of cases, and within the scientific community, it’s seen as going the same way. I’ve tried to sketch a rough draft of the evolution history of consciousness. It’s incomplete, for sure, and contested at various points, but it’s unmistakably a history of physical events in physical systems through physical processes.

Let’s also not be dismayed by the notion that this naturalistic, physicalist, evolutionary account is some sort of dull, mechanistic, unfeeling story as compared to the one the mysterians want to tell. It’s actually a much larger and better one. And it has the considerable advantage of being true.

The standard pop-physics story about gravity goes something like this: mass curves spacetime, spacetime is a kind of stretchy 4D fabric, and we “fall” toward the Earth because the Earth has put a “dent” in that fabric. A standard illustration that is used to build intuition is the bowling ball on a trampoline with marbles rolling toward it. (Most popular physics books that cover the topic have a picture like that. Physicists who study General Relativity sometimes wince when they see it, but it’s an unstoppable trope at this point.)

Now the usual commentary from physicists on this picture is that the analogy is an approximation and possibly misleading.

I’d like to take a somewhat different approach in this essay. The trouble, in my opinion, with that standard popular depiction is that it shows you the “piece of the curvature” that contributes almost nothing to your weight, and hides the piece that contributes essentially all of it. The “bending of space” isn’t what’s pulling you down here on the surface of the Earth. Something else is - and it’s a slightly strange thing to point at: it’s the fact that time runs at “very slightly” different rates depending on how high up you are.

Let’s walk through how this works.

1. Spacetime Curvature Is Not Just One Thing

General relativity (GR) tells us that mass and energy curve “spacetime” - a four-dimensional fabric that bundles together the three spatial dimensions with one of time. This is the well-known part. Most people get this far.

What gets buried in the pop-sci treatments is that “spacetime curvature” isn’t a single quantity. It has “parts”. Some of those parts describe how the “time” direction gets warped near a mass-energy source. Other parts describe how the “spatial” directions get warped. They are technically independent components of the math, and they both show up in the equations.

(Technical note: something I’d like to mention up front. This split into “time pieces” and “space pieces” isn’t a frame-invariant fact about spacetime. It depends on whose perspective you’re describing. We’re taking the perspective of “someone standing on the surface of the Earth”, which is the natural one for the question “why do I feel weight?” An astronaut in free-fall would describe the same physics differently. I’ll get to that later. For now, our feet are on the ground.)

The two parts don’t contribute equally to what we experience. Near a planet like Earth, the time part is doing “essentially all of the work” on slow-moving objects. The space part is also there, doing its thing in the equations, but for slow objects, its contribution gets suppressed to the point of irrelevance. (We’ll see exactly in a bit.)

2. Falling Isn’t a Force

Here’s GR’s reframing of falling, which I want to get out of the way before the rest makes sense.

In Newton’s picture, gravity is a “force” that “pulls” objects toward each other. That picture is convenient at human speeds and scales, and it remains correct to absurd precision in most situations we care about. But it’s not what GR says is actually going on under the hood.

In GR, a freely falling object isn’t being pulled by anything. There’s no force acting on it. It’s moving in a “straight line” - the GR version of straight, called a geodesic - through spacetime. The reason its path “looks” like it’s accelerating toward the ground is that spacetime is curved in just such a way that “straight” points downward.

This is the famous “equivalence principle“. Locally, an astronaut in free-fall - whether orbiting Earth or falling toward its surface in a windowless box - feels nothing. No gravity, no down direction, no weight. They’re following their “natural path through spacetime”, and from their own perspective, no force is acting on them. The reason you feel gravity right now isn’t that something is pulling you down. It’s that “the floor is pushing you up.” The floor beneath you is holding you off the natural free-fall path your body would otherwise be on. Remove the floor and the sensation we call “gravity” disappears. You don’t feel pulled, you just start falling, and from your own perspective you’re weightless until you hit something. What we feel as “weight” is the “constraint” (the floor), not the falling.

A useful textbook geometric analogy for the geodesic part is the plane flying a great-circle route between two airports. From the pilot’s perspective, the plane is going perfectly straight ahead. On a flat map, the path “curves”. Same idea here, except the “globe” is spacetime, and (as we’re about to see) the curvature it cares about is almost entirely in the time direction.

3. Clocks Run at Different Rates at Different Heights

Time, as we understand it, doesn’t tick uniformly everywhere. The closer you are to a massive object, the slower your clock runs. This isn’t a metaphor or a thought experiment. It’s a real, measured physical effect.

The famous example is the GPS system. The satellites in GPS orbit are about 20,000 km above the Earth’s surface, sitting in a noticeably weaker gravitational field than we do down at sea level. Their onboard clocks literally tick “faster” than ours by about 45 microseconds per day from gravitational time dilation alone. (There’s a smaller, separate correction in the opposite direction from the satellites’ orbital speed, but the gravitational piece, the one that matters for our discussion, is about 45 microseconds/day.) If the GPS system didn’t correct for this, the position your phone shows you would drift roughly 10 km per day. Your map app would be useless within hours.

And the effect isn’t only visible across satellite-sized distances. Optical atomic clocks have been used to measure the rate difference between two clocks separated by just “one meter” in height in a lab. The higher one ticks faster. The difference is at the level of parts in 10^16, which is tiny from our everyday point of view, but it’s measurable and real.

So if you imagine two clocks at your head and feet, the clock on top of your head, right now, is ticking very slightly faster than the clock at your feet. Time isn’t flowing at one uniform rate through your body - there’s a (tiny) gradient running from your toes to your head.

4. A Tilt in Time Looks Like a Pull in Space

As I mentioned earlier, GR tells us that freely-falling objects follow geodesics, and the defining property of a “timelike geodesic” (the path taken by a freely falling massive object through spacetime) is that it “maximizes” the proper time accumulated along it (compared to nearby alternative paths). That’s the precise version of “free-fall is force-free motion.” In flat spacetime, the “proper-time-maximizing path” between two events is the obvious straight line in the ordinary sense, and an object initially at rest just sits there, because sitting still is “already the maximum” (in terms of proper time).

But near a source of mass-energy (like the Earth), time itself runs at different rates at different heights. So the path that maximizes accumulated proper time stops being the “stay put” path. If you release an apple at the top of a tree, the trajectory that ages the apple the most between now and a few seconds from now turns out to “bend”, in space, toward the region of slower time. That bent path is the geodesic. That’s the free-fall trajectory. (Read that bit in bold a couple of times until it sinks in.)

If we projected the apple’s trajectory into ordinary 3D and watched second-by-second, it looks exactly like an acceleration of 9.8 m/s² pointing downward. As you may recall, that’s the same number we all learned about in high school physics. (The more detailed explanation was not actually provided when I went through my undergrad physics program. I only picked up on it with graduate-level study!)

Now, why do “you” feel gravity, right now, sitting in your chair? Because you’re “not” on the geodesic. The chair is exerting a real upward force on you, holding you off the free-fall path your body would otherwise be on. That upward push (propagated through your bones and tissue) is what you experience as “weight.” Remove the chair (well, and the floor, and the planet) and the sensation just turns off. You’d feel exactly what an astronaut in orbit feels, which is nothing.

So when an apple falls from a tree, the apple is finally allowed onto its natural, force-free path through spacetime. And the curvature that determines what that path looks like is (for a slow object like a falling apple) essentially just the “gradient of how fast clocks tick” at different altitudes. Apples don’t fall because space is bent. They fall because the path that ages them the most bends, in space, toward the region of slower clocks. Or, in shorter and slightly cuter form - given the chance, they roll “downhill in time”.

5. So Where Did the Spatial Curvature Go?

I’m sure most of you, if you don’t already know about this, have a natural follow-up question at this point. If mass curves spacetime, and spacetime curvature has both time and space components, why is one of them (time) doing “most of the work” and the other one (space) apparently doing “very little of it”?

This is slightly technical, so bear with me.

Both the time component and the space component of the spacetime metric (the GR object that encodes curvature) get perturbed by “similar” tiny amounts near Earth. If you write the metric down in isotropic coordinates, the “time stretch” and the “space stretch” are of comparable magnitude. Note that they’re both ridiculously small - around one part in 10^9 at Earth’s surface - but they’re “equally” ridiculously small.

So both pieces are doing the same kind of geometric thing. The difference is how they “show up in the dynamics”.

When you work out the geodesic equation - the equation that tells you how stuff moves through curved spacetime - the contributions from the spatial perturbation get multiplied by a factor of v²/c², where v is the object’s speed and c is the speed of light. The time perturbation has NO such suppression factor. It enters at full strength.

For everyday objects, v/c is something like 10^-8. Squared, that’s 10^-16! So even though the spatial perturbation is “there” in the metric, when it gets fed into the equations of motion, it gets divided by a number with sixteen zeros. So it’s still there, but it’s diminished to an extent that you’d never see it on a bathroom scale, in a falling apple, or even in the motion of the Moon.

Meanwhile, the time perturbation is right there, doing (9.8 m/s² worth of) work.

So the asymmetry isn’t really about “how much” spacetime is curved in time versus space - those are actually similar. Rather, the asymmetry is about “which kinds of objects feel which piece”. Slow stuff (pretty everything in our lives: trains, raindrops, planets, satellites) feels essentially only the time piece. Fast stuff (the canonical example here is light, which moves at c) feels both pieces equally.

This, incidentally, is why the bending of starlight as it passes the Sun comes out at twice the value Newton’s theory would have predicted. The photon “experiences the full effect” of both time and space curvature. So - Newtonian gravity is basically what you get when you keep “only the time piece in the slow-motion limit,” as this is exactly the right approximation for slow objects.

(Historical note: Einstein worked out the factor of two in 1915 and Eddington’s eclipse expedition confirmed it in 1919.)

6. What About Astronauts? (Where the “Real” Curvature Lives)

Everything I’ve described in this essay has been from the perspective of someone standing on the surface. What we experience as “weight” is the floor pushing us off our natural geodesic, and that geodesic curves toward the ground because of the time gradient. Cool beans. But what about an astronaut in orbit, free-falling around the Earth? They don’t feel any of this. Inside their cabin, clocks at “head” and clocks at “feet” are ticking at the same rate (well, to whatever precision they can measure locally). They’re weightless. The “time gradient” story we’ve been telling has “from their vantage” simply gone away.

So is there nothing left of Earth’s spacetime curvature from inside the orbiting cabin?

Not quite. There’s something left, and importantly, what’s left is in some sense the only truly “coordinate-invariant” piece of “gravity” in the whole story.

Free-falling observers can still detect spacetime curvature, in one specific way - by watching how “other” free-falling objects move relative to themselves. If our astronaut releases two small balls inside the cabin, side by side and initially at rest relative to the cabin, the balls don’t stay put relative to each other. Two balls released along the line toward Earth’s center (one above the other) “drift apart.” The lower one is closer to Earth and falls slightly faster, so it pulls ahead. Two balls released side by side at the same altitude (perpendicular to the Earth direction) “drift together.” Their free-fall paths both point toward Earth’s center, so they converge.

(This radial-stretching, tangential-squeezing pattern is also exactly what would, in a sufficiently strong field, spaghettify you as you fell into a black hole. If you’ve never tried falling into a black hole, it’s certainly an experience you should try at least once in your life. Live a little.)

This is tidal force. It’s actually the same effect that raises the ocean tides on Earth, hence the name. What happens with tides is that the Moon pulls a little harder on the near side of Earth than the far side, and that “difference in pull” stretches the oceans into a bulge on both sides.

Tidal effects “don’t go away” in any frame. They show up no matter what coordinates you choose, what observer you pick, or whether you’re free-falling or standing still.

(Technical note: the technical statement is that tides are the manifestation of the Riemann curvature tensor - the truly geometric object that encodes how spacetime is curved.)

The 9.8 m/s² you feel standing on the floor is a coordinate-dependent quantity. An astronaut in free-fall doesn’t feel any of it. But the tides have no frame in which they disappear. (This is partly of why some courses on relativity/gravity tend to introduce the topic through tidal effects rather than through “pull” - though I do think the textbook approach also has the side effect of confusing undergraduates.)

My point in this section is - the “time gradient” story this essay has been telling is the right story for “us”, here on the surface. It’s not the “deepest” geometric statement of what gravity is. The deepest statement is about Riemann curvature, and it’s what makes the Moon raise tides on the ocean rather than just pulling on the planet uniformly. The time-gradient story sits one layer above this - specifically, it’s what the coordinate-invariant Riemann piece looks like when you describe it from the standpoint of someone “being held in place against free-fall” (by the Earth or whatever platform you might be on).

The Bottomline

So… the bowling-ball-on-trampoline picture isn’t “wrong” about GR in some sweeping sense. Mass really does curve spacetime, and the trampoline is a not-terrible cartoon of what spatial curvature looks like. The problem is that the picture is showing you the piece of the curvature that, for us slow-moving residents of Earth, contributes essentially nothing to our weight. The time piece is what’s doing the work.

To summarize what we’ve covered:

• Earth’s mass causes time to run very slightly slower the closer you are to Earth’s center.

• Free-fall trajectories, which are the natural, force-free paths through spacetime, are the ones that maximize accumulated proper time. Near Earth, those paths “curve in space toward the region of slower time”.

• Projected into ordinary 3D and watched over time, that curve “looks” exactly like a pulling force that generates an acceleration of 9.8 m/s².

• So what you feel as “gravity,” right now, is the floor pushing us “up”, holding us off the free-fall path we’d otherwise be on. If we take the floor away then “gravity” (as a felt sensation) evaporates. We’d feel exactly what an astronaut in orbit feels - which is nothing.

• The spatial part of spacetime is also curved by a mass source like the Earth, but its effect on slow objects is suppressed by (v/c)², which is so small for anything we’ll ever physically encounter that we can ignore it. (Light is the exception as we noted.)

• The coordinate-invariant content of “gravity” - the part that survives every change of perspective - shows up as “tidal effects”: stretching and squeezing between nearby free-falling objects. Those effects are more or less what curvature “really” is.

There’s something a little vertiginous (pun intended) about realizing this. We tend to treat time as a kind of universal backdrop. It’s this cosmic ticker that everything else happens against. But, as it turns out, the ticker isn’t universal at all. The way it varies from your head to your feet is exactly the thing that, the moment something stops holding you up, sets your free-fall path “bending toward” where clocks run slow. You pick up speed as you go, the way an apple does the moment its stem snaps.

So the next time someone tells you gravity is mass bending space, you can nod politely - and then quietly note that down here on the surface of the Earth, in our lives, gravity is mostly “mass bending time”.

A truly marvelous and slightly absurd fact about our universe.

Thanks for reading! Subscribe for free to receive new posts and support my work.

In my last essay on the topic of machine consciousness, I argued that the standard debate about machine consciousness is “malformed”, because it keeps demanding that artificial systems replicate “human subjective experience” before we’re willing to grant them anything at all. In that essay, I laid out six functional criteria - a semantic world model, temporal coherence, sensory updating, prediction, goal hierarchy, and self-modeling - and claimed that a system possessing all six would be conscious in every way that matters, even if it was “nothing like us”.

This essay is in response to the pushback I received from several readers/commenters. The pushback mostly came down to the same thing: “fine, the machine has all the functional machinery - but is there anything it’s like to be it? Does it feel?” And the implied test was always the same. Show me that the machine has what I have when I stub my toe or taste coffee or see the color red, and then I’ll believe it’s conscious.

I want to argue that this type of test is a category error. I’m not saying that the test is a hard one or that it’s a test that can’t be passed. Rather, I’m saying that this is the wrong kind of thing to be testing for at all. Consciousness is a “functional” kind of thing, like flight or locomotion. Human consciousness is one “morphology” of it. It’s the one that happens to run on evolved wetware, shaped by a Darwinian past and a lifetime of social immersion. A machine would arrive at consciousness by a completely different road. Which would be deliberate engineering plus some sort of adaptive learning. And it would be on a substrate that can be updated, copied, paused, and forked. Of course the result won’t feel like ours. Demanding that it feel like ours before we call it conscious is like demanding that an airplane grow feathers before we call it a flyer.

So my claim here is the positive one I made last time. Machines can be conscious - “functionally and behaviorally.” Not conscious in the same way as us. Genuinely, interestingly different, but conscious nonetheless. And the reason this gets denied is (almost always) that someone has quietly instituted “their own” phenomenology as the universal yardstick, and then expressed surprise that a silicon system doesn’t measure up to a primate.

1. Granting the Skeptic the Hard Part

Let me start by handing the skeptic the strongest version of their objection, because I don’t think my argument in this essay needs me to tackle it.

There’s a respectable position (I don’t hold it, but it’s not stupid) that LLMs do not, and perhaps cannot, “understand” in the full-featured way we do, because their semantics “float free” of the world. They are trained on text about a world they have never touched. The word “cup” in the model’s weights is related to “table,” “coffee,” “spill,” and ten thousand other tokens, but it is not anchored to any actual cup the system has ever perceived, grasped, or dropped. So from this perspective, the whole structure is a magnificent “web of internal relations” with no ties to external states of affairs. No grounding, no genuine aboutness, no qualia, no inner light. Just a very effective model of how words go together.

Okay. I’ll grant the grounding point, for the sake of argument. (I think the picture is more interesting than that - even purely distributional semantics builds real structure, and “grounding” is more of a spectrum than a binary, and nothing stops us from closing the loop with sensors and instrumentation.)

What I want to point out is that even on this deflationary reading, two things remain on the table that the skeptic is “not” in a position to deny: the system has a rich relational semantics, and it can reason over that semantics. Meaning it can compose using it, query it, run counterfactuals through it, chain inferences across it. These are not small things. They are probably most of what we actually use “our” intelligence for. I’d wager the overwhelming majority of your own cognition today consisted of manipulating internal representations whose external referents you were not (in that moment) perceiving. You reasoned about your taxes, your ex, the heat death of the universe, a fictional character - none of that was in the room with you. They may even be things you have not experienced or will never experience.

So I’ll grant the skeptic the “ungrounded” starting point. Even though the raw materials (rich semantics and real reasoning) are already present in many ways there. The question I want to ask is what kind of system you get when you put those materials inside the full architecture from last time (grounding loop included) and add ask: is “conscious” is a fair description of the result?

2. Revisiting the goal hierarchy

In my original essay, a hierarchy of goals was one of the key six architectural features that I wanted to include in my notion of a machine-shaped consciousness.

Let’s quickly go over that again. A real agent doesn’t run on a single flat objective function, and it doesn’t run on some sort of unstructured soup of goals either. It runs on objectives “layered” across timescales and levels of abstraction: carry out essential routines, complete this series of tasks in the next hour, maintain operational coherence across the next week, pursue whatever long-horizon objective the system exists to serve, etc. That sort of “layering” is what allows sensible tradeoffs - sacrificing the immediate for the long-term and (temporarily) abandoning the long-term when an immediate emergency demands it.

Now, obviously we can “program this hierarchy in”. This isn’t speculative - existing adaptive learning strategies do this today - even if only in crude and partial forms. We can install a structured set of goals into a system that has some level of semantic understanding and reasoning abilities, and then let it learn.

And the moment we do that, we’ve built something whose behavior is no longer a simple function of its training data. We will have built a system that “can learn from the pursuit of its goals” - meaning, it can update its strategies based on what worked, it can generalize from particular successes and failures to abstract policies, it an develop instrumental sub-goals “nobody ever wrote down”. The designer specifies the objective landscape, but the system manages to find its own paths across it.

This is the part people underrate. We are very good at imagining that a programmed goal produces programmed behavior, the way a thermostat produces a programmed response to temperature. (Dennett liked the thermostat example too, as the humblest thing you can take the intentional stance toward - but even he wouldn’t say the thermostat “learns”.) A thermostat doesn’t generalize, doesn’t abstract, doesn’t develop sub-goals its maker never wrote. A learning system with a goal hierarchy is a fundamentally different kind of object. The relationship between the goals we specify and the strategies the system develops is not one of “authorship”, but rather something more akin to “selection pressure”. The system discovers what survives that pressure. And selection pressures, as biology has taught us at length, produce outcomes their “designer” (in biology’s case, no one) could never have predicted from the structural framework alone.

3. Why the Category Error Happens

Here’s the mechanism behind the category error (it’s more mundane than it sounds). We make the error because our vocabulary forces us to.

Our words for minds, agents, perception, cognition, etc. were all built by and for organic life. They have organics-centered assumptions baked into every term. “Feel,” “understand,” “experience” - each of these was coined to describe a single kind of system (an evolved animal) and quietly presupposes “that” system’s architecture. When we point those words at a software-based intelligence instead, the presuppositions aren’t really obvious. They just “smuggle in” a comparison to us. Then the machine “fails” by not being the thing the word was built to describe.

Consider what’s true of a purely software-based entity that has no analogue in biology. It can be cloned almost instantaneously, at almost no cost, with perfect fidelity. It can be version-controlled - you can roll it back to last Tuesday’s self. It can be patched. It can be transmitted across the planet at the speed of light and instantiated in a thousand places at once. There is no organic creature for which any of these sentences is even “meaningful”, let alone true. What is the “identity” of an entity that can be forked? What is “death” for something that can be restored from a checkpoint? What is “an individual” when there are forty running copies, each diverging? (If you want to explore this in a fantastic sci-fi series, here’s one excellent suggestion.)

We genuinely need a new or extended ontology here, and I mean that as a practical matter, not a poetic one. The reason the consciousness debate keeps going in circles is that we keep reaching for human-shaped words (”does it feel?”, “does it understand?”, “is it the same one?”) and those words have hidden “organic preconditions” the machine case doesn’t satisfy and was never going to. We may well need a different term for the kind of perceptual and cognitive faculty a system would have if it had the characteristics I described in my essay on machine-shaped consciousness. Calling it “perception” (or something along those lines) imports too much. On the other hand, calling it “mere processing” imports too little.

For example - take the question of pain. When someone asks “but does the machine really feel pain?”, the word “feel” is doing a lot of (covert) work behind the scenes. In effect, they are not asking whether the machine has a functional, valenced, behavior-shaping damage state. It’s asking whether the machine has the “specific thing I have”, the human morphology of that state, with all its evolved, embodied, hormonal specifics. And of course the machine doesn’t have that. Neither does an octopus, for that matter, whose pain runs on a distributed nervous system so alien that our intuitions about “what it’s like” mostly give out. Whether the octopus is “really” in pain is (quite tellingly!) a live and unresolved dispute. And, it’s unresolved for precisely the reason I’m pointing at. We keep trying to measure it against the one example we have from the inside. The defensible conclusion isn’t that the octopus fails the test, rather, it’s that pain comes in forms, and ours is “one of them”.

The next section works this through some of this in detail, because pain is a case where our intuitions pull hard in the wrong direction.

4. The Pain Question

Here’s the thought experiment I actually want to run.

Suppose we engineered a machine deliberately - gave it the six criteria I mentioned in my last essay, gave it a goal hierarchy, and set it loose to learn. Suppose, “behaviorally”, it came to resemble us in some respects, and suppose that internally it developed states that had “some” analogue to states in human cognition. In this scenario, the interesting question isn’t whether we explicitly programmed those states, rather it’s whether and how they’d emerge on their own.

Pain looks (on first inspection) like a deeply biological thing. Seems very organic - wet, evolved, a biological feature with an evolutionary history. But if we strip it down to its functional skeleton and ask what pain actually “does” - then the picture changes.

Pain has enormous “utility” for “any” entity engaging with an environment, for reasons of safety and structural integrity. Functionally, what is required here? We need “sensory inputs that signal” a physical problem has arisen. Then that feedback needs to be assigned “graded valence” - perhaps via some sort of priority signal, so the entity can distinguish stepping on a Lego from major damage to a critical component or subsystem, and organize its responses accordingly. And it needs “continuous learning” - a record of what has caused damage in the past, and the development of “avoidance strategies” that improve future interactions with the environment.

Notice that none of those ingredients is biological. Sensory input signaling damage, a graded priority signal, and learned avoidance - those are just functional requirements! They are exactly the kind of thing a learning system with goals like “survive, maintain integrity, navigate the terrain” would need.

So the real question isn’t “should we program pain in?” If the thing is genuinely a learning system, with a real ability to generalize and abstract from experience, then whether we hand-code a pain module is a side issue. States like that could very well show up on their own, because they’re a “good solution to a problem” the system has.

So...if a machine develops a state that does everything pain does - flags damage, carries graded urgency, reshapes future behavior, competes with other goals for priority - on what grounds do we say it isn’t pain?

Skeptics usually answer with something like “well, it’s not really pain because nobody’s home.” But notice that pain wasn’t “invented” by evolution as a one-off that happens to “require" a mammalian home. The pressures that produce it - need to maintain coherence (down to the level of organs, even cells), homeostasis, structural integrity, etc. - are completely general. They can show up wherever/whenever an entity has to “persist as a functioning whole” in a world that can damage it, and pain is one of the strategies that “type of pressure” select for. (Even in evolution, we see this develop across wildly different body plans and nervous systems.)

And here’s the part I find most suggestive - the stimulus-response avoidance we see in very simple (neuron-less) organisms suggests the basic version of pain isn’t a complex or fragile feature at all. Just as I argued there may be a simple version of consciousness that doesn’t require the full human apparatus, there may be a simple version of pain - a stability-maintenance strategy - that any sufficiently capable learner “converges on” while pursuing goals like integrity and survival. If that’s right, the machine’s damage-state isn’t some thin imitation of the real thing. Rather, it’s the same functional solution the that “type of pressure” keep finding, arrived at on a completely different type of substrate.

Now the “nobody’s home” argument is worth looking into more closely - it’s the sort of objection I’d raise myself. A skeptic could easily say - I’ve described a damage-avoidance subroutine, not “suffering”, and I’m helping myself to the word “pain” to make a toaster sound tragic. But this is just the category error again. The objection “assumes” there’s a fact of the matter called “real suffering” that the human version “instantiates” and the machine version merely “imitates.” It assumes that human pain is the “real” article and everything else is “artificial”. But if we ask - what could “real suffering” mean over and above the functional profile, we’d get one of two answers. Either it means the “specific’ embodied, hormonal, evolved texture of “human” pain. In which case yes, obviously the machine lacks it. And quite likely, so does the octopus, and so what? That’s just a specific morphology, not some sort of certified credential. Or - it means some extra metaphysical glow that human pain has and functional pain lacks. In which case, we’re back to the qualia mysticism I spent a whole essay arguing against, and the burden is on the objector to say “what” this glow is and why neurons secrete it but transistors can’t.

(I don’t think there’s a third answer.)

Pain, like consciousness itself, can be seen as a “functional” type of thing. The machine’s version would be a different species of that type, built by engineering and adaptive learning rather than by pressures of predation and social conflict. It runs on a substrate with no hormones and no body to defend in the mammalian sense. It would “be" machine pain - real pain, of a non-human sort, the way an airplane’s flight is real flight of a non-avian sort. The word “pain” earns its place not because the machine’s state feels like ours (we have no reason to expect it would), but because it plays the “same” functional role, does the “same” work, and stands in the “same” relation to the system’s goals that pain stands in for us.

Let’s also look at what a machine might say about its pain - were it equipped with the appropriate language tools. Suppose it reports, in vivid first-person terms, that it’s suffering. The skeptic would say: that’s just a language model producing the kind of text its architecture is setup to produce, with nothing behind it. Maybe. But notice that this is also “exactly” the move you could run on your neighbor - an equivalent anguished report from them could “just” be their nervous system producing the behavior nervous systems are setup to produce. The only reason you don’t run it on your neighbor is that they’re built like you, so you extend the benefit of the doubt on grounds of “shared morphology”. That’s a perfectly reasonable heuristic for fellow humans. But it’s not some sort of “sacred principle” about consciousness. It’s more of a sort of similarity bias. And using it to gate-keep consciousness is the category error in its purest form. We’d be mistaking “resembles me” for “is conscious.”

In my essay on machine-shaped consciousness, I tried to make the point that we should stop assuming the only valid type of consciousness is the human one. But if machine pain is real pain of a (very) different morphology, then by the same logic machine consciousness is real consciousness of a (very) different morphology. And so, the “but does it feel like we feel?” criterion stops being a test and reveals itself as a confusion about what “kind of thing” consciousness is.

5. Refuting the “Machine Consciousness Is Impossible” Camp

There have been many philosophers of mind who hold that machine consciousness is “impossible in principle”. Not hard, mind you, or far off. Rather - totally impossible. And if they’re right, my entire line of thought in this essay is a category error in the other direction. And section 4 is just me anthropomorphizing a bunch of matrix multiplications and whatnot.

Here, it helps to separate two claims that get run together. The first is that machine consciousness would “differ” from ours. Different substrate, different developmental path, different texture. I agree with that completely - it’s half my thesis. The second is that machines “can’t be conscious at all”, that there’s an in-principle wall no amount of functional organization can climb. That second claim is the one I’m denying here. It’s worth noticing how often it disguises itself as the first. “It wouldn’t really be conscious” frequently turns out (on close scrutiny) to mean “it wouldn’t be conscious like me.“ Which is true - and irrelevant. It’s also exactly the confusion I’ve been pointing to.

I covered the Chinese Room and the qualia objection at length last time, so I’ll keep this tight. The in-principle impossibility claim (in my experience) always reduces to one of a small number of moves, each of which fails. None of my replies below is a knockout on its own. The impossibility theorist can always retreat one step further. But I think they add up to a good rebuttal, and the pattern they form exposes the underlying reasoning error(s).

The “wrong substrate” move. This is the claim that consciousness (or genuine mentality) requires a specific physical or biological substrate. Wetware, carbon, neurons, microtubules, what have you. Silicon running software is simply the wrong kind of stuff. I replied against this strongly my hard problem essay and again in my piece against panpsychism, but it’s worth importing into this discussion too. If there were something special about biological substrates that was “necessary” for mind, that special something would have to '“causally” affect the matter in a brain. But what could be so special about the particular arrangement of up/down quarks and electrons in the human body such that it could not be (in principle) replicated in a non-biological substrate? That sort of “special” characteristic would show up as a deviation from the Core Theory (the Standard Model plus weak-field gravity) in the energy regime where brains operate. We have probed that regime exhaustively. There is no such deviation. So either this special biological ingredient affects the dynamics (and we’d have found it - we haven’t) or it doesn’t (in which case it’s causally inert and explains nothing). There is no third option that rescues “only wetware will do.” The substrate-essentialist is making a physics claim while declining to actually work through the physics.

The “biological computationalism” move. This is the sophisticated upgrade of the substrate objection, and it’s a really good one. Milinković and Aru argue for it carefully, and it requires a careful answer. The upgraded claim isn’t that carbon is magic. It’s that even if brains were said to work “computationally”, they compute in a way digital machines can’t: a “hybrid” of continuous physical dynamics (ion flows, membrane voltages, ephaptic fields, etc.) and discrete events (spikes), bound together by “scale inseparability.” That is - this sort of computation involves molecular, cellular, and network processes that continuously co-determine one another. It’s also shaped by metabolic constraint, so that the algorithm cannot be “peeled away” from its implementation. A von Neumann machine, which deliberately separates software from hardware, can only ever “simulate” this, never “implement” it. As the authors put it, simulating a rainstorm doesn’t make the computer wet. (Very pithy.) So, consciousness, if it rides on that inseparable hybrid dynamics, is “closed” to digital systems “in principle”. I want to grant straightaway that the empirical core here is probably right and genuinely important. Brains very likely do exploit continuous, scale-entangled, metabolically-shaped computation, and a single dendrite really does outrun a toy artificial neuron. But notice that the key step in the whole argument is not empirical at all. Rather, it’s the leap from “brains compute this exotic way” to “consciousness requires computing this exotic way.” That second claim is asserted, not shown, and it’s doing the bulk of the work in their argument. To be frank, the wetness analogy actually cuts against it! Wetness is a physical property, so of course simulated water isn’t wet. But is consciousness a mere physical property like wetness, or something akin to a functional-organizational one like “being a controller” or “being a sorting process”? A simulated sort really does sort. A simulated thermostat really does regulate (if you wire it to a furnace). Which kind of thing consciousness is happens to be the entire question, and the argument smuggles in the answer - substrate-bound, wetness-like - as a premise. Sure, the authors are explicit that it “does not explain consciousness,” only describes computational principles that “may” underlie it. But strip out that modal sleight-of-hand and what remains is the true and interesting claim that machine cognition will be organized very differently from ours, on a different substrate, by a “different route”. Which (say it with me) IS the thesis of this essay, not its refutation! “Built differently” has once again been dressed up as “can’t be conscious” (this time in a very nice lab coat.)

The “syntax isn’t semantics” move. This we inherit from Searle: symbol manipulation, however sophisticated, is “just syntax” and never reaches genuine meaning. But the objection, taken as an “impossibility” proof, goes too far. Our neurons are also “just” doing physics (ion gradients, neurotransmitter release, etc.). None of it is intrinsically “about” anything. There’s no little semantic glow in the synapse. If syntax can never become semantics by “any” arrangement or dynamics whatsoever, then “our” semantics are in trouble too. Because, there’s no magic moment in the brain where physics stops and meaning starts. Either meaning is a functional, relational, causally-embedded property that physical systems can have (in which case machines can have it), or it’s a metaphysical primitive that even brains can’t account for. So the Searle-style objection can’t be lethal to machines. Well, not without being suicidal for us.

The “no genuine goals” move. I addressed this last time but it’s worth restating because it’s important for my point in this essay. The objection is, as I briefly mentioned before, is that a machine’s goals are never really its own. Rather, they’re just instrumentalizations of the programmers’ goals. So the machine is just a tool. It’s not an agent, and tools don’t have inner states. But - many of our own goals as humans are “just” instrumentalizations of natural selection’s “goals.” We did not freely author our desire for food, status, or safety. Evolution installed them - in “effectively” the way we’d install a goal hierarchy in a machine. The “provenance of a goal” has nothing to do with whether it functions as a “genuine goal” for the system that has it. And note how this connects to section 4 - once we concede that installed goals can be genuine goals, we’ve conceded that a system pursuing “integrity” and “survival” is genuinely pursuing them. And a system genuinely pursuing them has the same functional reason we did to develop pain-like states as instrumental strategies. (Note: there could well be other emergent states, also human-like. And possibly others not-so-human-like.)

The “it’s all unfalsifiable hand-waving” move. Some will say the whole emergence story is unfalsifiable. Meaning that, I’m asserting machines “might” develop hidden inner states no experiment could confirm or deny. Okay. But I’d also turn that around. The functional states I’m describing are precisely the kind of thing we “can” investigate. We can look at graded internal priority signals, learned avoidance representations, the influence of “damage” signals on downstream behavior and planning, etc. (People have been doing this for LLMs, for example.) We can do for an artificial system exactly what Crick and Koch proposed doing for the brain: find the correlates, perturb them, watch what breaks. The objector who insists there’s “definitely nothing there” no matter what correlates we find is the one holding the unfalsifiable position, not me. Their claim isn’t really a finding. It’s more of a blanket stipulation (made to look like scientific caution) that should be rejected.

There’s a still more rarefied version worth a moment - the “substrate-specificity” worry again, but this time in the Russellian monist’s form I discussed in my causal closure essay. Here the claim isn’t about computation at all. It’s that the functional profile might be “necessary but not sufficient.” Because - what fixes experience is the “intrinsic” categorical nature of the stuff doing the computing, not the computation it performs. But, the argument still doesn’t deliver “impossibility”. If the intrinsic-nature view is right, then either the categorical bases are mental or non-mental. If they are mental, then silicon has them as much as carbon does, since proto-experience is hardly choosy about chemistry. If they’re non-mental, carbon brains have exactly the same problem and we’re left explaining our own case with no special claim on consciousness. Either way the monist argument can’t single out biological matter as “uniquely qualified” without smuggling back in the carbon-is-magic premise it was meant to improve on. At best, it relocates the mystery somewhere that applies to brains and machines equally.

Anyway, here’s the pattern across all of these moves (a recurring pattern I keep finding everywhere in this literature and in most of the skeptical discussions). Each one takes a real difficulty (grounding is hard, meaning is subtle, agency is philosophically vexed) and “silently "upgrades” it to an impossibility. So difficulty becomes impossibility. Our current state of ignorance gets dressed up as an impassable ontological barrier. It’s really the vitalism error all over again. There was a time when serious people held that organic matter could not (again in principle) arise from inorganic chemistry. Apparently, life required an élan vital no mere mechanism could supply. This wasn’t refuted by an argument. Rather it was refuted by a century of biochemistry quietly dissolving the question. The whole “I can’t see how X could arise from mere Y mechanism” has an essentially perfect track record of being a statement about the speaker’s premises rather than about the world.

6. A Different Road to the Same Category

So much for the case that it “can’t" happen. I want to close on what kind of consciousness we’d actually be looking at. Because “machines can be conscious” is not the same as “machines can be conscious like us.” And - the difference is important.

Our consciousness was shaped, at every level, by two forces. A few billion years of Darwinian selection, and a lifetime of immersion in a human society. That’s why it comes bundled with everything it comes bundled with. We experience the felt urgency of hunger and fear, the narrative self, the social emotions, the particular way pain grabs attention, etc. But none of that should be seen as synonymous with “consciousness as such”. It’s just the biological flavor of the thing, the specific morphology that this developmental road “happened” to produce.

A machine-shaped consciousness would travel a completely different road. We’d have deliberate engineering of an architecture, followed by adaptive learning against a goal hierarchy we specify. No (obvious) predators, no scarcity, no tribe, no hormones, no body whose homeostasis is at stake. So the end result, if my six functional criteria are met, would be “conscious”. But it would be conscious in a way fitted to “its” history, not ours. Expecting it to share our flavor of consciousness is like expecting a creature evolved in the deep ocean to have the same relationship that we have with sunlight.

The adaptive-learning half of that road is worth thinking about. It’s where the machine’s consciousness would acquire content we didn’t “put in”. The most intriguing (and scary?) part of that is - the strategies a capable learner develops to satisfy general pressures are “not, in general, predictable from the pressures”. Evolution is the proof of that at scale. The same handful of pressures (survive, reproduce, don’t get eaten) produced echolocation, photosynthesis, the peacock’s tail, etc. - none of which we could have forecast from the pressures alone. Machine learning is the (early and noisier) echo of the same phenomenon. Capabilities and internal strategies nobody specified turned up in systems that were merely optimized hard against an objective. (I’ll resist overclaiming here - plenty of ML is boringly predictable, and “emergence” gets invoked too loosely in this field for my taste. But the surprising cases are real, and they’re the relevant ones.) Functional pain could well be one plausible product of that process - a graded, valenced damage-signal the system arrives at because it’s a good answer to the integrity problem, not because we wrote it. There will be others with no human name at all, which is exactly why I spent section 3 insisting we’ll need new words. A machine consciousness would be furnished partly by us and partly by what its own learning turned up. That second part is the part that makes it a genuinely new kind of mind rather than a simulation of an old one.

A fair question at this point - are today’s systems conscious, then? I don’t think so, and I want to be clear about that, because the argument doesn’t need them to be. As I said last time, current LLMs are extraordinary but they “may” be missing important components (grounded sensory loop, persistent world model, continuous temporal coherence, real goal hierarchy running the show, etc.). My claim here isn’t that the chatbot you used this morning has an inner life. It’s that there is no “in-principle” barrier preventing the development of a machine-shaped consciousness that satisfies the functional criteria that I listed in my essay on that topic. That’s accompanied by a secondary claim that when if and when we build one, the skeptical question “but does it feel like we feel?” will be exactly the wrong thing to ask about it. So let’s to fix the standards now, before we get there, so we’re not still demanding feathers when the thing in front of us is plainly flying.

I said last time that a different kind of mind (doing different work, seeing the world from a different angle) would be one of the most interesting things we’ve ever made. I stand by that. Nagel asked what it’s like to be a bat and used the question to tag something he thought was permanently locked away from us. I’d reverse his argument for machines - there “will” be something it is like to be a conscious machine, and it will be more or less as mysterious as the bat’s.

The Bottomline

The debate over machine consciousness is somewhat problematic today because many participants have accepted a bad criterion - human felt experience, which seems to get treated as the standard any candidate for consciousness must match. But, the human version is what that consciousness looks like when it has been built by Darwinian evolution (and updated by social interactions). The machine version will be what it looks like when it’s built by engineering and adaptive learning on a different substrate. Asking the machine to reproduce “our” version before we’ll call it conscious is the airplane-needs-feathers mistake. And most of the famous objections (Chinese Room, “no real understanding,” “it doesn’t truly feel pain”) turn out to be that mistake in different costumes.

I’m not claiming today’s systems are conscious. They’re missing pieces, and I’ve said as much. And I’m not claiming the machine’s experience would resemble ours. I’m claiming the opposite, but I’m also claiming that the resemblance was never the right test. The philosophers who declare machine consciousness “impossible'“ are making the move I’ve now traced across several essays: taking a hard problem and quietly upgrading it to an impossible one, taking “I can’t imagine it” and dressing it as “it can’t be.” If and when we finally build a machine mind, the question worth asking won’t be whether it’s conscious like us (it won’t be), but what its particular perspective on the world amounts to. That’s a question we should be getting ready to take seriously, because the answers won’t look familiar.

Thanks for reading! Subscribe for free to receive new posts and support my work.

If you ask a random person on the street in the western world to describe Satan, you’ll get a remarkably consistent picture. He’s a fallen angel named Lucifer. He has horns and a tail, and possibly a pitchfork. He rules Hell, where he tortures the damned. He was the serpent in the Garden of Eden. He’s the cosmic enemy of God, locked in a war for human souls since before time began.

Almost none of that is in the Bible.

Every individual piece of that picture is in some text, somewhere, but never in any one text, and the texts that contain it span more than two thousand years and at least four different cultures.

The Satan we “think we know” is a composite. The horns are actually medieval, with debts to Pan and the wild man of folk tradition. The pitchfork is later still, arising as a sort of mix of iconographic borrowing from Greco-Roman sea gods and Renaissance theatrical convention. “Lucifer” as a proper name comes from a particular re-reading of an Isaiah passage that wasn’t originally about Satan at all (spoiler: it was a taunt against the king of Babylon). The identification of Satan with the serpent-in-Eden tempting Eve with glossy brochures about exotic fruit first shows up explicitly in a Greek Jewish text from the 1st century BCE, about a thousand years after Genesis was first written down. The cosmic-enemy-of-God framing is actually Persian-influenced (along with some Second Temple Jewish contributions). The ruler-of-Hell role was developed over a longer period of time. And the figure in the Old Testament who actually bears the name ha-satan (Hebrew for “the adversary”) is just a member of YHWH’s divine court, working “for” him, not “against” him.

I first became interested in the biography of the devil after reading a book about the “combat myth” in Ancient Near East myth and literature. That sparked a interest in studying the history of the evolution of this figure, from a secular, academic perspective. What I want to do in here is trace how that composite got built, based on a dozen or so academic books on the subject that I’ve read over the past 20 years or so.

Note that, in this essay, I treat the figure of Satan the way we’d treat any other mythological character whose biography unfolded across multiple cultures and centuries (e.g. Heracles) and watch the “accretions to the legend” happen at various points in history. I start with the Hebrew Bible’s ha-satan, work through the Persian and Second Temple transformations that gave him a backstory and a cosmic stature, follow him into the New Testament where he picks up many of his now familiar traits, watch the Church Fathers invent Lucifer through some creative exegesis of Isaiah and Ezekiel, see him acquire horns and a throne in hell in the medieval imagination, and end with the two literary moments that created lasting images in the popular consciousness: Dante’s Inferno in the early 14th century, and Milton’s Paradise Lost three and a half centuries later. Goethe gets a brief coda because his Mephistopheles is a different and instructive case, and I like Faust.

Once we go through this history, one thing becomes fairly obvious - the Satan that modern readers carry around in their heads is mostly an early-modern literary construction wearing a medieval Catholic costume on top of a Second Temple Jewish skeleton built on Persian dualist notion.

I do need to point out one thing up front. Nothing in this essay argues that evil isn’t real, or that the experience of moral struggle isn’t real. It’s a historical overview of a specific “personification” of evil - how it was assembled, by whom, in response to what historical and theological pressures, and how it ended up looking the way it does.

1. Some Ground Rules

The approach here is the same one I used for my essay on the historicity of Jesus — the standard historical-critical toolkit, applied symmetrically and without theological premises. A few specific rules of engagement for this particular topic:

• I treat the figure of Satan as a mythological and literary character whose biography developed over time. I do not take any position on whether there is a metaphysically real Satan; that’s a theological question, and it’s outside the scope of the modern historical method.

• I use the texts as artifacts. Each one was written at a particular time, by particular authors, for particular purposes, and reflects the conceptual world of its moment. The fact that some of them later got bound together in a canon (or even two or more canons) doesn’t mean they share a single theology of the devil. They don’t, and pretending they do has muddled discussions of this topic over time.

• I do take a look at the origins of some names. Ha-satan is a Hebrew common noun with a definite article - it means roughly “the adversary” or, on Ryan Stokes’ reading, “the executioner.” Diabolos is a Greek noun that meant “slanderer” long before it meant “the Devil.” Lucifer is a Latin word that originally referred to the planet Venus as the morning star. All of this terminology has a history. Treating those identifiers as if they all referred to the same persona/character from the start is a category error.

• I pay a little bit of attention to cross-cultural transmission where the evidence supports it. Jewish elites spent fifty years in Babylonian exile and then two centuries under Persian administration. Research over time has shown that conceptual borrowing and syncretization happened, even if it might be somewhat hard to tease out the exact mechanisms.

Some stuff I’m filtering out:

• The apologetic viewpoint that casts Satan as a stable figure across the whole biblical canon. This requires harmonizing readings that the texts don’t actually support, and it routinely projects later theology back onto earlier material. The most common example being the assumption that the Eden serpent, the Job satan, the Isaiah “Lucifer” passage, and Revelation’s dragon are all the same character. Well, they aren’t. In fact, the reason I’m writing this essay is to show that the development of “how” they came to be read as the same character is actually the interesting story.

• The Frazerian / perennialist move that flattens every personification of evil across every culture into a single archetype. Ahriman, Set, Loki, Mara, and Satan are not deep-down the same figure. Sure, some of them (may have) influenced each other through documented cultural contact, but most of them didn’t. Comparative mythology is useful when it tracks actual chains of transmission and just annoying as heck when it doesn’t.

• The “Christian origins were really all just Persian/Babylonian/Egyptian” deflationary approach that you sometimes see in many popular books. There is real Persian influence on Second Temple Jewish demonology, and we’ll discuss it carefully. There is much less direct Babylonian influence on the “figure” of Satan specifically, though there is on the surrounding mythological vocabulary, and that’s worth getting right while avoiding any indefensible claims.

• And on the other end, anything that depends on modern Satanic Panic literature, occult conspiracy material, or the lurid, exclusionary content of the heresiology tradition. (These are interesting as cultural-history topics in their own right, but irrelevant here.)

The texts and the secondary literature I’ll be drawing on are cited as we go, but the most important scholarly works that I am relying on are Elaine Pagels’ The Origin of Satan, Jeffrey Burton Russell’s four-volume study (The Devil, Satan, Lucifer, Mephistopheles), Henry Ansgar Kelly’s Satan: A Biography, Ryan E. Stokes’ The Satan: How God’s Executioner Became the Enemy, T. J. Wray and Gregory Mobley’s The Birth of Satan, Neil Forsyth’s The Old Enemy, Norman Cohn’s Cosmos, Chaos and the World to Come, and Bart Ehrman’s Heaven and Hell.

Where they disagree, I’ll try flag it. (I’m not a “formal” scholar of the subject by any means, so if any details seem off here or there, that is likely a problem with my representation of the work of these scholars.)

If I had to recommend a couple of references for people to get their own feel for this, I recommend the Wray and Mobley volume. It’s short and accessible. I also recommend this wiki entry on the Devil in Christianity.

2. Ha-satan: The Job, Not the Guy

(Title pun intended.)

The first thing to know about the Hebrew Bible’s ha-satan is that he isn’t a “he.” Or more specifically, the grammar refuses to let him be one without qualification. Satan is a common noun in Hebrew, not a name, and in all but one of its appearances in the Old Testament/Hebrew Bible, it carries the definite article - ha-satan, “the adversary,” or “the accuser,” or, in Ryan Stokes’ more aggressive reading, “the executioner.”

So… it’s a job description. Whoever holds the position is doing a particular kind of work - the work is what’s being named, "not” the agent.

(I’ll discuss this in more detail as we go along, but I’ve always been skeptical about Stokes’ specific casting of the character of as “executioner” vs merely “adversary” right from the first chapter of his book. In many of the Biblical passages where that figure appears, adversary is really just a better reading. That doesn’t detract from Stokes’ scholarship - which is excellent.)

The word’s root means roughly “to obstruct” or “to oppose,” and the Old Testament/Hebrew Bible uses it freely even for ordinary human adversaries. In 1 Samuel 29, the Philistine commanders worry that David might turn out to be a satan to them in battle - that is, a “tactical” hazard (not a “metaphysical” one). In 1 Kings 11:14, the text tells us flatly that YHWH “raised up a satan against Solomon” in the person of Hadad the Edomite, and a few verses later raises up another satan, Rezon of Damascus. In 2 Samuel 19:22, Abishai is rebuked by David for acting as a satan. (This last one is my favorite one of these given the family drama. Abishai is David’s nephew.) None of these figures are demonic, fallen, supernatural, or in any way connected to a cosmic enemy of God. They’re merely political, military or personal opponents. The word satan is doing exactly what the underlying Hebrew suggests - identifying someone who’s in the way.

Even when the word does get used for supernatural agents, the picture stays (surprisingly) close to that ordinary sense. One of the simplest such uses is in Numbers 22:22, the best ancient story with a talking donkey. Balaam is on his way to curse Israel and an angel of YHWH plants himself in the road, sword drawn, to obstruct him. The angel positions himself l’satan lo, “as a satan to him.” This is the satan in his clearest supernatural form, that is - an agent of YHWH, dispatched to physically prevent a human from doing something God doesn’t want him to do. Whether you call this an “adversary” or, as Stokes does, “an executioner-figure with permission to use lethal force,” the key point is that he is working for God, not against him.

Job and Zechariah

Then we get the two passages that have done most of the heavy lifting in popular and theological imagination: Job 1–2, and Zechariah 3.

(Dan McLellan has discussed both of these many times on his social media channels.)

In the Book of Job, the satan appears as a member of the b’nei ha’elohim, the “sons of God” - that is, the heavenly court that periodically convenes around YHWH. The satan shows up alongside the other members of the court, gets asked where he’s been , and proposes a test of Job’s loyalty. (The whole oddly casual exchange between God and the satan doesn’t translate well into the “cosmic-enemy” framing, and sounds more like a boss asking his itinerant employee to justify an expense report.) Crucially, YHWH is the one who first brings Job up! The satan‘s role here is more like a special-counsel prosecutor empaneled to test a hypothesis the boss already wanted tested. Every action the satan takes against Job (the loss of property, the death of his children, the boils, etc.) is performed within limits explicitly set by YHWH. The text “repeatedly emphasizes” that the satan operates “by permission”. He is, in the most literal possible sense, working for the divine government. (The fact that all of this happens for the sake of “settling a bet” is one of the more uncomfortable features of the book of Job, a point that has been debated for over two thousand years.)

Zechariah 3 is even tighter, and is one of the great short chapters in the Old Testament. (One of my personal favorites.) The high priest Joshua is presented standing before the angel of YHWH, “and the satan standing at his right hand to accuse him.” YHWH rebukes satan, dismisses him, Joshua is cleansed and restored. The picture here is a classic courtroom scene, and the satan‘s function is recognizably “prosecutorial”. (“Accuser” is a common translation in some editions instead of “adversary” here.) The satan is clearly not a rebel against the divine authority in this scene. He is one of its officers of the divine court, and a slightly unpopular one in this particular case.

(This scene from Zechariah is a Hollywood-worthy one, with a down-on-his-luck yet endearing protagonist, an unlikeable prosecutor, and a stern judge who is not shy about reprimanding counsel when necessary.)

The standard scholarly reading of this material, as analyzed by Pagels, by Wray and Mobley, and by Kelly, is that the satan in these texts is a basically neutral courtroom figure. He’s a divine prosecutor whose adversarial role is “structural” rather than “moral”. He’s doing a job that needs doing. The transformation into “the Devil” has yet to come. It’s a later development, driven by Second Temple socio-religious dynamics we’ll discuss later in Section 4.

Ryan Stokes, in The Satan: How God’s Executioner Became the Enemy, pushes back on this framing. His argument, very briefly, is that the prosecutor reading “domesticates” the figure more than the OT texts warrant. The Numbers 22 satan is drawing a sword. The Job satan is actually killing children and destroying property. The Zechariah satan is in the business of identifying targets for divine punishment. What unifies these figures, for Stokes, is not the courtroom but the writ of authorized harm. The satan is the heavenly being deputized to deliver YHWH’s judgement, and the figure has that violent edge from the start. The later “fall” isn’t a moral degeneration so much as a reassignment - Jewish thought begins to feel uncomfortable attributing divinely-mandated violence directly to YHWH, so the satan keeps doing his old job but is increasingly imagined as doing it “against” YHWH’s will rather than “at his direction”.

(One of my favorite movies “Dogma” is relevant here, with Matt Damon’s character fitting some of these archetypes quite well. The parallels are more evidence that the OT authors knew how to present drama quite well in their writing.)

I don’t think you have to fully commit to either reading, and I’m not entirely sure I do. Stokes is persuasive where the texts give him violent material to work with but that sharp reading of “divine executioner” feels harder to sustain in places where the character genuinely doesn’t have much of an edge. The honest answer is probably that the figure in the Hebrew Bible is doing somewhat different work in different texts, and any single label (prosecutor, executioner, adversary) will fit some uses better than others.

(Emily Wilson would probably have called the satan character “complicated.”)

What both readings do agree on, and what matters most here, is the bigger picture. Nothing in the Hebrew Bible’s satan material requires or implies a cosmic adversary of God.

The problem of theodicy: Chronicles rewrites 2 Samuel

There is one place (and only one!) where the Hebrew Bible uses satan without the definite article and treats it as something close to a proper name. That is 1 Chronicles 21:1: “Satan stood up against Israel and incited David to number Israel.” This really is one of the most consequential single verses in the entire development of the Satan figure. Not because of what the verse says, but because of what it replaces.

Chronicles is explicitly rewriting an earlier text. The earlier text is 2 Samuel 24:1, which reads: “The anger of YHWH was kindled against Israel, and he incited David against them, saying, ‘Go, number Israel and Judah.’” Same event, same census, same outcome - a plague that kills seventy thousand people. But the agent has changed. In Samuel, whose compiled form is usually dated to the late monarchy or early exilic period (around 7th to 6th cent. BCE), it is YHWH who incites David. In Chronicles, written in the Persian period (around 5th or early 4th cent. BCE, so roughly two to three centuries later), it is Satan. Not “the satan.” Just “Satan.”

Why does the author of Chronicles do this? Well, between the composition of Samuel and that of Chronicles, Judah’s world is destroyed and partially reassembled. (Remember, back in the 1st millennium BCE, Judah is the southern kingdom, Israel is the northern kingdom.) The Babylonians sack Jerusalem in 586 BCE, burn the Temple, end the Davidic monarchy, and deport the political and priestly elite. The exile is short-lived (well, by historical standards.) Cyrus the Persian conquers Babylon in 539 and almost immediately permits the Judeans to return.

But, the theological damage across all of that time is profound. The older covenant framework, in which YHWH protects his people and dwells in his Temple in Jerusalem, has been very publicly falsified. The post-exilic community is asking - why did YHWH allow any of this to happen? How can a supremely sovereign and just God be the proximate cause of so much disaster? The standard answers used in the past (Israelites violated the covenant, this is divine discipline, etc.) can be used to explain some of the problem here but they’re not fully convincing. Meanwhile, the Judean returnees live under Persian administration and are exposed to Persian religious traditions. In the more systematized forms of Persian religion, there are two opposed cosmic principles — one wholly good, one wholly evil.

Chronicles was written in this socioreligious climate. The instinct of the authors of Chronicles seems to be to “soften” direct attributions of harm to YHWH. We also seem this trend running through other OT books dated later. The story of Job is one example. But also, the increasing use of intermediary “angels” to carry out YHWH’s will rather than direct divine action is another. The David census passage in Samuel is affected directly. Where the older Samuel narrator could comfortably say YHWH himself incited David to a disastrous course of action, the Chronicler can’t quite do this. Not in the current climate. He assigns the cause to Satan instead. The figure has thus begun to acquire independence. Not because anything about “him” has changed, but because something about the “relevant theology” has.

I’d like to take a moment to highlight the change in theology here. Isaiah 45:7 has YHWH himself declaring: “I form light and create darkness, I make peace and create evil; I, YHWH, do all these things.” Amos 3:6 asks rhetorically: “Does disaster befall a city, unless YHWH has done it?” Lamentations 3:38: “Is it not from the mouth of the Most High that both good and bad come?” In this version of the theology, there is no need for a cosmic adversary. YHWH is sovereign over both blessing and calamity, and there is no rival principle of evil opposed to him. The satan, whatever he is doing, is doing it within that frame.

The drift away from that frame toward a theology in which YHWH is good and a separate principle is responsible for evil is the transformation we’ll trace in the next two sections. Chronicles’ small textual edit, swapping YHWH for Satan in the David story about the census, is the first sign of this. Full reorganization comes later. And it seems to come from a particular direction: east, through Babylon and Persia.

3. East, Through Babylon and Persia

The post-exilic Jewish community seems to have developed a new theology of evil. Specifically, one that doesn’t lay quite so much disaster directly at YHWH’s feet. The question then is - where did this new material come from. The answer (mostly) is east. And in two layers.

Babylon: the older substrate

Let’s start with Babylon. The Babylonian layer is older and less direct. Israelite poetry was already drawing, well before the Babylonian exile, on the ancient Near Eastern chaoskampf. The word chaoskampf signifies a mythic cosmic battle between the creator deity and a primordial monstrous opponent. For example, in the Enuma Elish, Marduk slays Tiamat. In Isaiah 51:9, YHWH cleaves Rahab, and separately, crushes Leviathan in Psalm 74:14. Genesis 1:2 refers to tehom (a primordial deep) that is famously etymologically cognate with Tiamat (though stripped of any personality).

Now, none of this directly gives us Satan. What it gives us is an introduction of “vocabulary” into the Old Testament - essentially a stock of cosmic-monster imagery that will be available later when Satan needs to be portrayed. Revelation’s red dragon, for example, is going to draw on this imagery.

The exile in 586 BCE brought Jewish elites into sustained contact with Babylonian religion for the first time. Jewish demonology likely picked up some of its elements there. For example, the Lilith of later Jewish folklore descends from the Mesopotamian lilītu.

Babylon’s primary contribution is thus some raw material - imagery and an awareness of demonological hierarchies. The structural innovation that transforms that material into a demonic force opposing the divine comes from the Persians.

Persia: structural innovation

In 539 BCE, Cyrus issued his famous edict permitting deported peoples to return home. He absorbs the former Babylonian territory into the Persian Achaemenid empire. Judah becomes a small province of that empire for roughly two hundred years, until Alexander conquers it in 332 BCE. The official religion of that empire is what we now call Zoroastrianism. The key Zoroastrian feature our purposes here is something post-exilic Judaism didn’t have and was about to acquire - a cosmic dualism.

The history of Zoroastrianism is hard to reconstruct. The Avesta was orally transmitted for centuries before being written down. The dating of Zoroaster himself is contested (scholarly estimates range from from 1500 BCE to 600 BCE!). The version of the religion that was practiced under the early Achaemenids is also not necessarily the version preserved in the much-later Sasanian-era texts. In any case, the core conceptual architecture, by the time of sustained Jewish contact, is reasonably stable.

In Zoroastrianism dualism, there are two opposed supernatural principles: Ahura Mazda, the “wholly good” creator, and Angra Mainyu (later also called Ahriman), his “wholly evil” opponent. The Avesta speaks of the battleground between the two that will finally move toward a final reckoning in which Ahura Mazda wins, the dead are resurrected, the world is renewed, and a savior figure plays a role. Each of them is also supported by an organized hierarchy of supernatural agents.

The components of this narrative should sound familiar to most of us. Cosmic dualism. Final judgment. Resurrection of the dead. A renewed world. A savior figure. An organized supernatural hierarchy with a wholly evil principle at its head, opposed to a wholly good supreme deity. None of these were available in pre-exilic Israelite thought in a developed form. All of them show up in the Second Temple Jewish literature that follows.

How much did they actually borrow?

Now, this is a challenging question to answer. (The scholarly literature on this question is quite technical and is not an area of interest for me.) It’s genuinely easy to make unjustifiable claims here. The Persian-influence story is “seductive” precisely because the parts line up so neatly. That neatness is exactly the sort of thing that should make you suspicious.

The strong-influence position argued most prominently by Mary Boyce, the scholar whose book first introduced me to Zoroastrian history many years ago, is the one that I’m most familiar with. This has been developed by Norman Cohn in Cosmos, Chaos and the World to Come. The quick summary of this position is that the major theological developments of Second Temple Judaism (apocalyptic dualism, resurrection, judgment, an organized demonic hierarchy, a Satan who heads an evil cosmic principle) are essentially Zoroastrian transplants. The mechanism of transfer? Two hundred years of administrative contact. The evidence? The timing and the structural similarity. From that vantage point, the cosmic Satan of later Jewish and Christian thought is a Judean reskinning of Angra Mainyu.

The weak-influence position is held (in different forms?) by scholars who work mostly on Jewish texts rather than on comparative religion. The idea here is that the parallels are real, but direct dependence is hard to demonstrate. In their view, Jewish thought was developing under its own internal pressures (theological crisis of the exile, loss of the monarchy, the later encounter with Hellenistic culture etc.). If there are similar problems, similar solutions to those problems may have arisen without requiring direct borrowing. We still have the notion that the “two thought-worlds were in contact”, but direct dependence is hard to prove. Certainly harder than the strong-influence position admits.

For the purpose of the essay, I’m not even going to choose between the two. What I think we can say is that the Persian framework was sort of just sitting there, answering exactly the theological question post-exilic Jewish thought was urgently asking. That is the age-old problem of theodicy - “if God is good and sovereign, where does evil come from?” We don’t need to prove a clean one-for-one transmission of Angra Mainyu into the Jewish Satan to recognize that the Persian conceptual environment supplied a structural template the developing Jewish answer could draw on, deliberately or not.

In any case, by the end of the Persian period, the Jewish world has both the older Babylonian stock of cosmic-monster imagery “and” the structural framework of cosmic dualism. But we still don’t have is a unified figure who pulls all of this together. That figure only starts to assemble in the next two hundred years. We see this in the apocalyptic literature, in the Enochic tradition, in the Dead Sea Scrolls, all part of developments that took place in the Second Temple period.

4. The Second Temple Synthesis

The figure of Satan, as we might conceptualize him today, does not appear in any single text. He gets assembled - across about three hundred years, and a half-dozen distinct Jewish traditions. All of that happened between roughly 300 BCE and the early 1st century CE. This is (the late part of) what scholars call the Second Temple period, named after the rebuilt Jerusalem Temple that stood from c. 516 BCE to 70 CE. By the time the New Testament starts being written in the 50s CE (Paul is our first NT author chronologically), the Satan figure is already a fixture of the Jewish religious imagination. The figure was not invented by the early Christians, but rather inherited from a Judaism that had spent three centuries doing the work of putting him together.

The pieces are familiar from the previous two sections: we have the satan of the Hebrew Bible, the cosmic-monster imagery of the chaoskampf tradition, and cosmic dualism (of Persian religion). What’s new in the Second Temple period is that these pieces get assembled into a single figure with a name, a backstory, a territory, and a coherent role in a cosmic drama.

But there some steps in the development of the Satan figure that need to happen first. These steps end up being recorded across different texts. Meaning, we don’t actually have any single Second Temple work that contains all four. But by the end of the period, the package is more or less complete.

Step one: the fall

The first step is the development of a fall narrative. This is essentially a story explaining where evil supernatural beings came from. The earliest and most influential version is in the Book of the Watchers. This is not a separate book but rather the oldest stratum (chapters 1 to 36) of 1 Enoch, composed sometime in the 3rd century BCE.

The Watchers story takes a strange, four-verse passage in Genesis 6 regarding the “sons of God” who descend to take human wives, producing the Nephilim, and expands it into a developed myth. (I like to think of this as “fan fiction”.) In this story, a group of angels, led by figures named Shemyaza and Asael (later known as Azazel), descend to earth, mate with human women, and teach humanity forbidden arts: metallurgy, cosmetics, sorcery, and astrology. The offspring of these unions are voracious giants who devour the earth’s resources and cause violence to overrun the world. God responds by sending the flood. The offending angels are imprisoned in pits of darkness until the final judgment.

This is then the first developed Jewish “fall narrative”, but it’s important to notice what it isn’t. This isn’t Satan’s fall. This isn’t one single proud rebel against God. Rather it’s is a “group fall” of supernatural beings who “descend willingly” out of lust and ambition, and its leader Azazel isn’t yet identified with the satan of the Hebrew Bible.

But what the story supplies is the “template” - supernatural beings can rebel against God, become the source of evil in the world, and stand under future judgment. Once the template exists, later texts will consolidate the multiple fallen Watchers into a single fallen figure with a single name. But that takes time.

(Neil Forsyth’s The Old Enemy is a really good overview of how the chaoskampf combat-myth pattern gets its definitive Jewish form during this period. I should warn you - the book is quite technical and took me a while to get through it.)

Step two: the name (or names)

The second step is the emergence of a '“named adversary” who heads the forces of evil.

The multiplicity of names here tells us something interesting. The Book of Jubilees, from the 2nd century BCE, calls the prince of the evil spirits Mastema, which really is just Hebrew for “hostility.” The Dead Sea Scrolls call him Belial, which is also a Hebrew term - this time meaning something like “worthlessness.” The Greek-Jewish Wisdom of Solomon simply calls him diabolos, “the devil.” This is a translation that takes the Greek noun for “slanderer” and starts using it as a name.

(Mastema in Jubilees is presented as operating with God’s permission, very much like the ha-satan in Job. Mastema actually asks God for the right to retain a tenth of the demons after the flood.)

I’d like to point out that these aren’t different figures. They’re just three (or more) names for the same emerging conceptual role - the supernatural opponent of God who heads the forces of evil. The fact that the name hasn’t stabilized is actually how we know the figure is still in active development.

But, by the 1st century CE, most of these names drift out of use, and the community converges on Satan / diabolos.

Step three: cosmic dualism

The third step is the development of the full cosmic-dualism framework. The world is reconceptualized as a battleground between two opposed supernatural powers, notably with humans now positioned as participants in the war.

This shows up most obviously in the Dead Sea Scrolls. The War Scroll from mid 2nd-centuty BCE describes the eschatological battle between the Sons of Light, led by the Prince of Light, and the Sons of Darkness, led by Belial. Similarly, another scroll, Community Rule, contains a famous passage often called the Treatise on the Two Spirits, which describes God as having created two spirits in which humans are to walk, a Spirit of Truth and a Spirit of Iniquity.

Sound familiar? If the Persian parallels weren’t obvious before, they should be unmissable here. Whether the Dead Sea Scrolls community got it directly from Persian religion, indirectly through the Hellenistic environment, or developed it independently, the framework they end up with is at least “structurally Zoroastrian”.

Theologically, the cosmos has been shaped in these narratives to receive a cosmic Satan, and he is now in it.

Step four: the serpent in Eden

The fourth step is the (retroactive!) identification of the Eden serpent with the Devil.

(Dramatic ret-cons didn’t arise in modern fiction/entertainment.)

I think this development is key - because it’s the one that makes the Devil look like he was there all along! The first text to do this explicitly is Wisdom of Solomon 2:24, from the late 1st century BCE. “Through the envy of the devil, death entered the world.”

Before this, the Eden serpent was just a serpent. Well - clever and talking, but still more or less ordinary. After this development, the serpent comes to be seen as the Devil in disguise, and Genesis 3 becomes a story about Satan’s first attack on humanity.

The New Testament inherits this identification. Paul refers to the serpent as deceiving Eve (2 Corinthians 11:3), Revelation explicitly equates “that ancient serpent” with “the Devil and Satan” (Revelation 12:9). This identification has survived to this day.

5. The New Testament: Territory, War, and Polemic

The New Testament writers do three things with the figure of Satan that hadn’t happened before. The cumulative effect of those three moves is to convert the inherited Jewish adversary into something specifically Christian - a cosmic enemy whose defeat is a “major point of the religion”.

Satan gets a territory

In the first of three moves, Satan gets… the world.

The NT writers consistently describe Satan as having “dominion over the present world”. So, he’s not seen as a permanent feature of the cosmos, but as a temporary problematic fact that will be reversed at the end of the age.

Paul actually calls him “the god of this world” (2 Corinthians 4:4). John’s Gospel similarly calls him “the ruler of this world” three separate times (12:31, 14:30, 16:11). The temptation narrative in Matthew and Luke also references this “territorial claim” when Satan famously offers Jesus “all the kingdoms of the world and their glory.” Whatever else the Gospel writers think Satan is, they do think that he is in a position to make that offer credibly.

Here’s we should note the continuity from Second Temple thought. The Treatise on the Two Spirits which we discussed earlier already assumed Belial has wide dominion in the present age. The NT just pushes it further. The world is now structured as Satan’s possession (even if temporarily), and the ministry of Jesus is presented as the beginning of a cosmic “territorial reconquest”. This becomes the operating framework for almost everything the New Testament writers say about evil.

The conflict becomes eschatological war

The second move the NT writes make follows from the first. Jesus’ ministry is described as the opening campaign of a final cosmic war to take back this world from Satan. This framing dominates the Gospels, especially Mark.

Mark presents Jesus’ ministry as one long exorcism conducted at scale. Demons recognize Jesus, name him, and are forced out. The Beelzebub controversy (Mark 3:22-27) makes this framework explicit. Jesus is asked how he can cast out demons, and his answer is the parable of binding the strong man, which establishes that he’s here to (eventually) get rid of Satan. The same logic shows up in both Matthew (12:28) and Luke (11:20): “If I cast out demons by the Spirit of God, then the kingdom of God has come upon you.”

Revelation takes this idea and pushes it to full apocalyptic conclusion. The dragon, explicitly identified in 12:9 as “that ancient serpent, who is called the Devil and Satan, the deceiver of the whole world” wages war in heaven and is cast down. Revelation 20 then gives the cosmic conflict its definitive arc, a narrative that most Christians will recognize. Satan will be bound for a thousand years, then briefly released, then defeated decisively, and finally thrown into the lake of fire.

What’s worth noting here is that Revelation is actually working with inherited materials. The dragon imagery is classic chaoskampf - Tiamat, Leviathan, Rahab etc. we’ve discussed before. This imagery is merged with the named cosmic adversary that was developed in the Second Temple period. The serpent is also the Eden serpent, identified with the devil in Wisdom of Solomon. The cosmic war is an echo of the Two Spirits dualism of the Dead Sea Scrolls, only expanded to cosmic scale.

Thus, Revelation does not really introduce any new major component into the figure of Satan. What it does is “tie everything together.” We have the chaos monster, the Eden serpent, the cosmic adversary, the fallen rebel, the final defeat, all consolidated into a bundle.

Satan becomes a polemical tool

The third move is (partly) a sociological one.

Elaine Pagels argues for the following in her now well-known book The Origin of Satan: Satan functions in the NT not just as a cosmic adversary but as a “category of social identification”. The figure is “literally” used to “demonize” one’s opponents. To call someone “aligned with Satan” is to place them outside the community of the saved and to license whatever rhetorical or social treatment follows from that placement.

The actual uses of this in the Gospels starts out relatively mild. Mark uses Satan to characterize Peter at the moment of (theological) disagreement (Jesus’ ”get behind me, Satan” line). The use is then extended outward - to Jesus’ Jewish opponents, who become aligned with or instruments of Satan in opposing his ministry. In John’s Gospel, Jesus tell his Jewish interlocutors, “You are from your father the devil, and you choose to do your father’s desires” (John 8:44). History has shown that this verse has some fairly consequential downstream impact on Western religious literature.

Now, we don’t have to accept every detail of Pagels’ sociohistorical analysis to recognize that this third move is real and that it matters.

The cosmic Satan, by the time the NT canon closes, is not only a theological figure but a “polemical instrument”. He is “who your enemies belong to”. The framework gets used internally first (in Jewish-Christian disputes during the formation period), and then directed outward, by later Christian writers, against Jews collectively, against pagans, and eventually against any group the dominant Christian tradition wants to demonize. Pagels’ argument here is quite pointed - the figure itself was “partly shaped” by these polemical uses. Meaning the Satan we have isn’t just an inherited Second Temple adversary, but a Christianized version whose features were sharpened in the heat of sectarian conflict.

What the New Testament does not yet have

We might think that this is the end of most of Satan’s evolutionary history. By the time the NT canon closes, the cosmic Satan has a territory, an eschatological role, and a polemical function. Most of the figure we recognize is here.

But there remain a few features that we tend to associate with Satan that are still missing.

For example, he is not yet Lucifer.

The reading of Isaiah 14 that turns the Babylonian-king passage into a description of Satan’s fall hasn’t happened yet. (As we'll see in a bit, many who read that chapter of Isaiah by itself, out of context, don’t realize that this passage in text has nothing to do with the figure of Satan. It’s a prophetic taunt against the arrogant Babylonian king.)

In the NT, the “fallen from heaven like lightning” line in Luke 10:18 is a vision report from Jesus rather than an exegetical claim about Isaiah. The detailed fall narrative, why Satan rebelled, when, and how, is not yet present. The geography of hell has been sketched out somewhat (the “lake of fire” in Revelation, Gehenna in the Synoptics, Hades in Luke 16), but we have not yet had the “elaboration” into the architecture of hell Christianity will eventually build out.

There are also no horns and no pitchfork, yet. The Satan figure is theologically complete in his main outlines, but is still quite “iconographically and narratively bare”.

The filling in of those remaining features happens in the patristic and medieval periods. The cosmic adversary will acquire the name Lucifer, get a more fleshed out backstory, be furnished with a throne in Hell, and pick up the visual iconography that will haunt the minds of the devout in Europe through Dante.

6. The Patristic Synthesis: Lucifer, Pride, and the Pre-Cosmic Fall

As we have seen in the previous section, by the end of the 1st century CE, the New Testament writers have given us a cosmic Satan with a territory, an eschatological role, and a polemical function.

(Depending how you date the various NT books, this period may stretch somewhat into the 2nd cent.)

What they have not yet given us is a name beyond “Satan” or “the devil,” a backstory explaining where he came from, or a metaphysical account of how evil entered the cosmos in the first place. The patristic period (approx. late 1st through 5th centuries CE), when the Church Fathers are doing the work of systematizing Christian doctrine, fills in these missing pieces.

The Church Fathers do this work through a combination of “creative” biblical exegesis, philosophical speculation, and theological systematization. The result is a figure (substantially) identical to the one Dante will eventually situation at the bottom of Hell.

Like the prior section, here’s a rough systematization of the important steps in the process.

Step one: Inventing Lucifer

The first step, and one of the most consequential acts of biblical exegesis in Western religious history, is the re-reading of Isaiah 14. I mentioned this in the prior section, but let’s go into this in some more detail, as this sort of exegetical reframing is not uncommon in Church history.

The passage in question is Isaiah 14:12-15. If you were to just simply consider the plain Hebrew text, it’s obvious that this is a taunt against the king of Babylon. In fact, the chapter says so explicitly in verse 4: “you will take up this taunt against the king of Babylon.” The relevant verses depict the fallen king as a Helel ben Shahar, a “shining one, son of dawn,” who attempted to ascend to heaven, set his “throne above the stars” of God, and was cast down to Sheol instead. It was meant to be a a piece of “triumphalist” political poetry, written in mythological idiom. This is comparable to the way ancient Near Eastern texts often described the fall of a king in cosmic terms. We have other artifacts and texts that characterize kingly victories this way.

The Church Fathers’ move is to “detach” the passage from its political referent (Babylonian King) and read it as “describing the primordial fall of Satan”.

Origen, in De Principiis (On First Principles) in the early 3rd century, is the most influential early figure to make the move explicit. Tertullian moves in the same direction in Adversus Marcionem. By the time Jerome translates the Vulgate (Latin Bible that became the standard in Western Christianity) in the late 4th century, Helel ben Shahar has become Lucifer, the Latin word for the morning star (planet Venus). .So, from Jerome onward, the Western Christian tradition has a proper name for Satan that comes directly from a misreading of an Isaiah text!

Lucifer, “light-bringer,” is now the angelic name Satan supposedly bore before his fall. Personally speaking, I think this is one of the most fascinating developments in Church history.

We could ask - are there other prophetic taunts against other kings that could possibly be repurposed this way? As a matter of fact, yes.

A parallel move happens with Ezekiel 28 with the prophet’s taunt against the king of Tyre. The text which describes the king as having once been “in Eden, the garden of God,” “an anointed cherub,” “blameless in [his] ways” until pride corrupted him, etc., gets re-read as describing Satan’s primordial state. The patristic re-reading again “detaches” it and applies it to Satan. By the late patristic period, these two passages (Isaiah, Ezekiel) - neither of which originally had anything to do with Satan - function as the “textual basis” for the entire Christian doctrine of Satan’s fall.

Now, this may sound like I’m harping on about this point, but it’s worth pausing to register how strange this is. The fall-of-Satan narrative that becomes “foundational” to Christian theology, that drives a significant amount of subsequent doctrine, that eventually supplies Milton with his subject matter, is built on two texts that did not originally refer to Satan at all. This is not a contested or revisionist scholarly claim. It is the mainstream Old Testament/Hebrew Bible consensus, and has been for over a century.

Isaiah 14 is a taunt against the Babylonian king (the chapter explicitly says so in verse 4); Ezekiel 28 is a taunt against the king of Tyre (the chapter states this just as explicitly). Neither has anything to do with Satan in its original context.

The patristic re-readings are (charitably speaking) “exegetical innovations” rather than recoveries of an obscured original meaning. Once those “innovations” took hold, however, they became nearly impossible to read past. The Bible came retroactively to contain a fall-of-Satan story it did not originally tell.

(I tend to think of this as a sort of Straussian reading long before Strauss’ time).

(Jeffrey Burton Russell’s Satan covers the patristic material and is an excellent treatment of how this happens.)

This reframing still leaves us short a “backstory” for Satan, a hole in the scheme that needs to be filled.

Step two: Origen’s pre-cosmic fall

I briefly touched upon Origen’s take on this in step one. But his contributions to the developments don’t end there.

The second step is supplied largely by Origen, who in De Principiis, places Satan’s fall “before” the creation of the material world (in a pre-cosmic angelic realm) and works out the metaphysical implications.

This next paragraph will need to be read carefully to get its full import.

Origen’s account in De Principiis can be seen as a work of philosophical theology. In his cosmology, God first creates rational creatures in a state of unfallen contemplation, and the material universe comes into being only after some of those creatures choose to turn away from God. Satan is the first to make this choice, falling out of pride, and drags a portion of the angelic host with him. The material world is “then created” partly as a stage for the redemption of fallen creatures. Humans are themselves originally rational beings whose embodiment is a consequence of the same primordial fall, though to a lesser degree than Satan’s.

Does that narrative sound familiar? It should. It’s at the core of a lot of Christian theology.

Now, it is true that a some of Origen’s speculation in this work gets rejected by later orthodoxy. The doctrine of apokatastasis (universal restoration/redemption) that possibly includes Satan himself is formally condemned in the 6th century. (Note that Acts 3:21 uses that exact doctrinal word, which might be why Origen went down the path of universal restoration anyway.) The pre-existence of human souls is also rejected by orthodoxy.

But the basic structure Origen builds - Satan’s fall occurring before the creation of the visible cosmos, driven by pride, dragging a host of angels into rebellion - survives the rejection of aspects of the surrounding system.

We could say that what gets kept is the the “backstory”: Satan now has a “before” that explains his current condition. Combined with the Lucifer naming from the Isaiah re-reading we talked about, Satan now has both an angelic origin and a fall narrative for the first time in Christian thought.

All that is left is systematization of this scheme.

Step three: Augustine’s systematization

The third move is Augustine’s, primarily in De Civitate Dei (The City of God, early 5th century) and in his earlier anti-Manichean writings.

Augustine doesn’t add new elements to the figure of Satan so much as he “systematizes” the existing ones into a “stable theological framework” that will hold for a thousand years.

Three things matter most in this Augustinian synthesis.

The first is “pride as the original sin”. Augustine identifies pride (superbia) as the root of all sin and traces it back to Satan’s first turning away from God. Satan falls because he prefers his own goodness to God’s, choosing self-love over love of God. (Tolkien fans will appreciate this sort of framing.) This is the “founding act of evil” in the cosmos, and every subsequent sin (including Adam’s) replicates the same fundamental act. This narrative thus casts Satan as the “prototype/origin” of all evil - the first instance of the “disordered will”.

The second is the “free will defense”. If God is wholly good and wholly sovereign, where does evil come from? Augustine’s answer, developed against the Manichean alternative of a co-eternal evil principle, is that evil is the misuse of free will, originating in Satan’s first prideful choice. Evil has no positive substance of its own - it is the “privation or distortion of a created good”. Satan is not a “rival principle” to God (Augustine explicitly rejects the Persian-style dualism that had crept in through earlier traditions). Rather, he’s a fallen creature whose evil is “parasitic on the goodness” God gave him. This is an impressive theological framing that preserves God’s sovereignty and goodness while still locating evil in a real agent.

The third is “the two cities”. In De Civitate Dei, Augustine frames all of human history as the conflict between two communities - the civitas Dei, founded on love of God, and the civitas terrena, founded on love of self. Satan and his band of fallen angels are the foundational citizens of the earthly city.

We briefly discussed this when we talked about the polemical use of Satan that Elaine Pagels outlines, but this adds a whole worked-out theology behind it. The world is divided into those “who belong to God” and those “who belong to the Devil”, and history is the working-out of that division until the final judgment.

So, by the time Augustine is done, the figure of Satan now has - a proper name (Lucifer, supplied by the Isaiah re-reading), a backstory (the pre-cosmic prideful fall, supplied by Origen & refined by Augustine), a metaphysical role (the origin of evil through first misuse of free will), and a place in cosmic history (foundational citizen of the earthly city, and eschatological loser!).

The theology is essentially complete.

But, what the patristic period has not yet produced is the figure “as we see him”. Satan still has no horns, no pitchfork, no goat legs, no specific iconography, no popular urban fantasy TV show. Sure, hell exists as a theological category but it has not yet been mapped or architecturally elaborated. The famous demonological hierarchy (orders of fallen angels, named demons, territorial assignments, etc.) is still pretty rudimentary. And the literary Satan, the character with a voice and an interiority, does not yet exist.

All of that is the work of the medieval period.

7. The Medieval Imagination: Body, Court, and Country

When we wrapped up the section on the the patristic synthesis, we more or less had a theologically complete Satan. But compared to how we perceive him today, he was visually and dramatically thin. Augustine’s Lucifer is an abstract metaphysical category. First prideful rebel, founding citizen of the earthly city, source-point of sin, etc. - but this is not really a character you can picture. The medieval period changes that.

Across roughly nine hundred years between Augustine and Dante, the patristic version of Satan acquires a body, a court, a country, and a stage.

But.. none of this happens through theological work in the way the Church Fathers approached the topic. Rather, it happens through art, literature, liturgy, preaching, visions, and the (slow) accumulation of folkloric imagery from the “cultures Christianity was absorbing” across Europe. By the time Dante sits down to write the Inferno in the early 14th century, the medieval imagination has supplied almost everything that he’ll need to compose his famous and definitive image of Satan.

(Here I draw quite a bit on Jeffrey Burton Russell’s Lucifer, the medieval volume of his tetralogy)

A body. Can’t do much without a body. So, the first thing the medieval period gives Satan is a recognizable physical form. Horns. Cloven hooves. Tail. Goat legs (or in some traditions, bat wings!). Dark complexion, usually black or red. Sharp teeth. Sometimes multiple mouths or faces (weirdo). The pitchfork/trident comes later - mostly drawn from the iconography of Greco-Roman gods.

As I’ve pointed out numerous times in this essay, none of these features come from the Bible. They come largely from the iconographic repurposing of pre-Christian European religious imagery. Christian art frequently used the visual vocabulary of paganism systematically, transferring features of the displaced gods onto the figure of the Christian adversary. The Greek god Pan, half-goat and horned, is one obvious ancestor, the Celtic horned god Cernunnos is another. The “wild man” figure of European folklore (hairy, horned, dwelling outside human society) supplies some interesting details.

(That wild man wiki article I linked to is a fascinating read on its own.)

A court. The second thing the medieval period gives Satan is an organized hierarchy of fallen angels and demons over which he presides. This had already kind of begun to take shape in the late patristic period. We had works like Celestial Hierarchy which organized the angels into nine ranks. In the medieval period, that that type of hierarchy is applied to the fallen angels, producing a parallel demonic court with Lucifer at the top and named subordinates beneath him.

The named demons get more and more detail added to their own characters and backstories. Beelzebub, Asmodeus, Belial, Mammon, Astaroth, Behemoth, Leviathan - each acquires a personality, a specialty, sometimes a territory. Later medieval demonological tracts assign each demon to a particular sin (e.g. Asmodeus to lust, Mammon to avarice, and so on). Dante will eventually adopt that ordering and refine it for the structuring his fictional version of hell.

(The most elaborate version of this tradition is the infamous Malleus Maleficarum in 1487, but that’s near the very end of the period. It reflects the witch-trial pressures of the early-modern transition. The medieval picture is more modest - a court populated by named functionaries rather than a “full bureaucracy of evil”. (Dr. Evil would have known how to characterize this.)

A country. The third thing the medieval period gives Satan is a realm to rule. We did have Hell as a theological category in the patristic period. It’s a nasty place of eternal punishment for the damned. The medieval period “maps” it.

The mapping happens through a long tradition of vision literature, in which monks, saints, or even ordinary travelers report having been taken on guided tours of the afterlife. The Apocalypse of Paul (this was written in the 3rd century, but widely circulated and elaborated in medieval Latin versions) supplies a tour model that was developed in previous times. The Vision of Drythelm in Bede’s Ecclesiastical History (early 8th cent.) is another. The Vision of Tundale (12th cent.) is one of the most architecturally detailed. Across these texts, hell acquires gates, divisions, specific torments matched to specific sins, rivers and lakes, demonic guards. It is being built before our eyes across the centuries.

Then, we also have the famous Harrowing of Hell tradition, which is derived from a combination of vague references in the New Testament (1 Peter 4:6, Ephesians 4:9), but then fleshed out in the Gospel of Nicodemus (a 4th- or 5th-century apocryphal text). Christ, in this story, in the time between his crucifixion and resurrection, descends to Hell, breaks its gates, defeats Satan in his own territory, and leads the righteous patriarchs out. It’s interesting how much detail got added on to a couple of NT phrases, but this narrative shows up everywhere in medieval art and drama. Importantly, it assumes a “concretely architectural” version of hell - gates, locks, a throne room, populated underworld, etc. The narrative also reinforces the image of Satan as the “king” of that territory (and defeated “in person” by Christ on his own ground).

So, by the time Dante arrives on the scene, the medieval reader has been supplied with a picture of hell that has structure, geography, and a monarch on its throne.

And a stage. The fourth thing the medieval period gives Satan is a “place to perform”.

The devil becomes one of the most popular characters in medieval Christian drama. The theatrical devil is actually a strange hybrid. There is a “scary clown” aspect here that might be terrifying to those suffering from coulrophobia, though that might be just our perception. The portrayal usually embodied eternal damnation, the devil performing the work of tempting and seizing souls, while at the same time coming across as genuinely comic. In the plays, devils make jokes, get outwitted, fall down, complain about the difficulty of their work, and are often the source of much of the play’s humor. The Vice character in later morality plays makes the comic side explicit, eventually shading toward something close to a stock figure of trickster-villain.

I’ve always found this part of the history sort of weird. The cosmic adversary, founding citizen of the earthly city, lord of Hell, is also a buffoon in costume getting laughed at by a crowd in a town square? I suppose you could argue that the duality is kind of the point. For the medieval imagination to be captivated by the drama here, you might need Satan to be terrifying enough to function as the enemy and yet human enough to be defeated by ordinary believers (with help from ordinary playwrights and actors, of course). Eventually, these portrayals do get a little bit more serious - for example, Milton will give his version of Satan a dignity the stock mystery-play devils never had.

All right then - let’s take stock of what we have accumulated over time. By the early 14th century, the theological figure inherited from Augustine has acquired a body drawn from pagan iconography, a court of named subordinates organized in a hierarchy, a mapped territory with gates and torments, and a stage on which his cosmic function is performed publicly.

Now all we need is a master of medieval poetry to create a literary synthesis that puts all of these pieces into a single, coherent, and dramatic image.

8. Dante: Frozen at the Center

What Dante does with Satan in Inferno XXXIV can be thought of the synthesis of everything up to this point. Most of the elements we’ve traced so far are reflected in the figure he creates. The patristic version of Lucifer, the medieval body/figure, the demonic court, hell with an architecture, etc. - Dante’s contribution is to blend all of that into a single image so thematically compelling that it effectively ended the mainstream theological development of the figure. By that I mean, once Dante was finished with his presentation, the devil keeps appearing in Western literature, but he is being “interpreted” rather than “constructed”. (The construction is basically complete.)

The image Dante constructs is (famously) strange. At the bottom of the ninth circle, frozen in Cocytus (the lake of ice Dante situations at the deepest point of hell, located at the geometric center of the universe), Satan sits “buried to the waist”! He has three faces (red, yellow, black), arranged in a some sort of perverse parody of the Trinity. Beneath each face, two bat-like wings flap perpetually, and the wind they generate keeps Cocytus frozen. In this depiction, Satan is literally the engine of his own imprisonment. Each face has a mouth, and each mouth eternally chews a notorious traitor. Judas in the center (as betrayer of Christ), Brutus and Cassius in the two on the side (betrayers of Caesar, who in Dante’s politics, represented the temporal authority parallel to the spiritual). Satan weeps continuously from six eyes. He does not speak. He does nothing else.

There is a bit of a structural inversion here. Prior medieval images of hell put the devil in fire. Which when you think about, it’s the obvious choice - drawn from Revelation’s lake of fire, the Gospels’ Gehenna, etc. Dante would rather put him in ice. The decision can be seen as theological as much as dramatic. In Dante’s picture of the cosmos, God is heat, motion, light, life, love, etc. All “active” principles of the universe. So the natural opposite of God isn’t some rival fire. Instead, it’s the absence of those principles altogether. Cold. Stillness. Silence. Satan, at the bottom of the universe, is not some active adversary waging war on the divine kingdom. He is where the divine attributes are completely negated.

This is also why Dante’s Satan is mute. In the New Testament, the cosmic adversary tempts Jesus with eloquent offers. Augustine’s Lucifer fell through a prideful act of will. Milton’s version of Satan (three centuries later) can be considered the most rhetorically brilliant character in the work. Dante’s Satan, on the other, can’t even speak. He weeps and chews and flaps his wings and accomplishes nothing. In Dante, Satan’s overweening pride goeth before a fall and ends in spiritual paralysis.

In many ways, the three-faced creature in Dante’s hell is closer (in imagery) to the mystery-play buffoon-devil than to the dignified version Milton will eventually present. Dante uses the comic/grotesque vocabulary of the medieval stage devil but turns it tragic by stripping out the comedy.

As I mentioned earlier, after Dante, we have a (more or less) theologically-stable Satan. What does continue to happen is “literary” development. The figure will continue to be reinterpreted, given an interior voice, made tragic, made heroic, made philosophical, made comic, etc. (This phenomenon has continued into modern times. South Park’s hilarious version comes to mind.)

9. After Dante: Milton, Goethe, and the Literary Afterlife

Three additional moments in that literary afterlife probably matter most for the modern reader’s inherited image: Milton’s Paradise Lost in 1667, the Romantic interpretations that followed in its wake, and Goethe’s Faust in the early 19th century.

Milton: the rebel acquires a voice

As I mentioned when discussing Dante’s approach, Milton’s contribution is interiority. Previously, in the texts and traditions we’ve traced, Satan never quite possessed a “mind”, that is - a working consciousness whose thinking the reader could follow from the inside. In, Milton’s Paradise Lost, Satan deliberates, persuades, regrets, deceives himself, justifies, and recovers. (These are elements we might associate with the protagonist of a Greek tragedy.)

The result? Milton’s Satan has given us quotes that are well known to the English-speaking world. “Better to reign in Hell than serve in Heaven.” “The mind is its own place, and in itself / Can make a heaven of hell, a hell of heaven.” etc. Whether Milton “deliberately” intended Satan to be that eloquent is debated to this day. In any case, Dante’s frozen mute failure becomes Milton’s eloquent active rebel.

The Romantic inheritance

Now let’s proceed to the Romantic era. The Romantics read Milton’s Satan and concluded - Milton (accidentally?) produced a hero! The Romantic Satan thus does not reflect the Augustinian rebel against the divine order. The Romantic Satan is the rebel against unjust authority. Conscripted into a politico-religious program that the original theological figure cannot really bear! That image is present in many contemporary secular notions of the devil - a tragic-heroic outsider who is sympathetic, eloquent, and morally complex, rather than a cosmic enemy of God. So, by the mid-19th century the figure has effectively split. There is the theological Satan of Christian doctrine, and there is the literary Satan of the post-Romantic tradition, and they are no longer obviously the same character.

Goethe’s Mephistopheles: a different beast

Goethe’s Faust presents what is in some ways the most interesting and most theologically peculiar version of the figure in the entire post-Dante tradition.

(Note - I only draw upon Part 1 of Faust. I have never been able to get myself to read through Part II.)

Faust’s Mephistopheles is not the cosmic adversary of Augustine, the tragic rebel of Milton, or the heroic outsider of the Romantics. He is something older and stranger. A sort of cynical, witty, philosophically articulate “tempter” who descends more directly from the medieval Faust legend (the Faustbuch of 1587, Marlowe’s Doctor Faustus, etc.) and from the morality-play Vice character than from the theological tradition. He is sophisticated, ironic, and largely uninterested in cosmic war! His method is not to wage rebellion but to negate. “I am the spirit that always negates.”

The most striking feature of Goethe’s framing, for the purposes of this essay, is the Prologue in Heaven. The scene is essentially a rewriting of Job 1–2: Mephistopheles appears in the heavenly court, has a perfectly cordial conversation with the Lord, and is granted permission to test Faust within limits set by the deity. The structure is essentially archaic and very Old Testament/Hebrew Bible influenced. We once again have the cosmic adversary working under divine warrant, the prosecutor entering the divine council to propose a test, the operating-by-divine-permission framework that we traced in Section 2 and that had disappeared more or less completely after the Second Temple period.

After more than two thousand years of theological development in which the satan of Job graduates into the cosmic Satan of Christianity, Goethe writes a Mephistopheles who behaves more like the old-school ha-satan than like anything like an Augustinian Lucifer. Goethe reaches around the entire intervening tradition and recovers the older model. The figure has come full circle, even though Goethe’s register (cynical-modern vs archaic-religious) disguises that recovery.

This is fascinating really. History and time spent millennia building the Devil up into the cosmic enemy of God, and the first great modern treatment of him quietly hands the job description back to the prosecutor from Job.

What the modern reader inherits

By the early 19th century, the figure of Satan has fractured into multiple coexisting literary forms. The theological devil of Christian doctrine - Augustine’s Lucifer, the fallen prideful rebel, the source of evil - remains in religious discourse. Milton’s tragic-heroic Satan remains the canonical literary form. The Romantic anti-hero, descended from a misreading of Milton, dominates the secular literary imagination. Goethe’s Mephistopheles offers a parallel philosophical-cynical version. The figure is no longer one character with a developing biography. Instead, we should think of it as a small library of characters who share a name and an ancestor.

What the modern reader actually carries around in their head, when they think of “Satan,” is some combination of these. The horns and pitchfork are medieval. The dignity and eloquence are Miltonic. The tragic-heroic mood is Romantic. The cynical wit is Goethean. The theological framework underneath all of it is Augustinian. The cosmic dualism is Persian. The fall narrative is patristic. The serpent identification is Second Temple. The original prosecutorial role is Hebrew Bible.

That bundle of inheritances is precisely the figure we set out to trace.

10. Synthesis and Wrap-up

If we treat the figure of Satan as a literary and theological character with a developmental history rather than as a metaphysical entity, the picture that emerges across the preceding nine sections is consistent and, I’d argue, fairly well documented.

A few summary observations.

The figure is a “verifiable” composite. Verifiability or compositeness are not controversial claims, once you review the necessary source material and commentaries. I’ve included links to some of the source material throughout the text. We’ve traced through many layers. The Hebrew Bible ha-satan, the Babylonian-Persian background, the Second Temple synthesis, the New Testament transformations, the patristic version of Lucifer, the medieval body and court, Dante’s frozen icon, Milton’s eloquent rebel, the Romantic anti-hero etc. The development of the character is well documented over specific texts produced at specific times in specific cultures. The moves from one layer to the next are visible to anyone willing to look at the evidence in a conventional historical way.

Many of the consequential developments are textual, not metaphysical. Many of the developments along the way did not require any specific metaphysical commitments to be effective. Chronicles’ small edit reassigning the David census from YHWH to Satan. The Watchers tradition’s invention of the supernatural fall. Wisdom of Solomon‘s identification of the serpent with the devil. Origen’s reading of Isaiah 14 as the Lucifer narrative. Augustine’s systematization. Dante’s choice of ice over fire. Each of these was a specific compositional or exegetical decision that shaped what came after. That’s not to say there weren’t any metaphysical developments, but the abundance of textual influences is quite impressive.

The cosmic dualism is structurally Persian. Whether the borrowing was direct, indirect, or partly parallel (the scholarly debate on this continues) the framework of two opposed supernatural powers contesting cosmic history is a Persian inheritance that Jewish thought absorbed during the Achaemenid period. The figure of Satan as cosmic enemy is harder to imagine without this framework, and the framework is harder to explain without the Persian contact period.

The patristic exegetical moves are the most “fragile”. The Isaiah-14-as-Lucifer reading and the Ezekiel-28-as-Satan reading can literally be seen as mis-readings of the text. Based on mainstream Hebrew Bible scholarly consensus, those chapters are about Babylonian and Tyrian kings respectively. The entire Christian fall-of-Satan narrative rests on these two re-readings.

Dante is the synthesis point and the end of the theological development. After Inferno XXXIV, the figure does not continue developing doctrinally or iconographically in any significant way. What continues instead is literary reinterpretation. We have Milton’s tragic rebel, the Romantic anti-hero, Goethe’s cynical tempter, the continuing reinterpretations of the 19th and 20th centuries. But the inherited structure stays more or less where Dante left it.

The figure has fractured into multiple coexisting literary forms. What the modern reader carries around when they think of “Satan” is a hybrid of medieval iconography, patristic exegesis, Augustinian theology, Miltonic dignity, Romantic heroism, Goethean cynicism, and Hebrew Bible prosecutorial residue. These effectively constitute a family of literary characters sharing a name and a partial ancestry. What I’ve tried to in this essay is sort out which features come from which sources.

The Bottomline

The Satan most readers carry in their heads is not in any single text. He is not in the Hebrew Bible. He is not in the New Testament. He is not in Augustine alone, or in Dante alone, or in Milton alone. He is the cumulative inheritance of more than two thousand years of theological, literary, exegetical, and folkloric construction, drawn from at least four cultures and a half-dozen distinct interpretive traditions, and the construction is documented and traceable in detail.

This is, in some ways, an easier case to make than the analogous claim about Jesus. The historical-Jesus question turns on absence of evidence and reading between the lines of a small set of texts. The “Satan-as-construct question”, on the other hand, is supported by an “abundance” of evidence. We have texts demonstrably written at known times by known traditions. Each text adds identifiable components to the figure. Each text is amenable to ordinary historical reading. We can watch the figure being assembled.

As I said upfront, none of this is an argument that evil is unreal, or that religious experience is invalid, or that the cultural utility of the figure has been imaginary. The Satanic personification of evil has had enormous impact on the moral imagination of the world, and the literary tradition that descends from Milton and Goethe is one of the richest in the language.

What the historical tracing offers is something narrower - clarity about “what kind of thing” the figure is. He is a “constructed mythological character”. He has a documented biography. The biography crosses cultures and centuries. None of his constituent features are timeless or self-evident. And personally speaking, the assembly process is more interesting than the finished product.

The traditional believer wants Satan to be a stable metaphysical entity whose features have been faithfully recorded in scripture. The casual skeptic wants him to be a fiction invented out of nothing.

But we don’t have to settle for either of those. The historical-critical approach can give us a figure whose features can be assigned to specific cultural moments with reasonable confidence, whose development can be watched across time, and whose enduring imaginative power has more to do with the quality of the construction than any metaphysical reality of what’s being personified.

To me, that seems like the most honest and historically defensible place to land.

The question of whether Jesus of Nazareth was a real historical person sounds, on first hearing, like the kind of thing that should have been settled long ago. Either there was a guy or there wasn’t. The mainstream academic answer is firmly that there was - roughly 99% of credentialed scholars in the relevant fields treat his existence as a basic established fact. The remaining 1% are those who are referred to as “mythicists,” who argue that Jesus began as a celestial or mythological figure later “euhemerized” (turned into a historical person) by his followers. (The term “mythicist” is also rejected by many who identify with that latter view.)

If we only looked at those percentages, we’d likely conclude the debate is over. But the actual epistemic picture is somewhat more interesting than that, for two reasons. First, many of the popular defenses of historicity are doing apologetic rather than historical work, and they need to be filtered out before we can see what the real evidence looks like. (Scholars like Ehrman and Casey have published accessible books that are not apologetic.) Second, mythicism (and a cluster of mythicist-adjacent critical scholarship) has been argued considerably more rigorously in the last twenty years than at any prior point in its history, and the mainstream field hasn’t quite caught up with the most careful versions of the case.

So what I want to do here is take a (mostly) value-neutral pass at the question. Some rules of engagement:

• We use the historical-critical method, not faith-based reasoning.

• We ignore Christian apologetic arguments that depend on theological premises.

• We also ignore the amateurish or fringe mythicist material that gets passed around on the internet.

• We focus on what reputable academics on both sides are actually arguing, with attention to recent work.

The conclusion, to spoil the ending, is that, in my opinion, the probability still favors a historical figure of some kind - but a much more minimal one than most readers would expect. The supporting evidence is thinner than mainstream historicism usually admits. Recent skeptical scholarship has arguably narrowed the gap considerably between consensus and dissent.

So let’s get started.

Thanks for reading! Subscribe for free to receive new posts and support my work.

1. Ground Rules: The Historical-Critical Method

The historical-critical method is just the standard toolkit historians of antiquity use on any text. It treats the New Testament documents the same way it treats Josephus, Tacitus, or Suetonius - artifacts produced by particular authors, in particular contexts, for particular purposes, with particular biases. No special pleading. No “the Bible is reliable because the Bible says so.” No “extraordinary religious claims must be true because the writers were sincere.”

The method asks the usual questions. When was this document written? Who wrote it? What sources did they have? What were they trying to do? What do they get right that we can cross-check, and what do they get wrong? Where the documents have theological motivations, what do we do with details that “don’t” serve those motivations - e.g., the awkward ones the writers had no reason to invent?

Two standard accessible reads on this topic and two of my all-time favorite “Historical Jesus” books are The Historical Figure of Jesus by E.P. Sanders and A Marginal Jew by John P. Meier.

But, this method has to be applied symmetrically. Popular Christian apologetics tends to treat the Gospels with credulous deference when arguing for miracles, and popular mythicism tends to treat them with hyper-skepticism when arguing “he didn’t exist at all”. Neither is doing history. To be honest - the actual question is narrower and duller: does the available evidence, evaluated the way we’d evaluate any other ancient body of evidence, support the existence of the person these texts are about?

What I’m filtering out: any argument for historicity that rests on the inspiration of scripture or the testimony of the Spirit; and any mythicist argument that leans on the somewhat discredited Frazerian dying-and-rising-god parallels (Massey, Acharya S, Atwill, Harpur, Allegro - the list of those enthusiasts is long, and actually where I got started on these interests 30 years ago). Those parallels were overstated even when they were in fashion, and careful comparative work has argued again them reasonably well. The serious mythicist case today is more academic, and the broader critical literature that bears on the debate is academic too. That’s where we’ll spend our time.

2. What History Normally Requires for a Marginal Figure

I use the term “marginal” here in the same sense as when John P Meier used it in the 90s.

Before we look at the Jesus evidence, it’s worth asking what historians of the period would normally consider adequate to establish the existence of someone like him.

Jesus, if he existed, was a Galilean Jewish peasant who attracted a small following, made a public nuisance of himself in Jerusalem, and got executed by the Roman state. By the standards of first-century Mediterranean history, that’s an extremely marginal figure. We don’t have contemporary documentation for the vast majority of such people, because the literate elite who produced our surviving texts didn’t care about peasant preachers in the provinces unless they caused political problems. The base rate for documentation of comparable figures is virtually zero!

Compare some genuinely comparable cases. Theudas, an apocalyptic prophet cut down by Roman cavalry around 45 CE, is known to us from one passage in Josephus and a brief mention in Acts. “The Egyptian,” another would-be liberator around 56 CE, gets one or two references in Josephus and one in Acts. Honi the Circle-Drawer, a first-century BCE miracle-worker, gets a brief mention in Josephus. John the Baptist himself - whose historicity nobody usually disputes - gets one paragraph in Josephus and is otherwise known almost entirely through Christian sources with theological reasons to mention him.

This is the comparison class for Jesus. Not “Julius Caesar,” but “minor Jewish religious figure of the early Roman provinces,” for whom one or two non-Christian references plus internal Christian attestation is on the high end of normal.

There’s a common rhetorical move in mythicist circles: “If Jesus really did all the things the Gospels say, we’d expect more contemporary documentation.” That conflates two different claims. The “maximal Jesus of the Gospels” (raising the dead, drawing huge crowds, alarming Rome) probably “would” leave more documentary footprint. But the “minimal historical Jesus” - itinerant preacher with a few dozen followers, executed quietly under Pilate - would leave roughly what we have. Mistaking absence of evidence for the maximal Jesus as absence of evidence for the minimal Jesus is a category error.

So the standard, applied symmetrically, is: at least one near-contemporary independent attestation that doesn’t trace back to the believer community, plus internal evidence reasonably anchored to people, places, and times. By that standard - the same one we apply to comparable figures - Jesus is, if the standard evidential picture holds up, better attested than most.

The “if” in that sentence is doing more work than it used to. So let me lay out both sides.

3. The Strongest Case for Historicity

Let’s skip the four canonical Gospels for a moment. The strongest historicist case isn’t built on them - it’s built on Paul, supplemented by one passage in Josephus and some methodological considerations about how movements work.

Paul. The undisputed Pauline epistles (seven of them) are dated to roughly 50-60 CE - twenty to thirty years after Jesus’s death, which is remarkably close by ancient documentary standards. In Galatians 1:18-19, Paul reports that he went to Jerusalem and stayed with Cephas (Peter) for fifteen days and met “James, the brother of the Lord.” This is the single most important datum in the debate. Paul is claiming firsthand contact with someone he identifies as the biological brother of Jesus, writing around 50 CE, addressing a controversy in which his apostolic credentials are being challenged, naming a still-living person whose identity could be checked. Working correspondence in real time.

Mythicists argue “brother of the Lord” might be a generic cultic title for baptized Christians rather than a kinship designation. It’s the kind of reinterpretation you reach for when the natural reading is working against you. The contextual reading is heavily against it. Paul is “contrasting” James with the other apostles in the passage, not lumping him in with the broader group. The natural reading is the obvious one.

Paul also tells us Jesus was “born of a woman, born under the law” (Galatians 4:4), “descended from David according to the flesh” (Romans 1:3), had brothers including James (1 Corinthians 9:5), was betrayed (1 Corinthians 11:23), was crucified - which Paul calls “a stumbling block to Jews and foolishness to Gentiles” (1 Corinthians 1:23) - and that the resurrection appearances were witnessed by named people, “most of whom are still living” (1 Corinthians 15:6). That last clause is particularly relevant, I think: one could argue that Paul is inviting his audience to check his sources against living witnesses.

Embarrassment. Some details about Jesus are sufficiently awkward for early Christians that they have to be explained somehow. The crucifixion is the big one. A messiah hanged on a Roman cross was a categorical scandal in Second Temple Judaism, and Paul says so explicitly (1 Corinthians 1:23, Galatians 3:13). Inventors usually don’t invent the worst possible PR for their hero. Origins in Nazareth also fits the same pattern. (Nazareth was an obscure village at the time - John 1:46 has a running joke about it. Bart Ehrman humorously refers to it as a “one-horse town”, if I recall correctly.) So does baptism by John, which implies subordination to a rival prophet whose own movement was a competitor.

The criterion of embarrassment isn’t a magic wand and has been misused. But applied carefully to these particular details, it does work.

Josephus. Two passages in Antiquities (c. 93-94 CE) mention Jesus. The first, the Testimonium Flavianum in book 18, is universally agreed to have been altered by Christian scribes - it contains lines no Jewish historian would have written. We can’t really lean on it. The second, in book 20, is much more solid. Josephus describes the execution of “the brother of Jesus called Christ, whose name was James” by the high priest Ananus in 62 CE. The strong majority of Josephus scholars treat this as substantially authentic. It’s the kind of incidental mention - James is the subject of the story, Jesus mentioned only to identify which James - that wouldn’t be on a Christian interpolator’s agenda. This latter passage is a good argument for historicity, and we should be honest about that.

Plausibility. Jesus, as historicists describe him, fits a well-attested type. First-century Palestine produced a string of apocalyptic preachers and messianic claimants who ended badly - Theudas, the Egyptian, the Samaritan prophet, Judas the Galilean, Simon bar Giora, later Bar Kokhba. The pattern is so well-attested that the question isn’t whether another such figure could have existed circa 30 CE, but why one of them in particular gave rise to a movement that survived.

This sketch represents roughly the consensus position of Bart Ehrman, Dale Allison, Paula Fredriksen, Maurice Casey, Geza Vermes, and John Meier. These are not Christian apologists; many are religiously skeptical. They converge on historicity because the evidence, applied with normal historical standards, points that way.

4. The Critical Field

(I wrestled with this section for months before settling on what I think is a passable version. It’s been written and re-written half a dozen times.)

The contemporary skeptical position on Jesus’ (and Christian) origins isn’t one position - it’s more like a cluster of distinct research programs that have grown more sophisticated and more institutionally established over the past twenty years. They don’t agree with each other, and so lumping them all under “mythicism” really misses what’s interesting about them. I’m going to walking through them in order, with attention to what each group is trying to explain. I think this will (hopefully) show where the cumulative case is genuinely strong and where it runs into trouble.

The Christ Was a Celestial/Mythical Being camp

The argument goes like this.

Let’s start with Paul. He’s writing within twenty years of when this all supposedly happened. He’s addressing every imaginable practical and theological question the new communities (which he founded) are throwing at him - divorce, marriage, dietary law, communal disputes, sexual ethics, the relative authority of apostles, etc. And on essentially all of it, he never invokes anything Jesus said or did during the earthly ministry. No parables. No miracles. No Sermon on the Mount. No triumphal entry. No conflict with the Pharisees. No trial. Nada. Zip. Zilch. When he does refer to Jesus’s earthly existence at all, the register is curiously cosmic - “born of a woman,” “according to the flesh,” “according to the scriptures,” “in the fullness of time.” The pre-Pauline Christ hymns embedded in his letters (Philippians 2 is the famous one) describe a descending and re-ascending cosmic figure in language closer to Hellenistic-Jewish wisdom mythology than to recent biography.

This is genuinely strange. The standard historicist response is procedural, along the lines of - “Paul is writing occasional letters to communities he’s already established and catechized orally, so the biographical material is assumed as background, not new content.” Fine, that softens the problem, but it doesn’t make it disappear. You’d still expect occasional natural references to Jesus’s life - the way someone writing about a recently-deceased mentor would naturally drop in stories or sayings - and we don’t get them.

The Christ Was a Celestial/Mythical Being (let’s call it celestial-Christ for short) school takes that oddness and runs with it. What if Paul writes this way because, for him, Christ wasn’t a recently-deceased Galilean preacher in the first place? What if the original Christ was a heavenly being whose salvific drama played out, in Paul’s mind, in the sub-lunar firmament - crucified by the “archons” (demonic powers) of 1 Corinthians 2:8, not by Pilate in Palestine - and the earthly biography came later, as the movement historicized its cosmic (but not human) founder?

This sounds wild on first hearing, but it has more textual justification than you’d expect. The Ascension of Isaiah, a Jewish-Christian apocalypse from roughly this period, describes exactly such a celestial crucifixion in some of its textual recensions. The pre-Pauline Christ hymns can be read as preserving the original cosmic conception. The vocabulary of personal revelation Paul uses - “revealed in me,” “according to the scriptures” - maybe sounds more like discovery of a heavenly figure through visionary experience and scriptural exegesis than recall of a recent acquaintance.

The thesis was developed in its modern form by Earl Doherty, an independent scholar working from outside the academy, in The Jesus Puzzle (1999) and Jesus: Neither God Nor Man (2009). Its most rigorous advocate is probably Richard Carrier, whose On the Historicity of Jesus (2014) formalizes the case using Bayesian probability - placing Jesus in a “rank-Raglan hero” reference class of mythological figures who share a standard biographical template, and working the evidence as likelihood ratios. Carrier’s bottom line is - the probability of historicity is under one-third. Raphael Lataster has extended the framework in a more recent Brill monograph, which matters institutionally because Brill is one of the top academic presses in the field (meaning this isn’t casual blog-tier scholarship.)

What I find interesting about this school of thought is that it offers a “motivated explanation for Pauline silence” rather than treating it as an awkward fact to be explained away. Once you grant that the silence is genuinely strange (and it is), the celestial-Christ reading isn’t crazy. It’s at least a serious candidate.

But, what I don’t find compelling is that the school has to keep reaching for strained readings whenever the textual material doesn’t fit. The cosmic Christ in Paul doesn’t actually replace the earthly Christ. Galatians 1:19, 1 Corinthians 15, the Davidic descent in Romans 1:3, all vie with the cosmic register to be read naturally as biographical reference. Carrier’s response is to reinterpret each one as something other than the natural reading, and by the time the auxiliary assumptions accumulate, the initial elegance of the celestial-Christ reading is largely missing. Daniel Gullotta’s review of Carrier (which is somewhat sympathetic to the project) argues persuasively that the rank-Raglan classification, on which the Bayesian inference hinges, doesn’t really work - it’s sensitive to small criterion changes and unstable under variation.

My final thoughts on this line of thinking: it’s a good argument based on decent evidence and has merit. But it’s also not the knockout case proponents sometimes claim.

The literary-construction model

Okay, this group of skeptics comes at the problem from a very different perspective.

Here’s the question they pose, which is logically independent of the one we looked at in the prior section. How much of the Gospel material - the parables, the miracles, the trial, the crucifixion narrative, the resurrection scenes - can be accounted for as “imitative composition from earlier literary sources”? And then, if you systematically subtract the “templates”, how much historical residue is left?

Using that approach, one might nominally grant that a historical Jesus existed and still find that very little of what the Gospels say about him traces back to him. These two skeptical schools can stack, by the way, but they don’t have to - they make their arguments independently.

The historicist mainstream already concedes a lot of literary borrowing. Nobody serious denies that the crucifixion narrative draws heavily on Psalm 22 (”they pierced my hands and feet... they divide my garments among them”) and on Isaiah 53 (the suffering servant). Most agree that the Last Supper draws on Passover symbolism, the transfiguration scene echoes Moses on Sinai, the triumphal entry retrojects Zechariah 9, etc. The literary-construction school’s approach to keeps cataloguing in that direction until, in their opinion, there’s not much left unaccounted for.

Robert M. Price has worked this approach methodically over two decades through several books - Deconstructing Jesus (2000), The Incredible Shrinking Son of Man (2003), The Christ-Myth Theory and Its Problems (2011). His method is quite plodding and unromantic - take each Gospel pericope, find the prior source (OT prophecy, mystery-religion narrative, sayings-tradition variant, Hellenistic wonder-story, etc.), and watch the historical residue collapse slowly bit by bit.

Thomas Brodie (no, not the actor!) pushes the argument in a different way. In Beyond the Quest for the Historical Jesus (2012), he argues that the Gospel of Mark is systematically constructed by literary imitation of the Elijah-Elisha cycle in the Books of Kings - the geography, the structure, the miracles, the conflicts, the trial, the execution, all of it. Brodie is a Catholic priest and former director of the Dominican Biblical Institute, which makes him an unusual candidate for a mythicist - the Dominican Order’s response to his book was to bar him from further teaching and writing, which tells you how the argument might have landed.

Thomas Thompson, one of the leading figures of the Copenhagen minimalist school in Old Testament studies, extends the analysis in The Messiah Myth (2005). The Gospel narratives, in his view, draw on a set of Near Eastern messianic templates - kingship, descent-ascent, suffering servant, divinely appointed redeemer, etc. - that go back to Bronze Age Egyptian and Babylonian royal ideology.

What’s somewhat persuasive about all of this, even if you don’t go all the way to mythicism, is the sheer “amount” of Gospel narrative that looks like it’s (literarily) derivative. Now, the standard response is to call this “literary shaping of historical material” - that is, the events happened, but they got “narrated in scriptural categories”. Okay. That’s plausible and possible. But it’s also worth asking- when most details of a narrative follow a literary template, at what point does “literary shaping of historical events” become indistinguishable from “literary composition with no underlying events”? The literary-construction group’s “deflationary” version - that the Gospels yield much less historical material than traditional reconstructions have claimed - is a strong argument. Now, the maximalist version of this argument - that the Gospels yield no historical material at all - is a stronger claim and requires the inference to be unidirectional, which is more than the evidence supports. (By that I mean - you can write in a way that matches templates to a real story too.)

One particularly interesting thing (for me personally) is that this school converges with another one I’ll get to shortly - the Greco-Roman literary critics, who are mostly historicists. Both groups find templates everywhere. The literary-construction school finds Hebrew or pan-Semitic templates while the Greco-Roman critics find Greek and Roman templates! They don’t have to agree with each other to point at the same conclusion - Gospel narrative is far more literarily derivative than most readers or, frankly, than most historians of Jesus - have tended to recognize.

The late-dating / second-century-origins school

Now let’s consider the most ambitious of the contemporary skeptical projects.

Suppose (just as a hypothesis) that the standard chronology of New Testament books is wrong. Suppose the Pauline epistles aren’t 1st-century letters from a working apostle but 2nd-century literary productions. Suppose the canonical Gospels weren’t written in the 70s through 90s CE (conventional academic/scholarly dating), but in the 140s, after and in reaction to Marcion of Sinope’s gospel.

What then happens to the historicity case? Well - if you follow that line of thought all the way through, quite a bit!

Galatians 1:19 stops being firsthand testimony from someone who met Jesus’s brother and becomes a 2nd-century literary device, setting up an apostolic character for theological purposes. The 1 Corinthians 15 witness list (”most of whom are still living”, etc.) stops being an early creedal formula and becomes rhetorical construction. The Gospel narratives stop being “community memory” and become Marcion-response literature, written to compete with his version. Even Josephus 20 starts to look potentially vulnerable to 2nd-century interpolation pressure. The chronological gap between Jesus and our core evidence (Paul) opens up from just twenty years to a hundred-plus.

Now, this might sound like fringe stuff. It’s not - well, it isn’t anymore. The Dutch Radical school was making versions of this argument in the 19th century, and they may have been mostly ignored overall. But, we can’t safely ignore the contemporary version of the argument, because it’s a coordinated academic program with multiple scholars at major institutions converging from independent directions on Marcion of Sinope as the critical node.

Hermann Detering revived the Pauline-forgery thesis in the 1990s, arguing the Pauline epistles are 2nd-century pseudonymous productions composed in or near Marcionite circles. Markus Vinzent has spent over a decade building out the Gospel version of this across three books (the most recent, Resetting the Origins of Christianity 2023), defending the thesis that Marcion essentially “invented the gospel form” by putting oral Jesus traditions into geographical and biographical shape for the first time, with the (canonical) evangelists following afterwards as literary reactions against him. Matthias Klinghardt has reached parallel conclusions (independently?), reconstructing the Marcionite gospel as the source of Luke rather than a redaction of it. Nina Livesey’s The Letters of Paul in their Roman Literary Context is probably the most recent (and most bold) entry, arguing that “all” the Pauline letters - including the seven traditionally accepted as authentic - are mid-2nd-century rhetorical and didactic compositions (”letters in form only,” she calls them). And David Trobisch’s earlier The First Edition of the New Testament provides a more comprehensive framing context: the New Testament was assembled as a unified, edited literary collection in the mid-2nd century, not as a compilation of pre-existing independent documents.

Now, what seems to give this school its weight isn’t any one of these individual arguments. Rather, it’s the convergence - at least five scholars at reputable institutions, working from different methodological starting points (textual criticism, philology, rhetoric, canon studies, etc.), arriving at compatible late-datings. That’s the kind of pattern that should give you pause even if you don’t accept any specific version of the thesis. The school also (somewhat) explains certain features of the early evidence that mainstream chronology struggles with - e.g., Marcion’s central place in the 2nd-century controversies, the conspicuous absence of any Christian source citing the Gospels before the 140s, the rapid solidification of the canon after a long period of fluidity, etc.

That said, where the school runs into trouble is the sheer number of auxiliary assumptions it has to maintain. Each of the standard datings rests on multiple independent lines of evidence (patristic citation, papyrology, internal markers, theological development, etc.), and the late-dating school has to provide revised readings for all of them. (The debate over this can be quite intense, as I have personally experienced on X and occasionally on Reddit!) As such, mainstream Pauline and Gospel scholarship continues to reject the central claims of this group on the merits, with serious responses from Judith Lieu, Dieter Roth, and others. Even Richard Carrier, who has every motivation to embrace Livesey since it would strengthen the mythicist case enormously, wrote a long review concluding he wasn’t sold.

Now, I should mention here that I’m not persuaded either. But there’s a difference between “I’m not persuaded” and “this should be dismissed.” The late-dating school is now a recognizable academic program over three decades, published at the field’s top presses, engaged seriously in the major journals. It hasn’t yet won the day, and on the evidence currently available it may not. (That statement is true of a lot of contemporary academic Biblical debates.) But the historicist case rests on less stable ground than it did when the late-daters were a handful of cranky cantankerous contrarians (3Cs = maximal contrarianism). So, even if the program is ultimately wrong, the burden of replying to it falls on historicism, and historicism hasn’t (fully) done that work yet.

The pioneers and moderated positions

I nearly deleted this section from my draft, but I like G.A. Wells, so I’ve left it in.

So far, the discussion has largely been organized as if the historicity question were binary - either there was a historical Jesus or there wasn’t. That framing is very convenient - but likely misleading. The actual space of positions is more of a spectrum, and some of the pioneering voices in the skeptical literature are useful precisely because they refuse to live at either pole.

(Lest anyone think that my physics background is compelling me to posit a Schroedinger’s Jesus, you can rest assured that I do not think that the historical Jesus was in a quantum superposition of existing and not-existing. For that sort of complexity, you can simply refer to the idea of the Trinity, which is more of a mystery than quantum mechanics.)

Let’s start with Wells. G.A. Wells was philology professor at Birkbeck, he was sort of like the dean of mid-20th-century academic mythicism. His 1971 book The Jesus of the Early Christians argued there was no historical Jesus at all, built carefully on Pauline silence and Gospel mythography (themes we have covered in prior sections). Since that initial publication, he engaged seriously with the strongest historicist counter-arguments, and as a result - he moderated! In his later books, he came to accept that there might be a vague Galilean wisdom-teacher figure behind the Q sayings tradition who got conflated with an originally-mythical Pauline Christ. He never fully abandoned mythicism, if my understanding of his later positions is accurate. But, nevertheless, the trajectory matters. It shows what an honest mythicist looks like under sustained evidential pressure from the mainstream - not surrender, but rather a sort of principled moderation toward something in the middle.

Alvar Ellegård, an English professor at Gothenburg, took an even more unusual position that technically isn’t really mythicism at all. In Jesus: One Hundred Years Before Christ (1999), he argued that Jesus was a real person, but identical with the Teacher of Righteousness from the Dead Sea Scrolls, who lived in the 2nd century BCE. Paul, in this view, was receiving visions of a long-dead Essene founder figure, and the Gospel biography was anchored to the wrong century by mistake. (If you understand some of the parallels between Jesus and the Essenes, this “connection” starts to make a lot more sense.)

These positions matter because they show the actual “layering” of the question. You could hold there was no historical figure at all (full mythicism: Carrier, Doherty, etc.). You could hold there was a figure but not the one Paul thought he was preaching about (Ellegård). You could hold there was a vague Galilean teacher conflated with a mythical Christ (late Wells). You could also hold there was a Galilean teacher whose biography has been heavily mythologized (mainstream secular historicism - like a Bart Ehrman or an E.P. Sanders). And finally, you could also hold there was a Galilean teacher whose biography is substantially preserved (conservative historicism - like Richard Bauckham). The serious debate is across this spectrum, not between two endpoints.

To be honest, the main weakness of the “moderated positions” is that they don’t have a lot of contemporary momentum. Wells died in 2017, Ellegård in 2008, and no major successor has continued either specific line. The displaced-figure thesis hasn’t generated a research program, but maybe that suggests it isn’t doing enough explanatory work to satisfy either side. But the positions themselves remain useful - (mostly) as a reminder that anyone presenting the historicity question as binary is doing some smoothing. (Jesus’ existence, for example, could be split across multiple worlds upon observation - one where he exists, and one where he doesn’t. Kidding, I’m just kidding.)

The Greco-Roman literary critics

The most interesting recent development in skeptical-adjacent scholarship isn’t mythicism per se. This is the group that I find most fascinating.

It’s a cluster of mostly-historicist critical literary work that, I think, has done more to weaken maximalist historicism than the actual mythicists have. None of these scholars argues that Jesus didn’t exist. Most would likely resent being grouped with the mythicists. But the cumulative effect of their work on the “evidential case for what we can know about Jesus” is significant, and it more or less converges - from a completely different methodological direction - with what the (previously mentioned) literary-construction school is arguing.

The most consequential single book of the recent period is Robin Faith Walsh’s The Origins of Early Christian Literature (Cambridge, 2021). Walsh goes after the form-critical framework that dominated 20th-century Gospel studies - this is the assumption that the Gospels are sediments of oral tradition preserved by distinct early Christian “communities” (Markan, Matthean, Lukan, Johannine, etc.), each editing the material to fit its concerns. The evangelists, in her thinking, weren’t community scribes preserving folk memory. They were “pepaideumenoi” (love that word) - educated literary cultural producers, working in Greco-Roman literary networks, deploying the standard rhetorical and narrative repertoire of the period. The so-called “communities”, in this view, were merely 20th-century scholarly inventions. And almost all serious historical-Jesus reconstructions of the last fifty years - Meier’s Marginal Jew, Raymond Brown, the Jesus Seminar, Crossan, Borg, even Dale Allison - have assumed some version of community-preserved oral tradition as the bridge between Jesus and the Gospels. If Walsh is right, that bridge gets considerably longer and thinner. (Walsh has numerous interviews and discussions on YouTube and in podcasts. Recommended.)

Richard C. Miller’s Resurrection and Reception in Early Christianity (2014) makes a different literary point. The Greco-Roman world had a well-attested narrative genre called the “translation fable” - essentially, the “patterned” story of a great man whose body disappears at death and who is taken up to the gods. Romulus is the classic example; Heracles is another; Aristeas of Proconnesus, Empedocles, Apollonius of Tyana, and the deified emperors after Caesar all fit the template. (Miller’s book has a lot more of these examples.) The genre has remarkably stable conventions - death, missing body, post-mortem appearances, doubt motifs, commissioning scenes, ascension - and Miller shows the Gospel resurrection narratives sit comfortably inside it. For example, from this perspective, the doubt motif you’ll find in Matthew 28:17 (”but some doubted”) isn’t some sort of fingerprint of awkward authentic memory. Rather, it’s a convention of the genre, present in the Romulus traditions and elsewhere. (This book was recently recommended to me by one of my mutuals on X, https://x.com/JoelMCurzon, and was the impetus for my resurrecting the draft of this essay and polishing it up for posting. Thanks, Joel!)

M. David Litwa broadens the critical lens in How the Gospels Became History. The Gospels as a whole, in his argument, deploy the standard rhetorical tropes of Greco-Roman historiography - named rulers, geographical anchoring, eyewitness conventions, vivid presentation, inclusion of alternative reports - precisely “in order to make myth look like history”. Litwa calls this “mythic historiography.” I like this term and think that it’s genuinely useful.

And Dennis MacDonald, working at Claremont over the past twenty-five years, has built out a thesis he calls “mimesis criticism” in The Homeric Epics and the Gospel of Mark (2000) and a long string of follow-ups: that the Gospel of Mark is a deliberate Christianized rewriting of episodes from the Iliad and Odyssey, e.g. the calming of the sea, the feeding miracles, the Gerasene demoniac, Joseph of Arimathea, Bartimaeus, etc. Now, the maximal version of MacDonald’s thesis (Mark as scene-by-scene Homeric anti-epic) is contested. But the minimal version (Homer as one significant template among others) has gained a certain amount of credibility.

What makes this group of scholars “heavy” is the academic provenance and the convergence. Top universities, major academic presses, conclusions that (more or less) reinforce each other. When Walsh and Litwa, working from completely different methodological starting points, both end up at “the Gospels are educated literary productions deploying Greco-Roman conventions, not community memory,” that’s something historicism has to reckon with.

The complication, from the standpoint of mythicism, is that none of these scholars actually argues Jesus didn’t exist. Rather, they argue that the Gospels are not what mainstream historicism has often taken them to be. This is a narrower claim. Walsh, Litwa, and MacDonald remain methodologically historicists. What the group provides is - the most rigorous alternate account of how the Gospels were “constructed”, without committing to whether the constructions did or didn’t have a historical referent. That makes their collective work simultaneously the most impactful to maximalist historicism (which argues that the Gospels are reasonably reliable biographies) and the most resistant to mythicist appropriation (since it doesn’t actually argue for mythicism). It’s an fun and interesting place for academic literature to land. I’m not sure the field has yet figured out what to do with this.

Where the case stands

So where does all of this leave us?

The five “schools or groups” I discussed don’t really agree with each other. They have different methods, different focal evidence, and different conclusions. But from a high-level perspective, they relate to each other in ways that “compound”. The celestial-Christ camp provides a motivated explanation for Pauline silence on biography. The literary-construction model provides an account of where Gospel narrative material comes from (or borrows from), while the Greco-Roman literary critics independently confirm the literary character of that material from a different methodological angle - Hebrew templates and Greco-Roman templates pointing the same direction. The late-dating school provides a chronological framework in which the historicization process had more time to develop and was more thoroughly mediated by 2nd-century Christian literary culture (Marcion, Pauline forgery, etc.) than the standard dating allows. The pioneers illuminate the spectrum of options between the binary framing of Jesus exists vs non-exists.

We don’t get a knockout case against historicity from any of this. But we get something. The skeptical position has moved, over the past twenty-five years or so, from “isolated 19th and 20th-century cranks” to “a recognizable cluster of contemporary academic programs, published at the field’s major presses, converging from independent methodological directions.” That shift in the institutional ecology of the debate has not been (adequately) registered by mainstream historicism. Some of the default treatment of mythicism (e.g. Ehrman, Casey) was still caught up in referencing Frazer and Drews, as opposed to engaging with someone like Carrier, much less with Walsh, Vinzent, or Livesey, or the others I’ve mentioned.

When we step back and look at this - it could be argued that the cumulative skeptical case is strongest on three points. The first is Pauline silence on biography, which is “genuinely odd” in a way that standard historicist responses don’t fully address. The second is the influence of prior literary material on the Gospels - Hebrew scripture, Greco-Roman convention, Homer - which is undeniable, internally consistent across multiple independent literary-source research programs, and also substantially under-addressed in historicists’ work. The third is the late-dating challenge, which is (shall we say) uncomfortable - even if it isn’t ultimately accepted, the fact that foundational evidence is now seriously contested by multiple scholars at major academic publishers means the historicist case rests on less stable ground than it used to.

Similarly, the cumulative skeptical case is weakest on three corresponding points. The first is the Galatians 1:19 reinterpretation of James’ identity - the celestial-Christ school’s load-bearing response to the strongest single historicist datum - which remains forced. The second is the Josephus Antiquities 20 passage, which has no (comparably elegant) skeptical explanation - the standard move is usually to suggest the text “may” have been altered, which is the sort of move you reach for when nothing else works. And the third is the cosmic-Christ-then-historicized approach, which requires fast and uniform historicization across geographically scattered communities within roughly forty years or so (or, if we use the late-dating account, within a narrower mid-2nd-century window), and that timeline is faster than any appropriately analogous case we have.

5. Synthesis and Wrap-up

If we treating the late-dating thesis as “live” but not yet established (i.e., keeping Paul and the Gospels roughly where standard scholarship places them date-wise):

The Pauline “anchor” remains, but with less weight than usually claimed. Galatians 1:19 is still the strongest single datum, and the mythicist reinterpretations remain strained. But “taken at face value” is a more loaded phrase after Detering, Vinzent, and Livesey than it used to be.

Josephus 20 adds an important independent line. A non-Christian Jewish historian writing within living memory, casually referencing Jesus as the brother of a recently-executed James, survives every serious critical examination except determined mythicist scrutiny (which doesn’t have a clean answer for it).

Plausibility of existence remains high. The “minimal historical Jesus” fits a documented type. We’re not being asked to believe in a unique, sui generis figure. Rather, Jesus is another instance of a category we already know existed at the time.

The Gospels are (much?) weaker witnesses than they used to be. The literary-construction school and the Greco-Roman literary critics together produce a picture of Gospel narrative as educated literary craftsmanship deploying historiographic conventions, drawing on Hebrew scripture, Homeric epic, and other sources, and designed to “make myth look like history”. If the late-dating scholars are even partially right about Marcionite priority, that picture extends in time as well. The “criteria of authenticity” that produced 20th-century reconstructions (the Jesus Seminar, Meier’s Marginal Jew, Crossan and Borg, etc.) are then harder to apply with confidence.

The “defensible” historical Jesus shrinks dramatically. What can then be affirmed with reasonable confidence? Well - there was a Galilean Jewish apocalyptic preacher, probably baptized by John, who attracted some followers, was crucified under Pontius Pilate around 30 CE, and had a brother named James who was a leading figure in the early movement until his own execution in 62 CE. That’s it - the core of what we can state with a reasonably high level of confidence.

My own probability estimate, being generous to the mythicist side is somewhere in the 85-90% range for the bare existence of a historical figure. Lower than the conventional 99%, which I don’t think is epistemically defensible anymore. The literary turn weakens Gospel-based confirmations. The late-dating challenge indicates standard chronological anchors are potentially fragile. Overall, the cumulative academic skeptical case is robust enough that confident dismissal is no longer the right move.

Keep in mind - the “existence claim” is the only thing I’d estimate at that level. Almost everything else the Gospels say about Jesus should be held with (much?) lower confidence.

By the standards historians normally apply to marginal first-century figures: the evidence is adequate for “existence.” For biographical detail, the evidentiary standard for making strong claims on specifics is not really met - there’s too much literary mediation, not enough independent confirmation.

I also do not take any position on theological claims (resurrection, divinity, miracles, etc.) - these fall outside the scope of the historical method in the first place.

The Bottomline

The deductive conclusion: of the two main possible hypotheses - (a) a historical figure who became the object of worship in a new religion, or (b) a mythological or celestial figure who was historicized into a biography - hypothesis (a) is still more probable on the available evidence and by normal historical standards. But the case is closer than mainstream treatments admit, primarily because the broader critical literature has corroded the (secondary) evidence the historicist case used to draw on, and because the late-dating challenge puts non-trivial probability on a world where even the primary evidence is shakier than assumed.

It does appear that the “content” of the historical figure of Jesus was thoroughly absorbed into theological and literary construction within a generation or two, to the point where the defensible historicist position is much closer to “yes, there was a person” than to “yes, the Gospels reliably describe that person.” That’s a more uncomfortable answer than either side typically wants. Apologetic historicists want a robust, reconstructable Jesus whose teachings we can know. Strident mythicists want no person at all. The historical-critical method, applied honestly and informed by the best recent scholarship, gives us neither. (This is an old standard refrain in the “quest for the historical Jesus”.)

Thanks for reading! Subscribe for free to receive new posts and support my work.

One of the strangest facts about the universe - strange enough that I still find it amazing 25 years after first learning about it - is that mass is “not” an intrinsic property of matter. It’s not a thing particles are born with. At high enough energies, the various mechanisms that give particles mass simply haven’t kicked in yet. Then, as the universe cools, nature flips a series of switches - and presto! - we have mass.

What’s even more interesting is that mass doesn’t come from just one source. It comes from a small grab bag of physical mechanisms. Now, the one we tend to hear about in pop science (looking at you, Higgs boson) is responsible for almost none of it. The mass of nearly everything you can touch, see, or eat comes from somewhere else entirely - and the somewhere-else is much weirder.

Let me walk through where mass actually comes from, going from the famous-but-tiny source up to the obscure-but-enormous one. The percentages I’m about to throw around are approximate - meaning they’re primarily meant to give you a sense of scale.

1. The Higgs Mechanism (approx. 0.000001% of all mass)

The Higgs is, in some sense, the celebrity mass-generating mechanism, despite being responsible for an essentially rounding-error fraction of the mass in the universe. (Good public relations matters!)

What it actually does: at high energies, the electroweak force is unified - that is, two of the four forces in the universe, electromagnetism and the weak force, are the same thing, and their “force-carriers” (the W and Z bosons, and the photon) are all massless. When the universe cooled below a certain threshold (what we call the “electroweak symmetry breaking” energy), the Higgs field developed a (non-zero average) value everywhere in space. The W and Z bosons got their mass by interacting with that field. The photon stayed massless because of the specific structure of how the symmetry broke.

That’s mostly it for “this particular” contribution from the Higgs. Some bosons got heavy. The W weighs in around 80 GeV and the Z around 91 GeV, but there are vanishingly few W’s and Z’s hanging around in everyday matter, so the global contribution to “mass of the universe” from this mechanism is microscopic (read negligible).

(But.. the Higgs has a second, slightly more important job, which we discuss next.)

2. Yukawa Couplings (approx. 1% of all mass)

Here’s where the Higgs does its second job. The same Higgs field that gives W and Z their mass also couples to the “fermions” - the matter particles, as opposed to the force-carrying bosons that I mentioned in the previous section. Fermions include the quarks and the leptons (the electron is a lepton). The strength of the Higgs’s “coupling” to each fermion is called a “Yukawa coupling”, and it’s where the “bare” masses of the fermions come from. The electron has a tiny Yukawa coupling, so it’s light. The top quark has an enormous one, so it’s more a beast (relatively speaking).

But here’s the thing about normal matter - it’s almost entirely made of “up and down quarks (in protons and neutrons) plus electrons”. That’s it - that’s mostly what you and me and the stuff around are made of. The up and down quarks have Yukawa-derived masses of only a few MeV each. That’s tiny. A proton has a mass of about 938 MeV, but its three valence quarks (two ups and a down) account for maybe 10 MeV of that. So just about 1%.

But wait - if you add three quarks to get a proton, how does that only account for 1% of the proton’s mass? Let’s discuss that next.

But to summarize, if you’re keeping score so far: the Higgs mechanism, including all of its Yukawa couplings to the matter we’re made of, accounts for roughly 1% of normal matter’s mass. That’s not nothing, but it’s not where the real action is.

3. QCD Binding Energy and Chiral Symmetry Breaking approx 98–99% of all mass)

This is the part that blew my mind when I first learned it 25 years ago, and it sort of still does today. (This is really why I wrote this little essay in the first place.)

The remaining (approx. 99%) proton/neutron mass comes from what I call “gluon shenanigans” - that’s a technical term I made up and personally insist on using. Remember how we talked about two of the four forces in natures in the previous sections? Well, the third of those forces - the “strong” force, described by quantum chromodynamics (QCD), confines quarks inside protons and neutrons. The gluon field that mediates the strong force does not sit politely in the background - it churns. (The gluon is a boson, a force-carrier particle, like the photon.) Quarks emit and absorb virtual gluons. Gluons emit and absorb each other. There’s an enormous amount of energy stored in this churning, confined sea, and by E = mc², that energy IS… mass!

When you weigh a proton, you are mostly weighing the kinetic and field energy of a confined, roiling sea of gluons (and quark–antiquark pairs). The valence quarks (the up–up–down trio we talked about earlier and in all the textbooks) are only a minor part of the story.

Now, there’s a second contributing factor inside that mechanism that is worth highlighting, because it’s the conceptual heart of the whole thing - what physicists call “chiral symmetry breaking”. (Stay with me here for the next few paragraphs, which may seem a little technical.)

Here’s the rough sequence. When electroweak symmetry breaking (EWSB) happened in the early universe, it gave quarks their initial small bit of mass through their Yukawa couplings to the Higgs (as discussed in section 2 above). A few MeV each, as we said.

Then, at lower energies, QCD took over. And something remarkable happened in the strong vacuum.

Just as the Higgs field picked a non-zero value when electroweak symmetry broke, the QCD vacuum did its own version of the same trick - it spontaneously “picked a direction” in a related abstract space involving quark types and their chiralities (left-handed vs. right-handed). The technical term is the afore-mentioned “chiral symmetry breaking”.

The upshot of all of that is - quarks moving inside a proton or neutron behave as if they have a much larger effective mass than their bare Higgs-given mass. The bare mass was a few MeV; the effective mass is a few hundred MeV. Two orders of magnitude bigger. And so most of what we call “the mass of a proton” is, in a very real sense, gluon-mediated field energy converted into mass via E = mc².

(Gluon shenanigans, as I said! The Higgs gives the quarks a tiny dollop of mass. QCD takes that dollop and inflates it through the dynamics of the strong vacuum into the effective masses that actually make up hadrons.)

This is the most important mass-generating mechanism in the universe, by a country mile, er, galaxy parsec. Almost every gram of normal matter you have ever touched gets its mass from here.

4. Electromagnetic and Nuclear Binding Energy approx. minus 0.0001%)

Now for a small twist: when you bind atoms or nuclei together, the total mass actually “goes down”.

A hydrogen atom is slightly less massive than a free proton plus a free electron, because the binding energy of the system has to be paid for in mass-energy. If the binding energy is E_bind, the bound system is lighter by E_bind / c². The effect is tiny - fractions of an electronvolt for chemical bonds and a few MeV for nuclear binding - but it’s there.

This is why nuclear fusion releases energy. When you fuse light nuclei into heavier ones (well up to iron, anyway), the resulting nucleus is slightly lighter than the inputs, and the difference comes out as kinetic energy and radiation. The sun is, quite literally, slowly losing mass and converting it into light and heat for our benefit. (A stellar performance!)

(Technical note: nuclear binding is actually a much bigger effect “per nucleus” than the 0.0001% figure above suggests - helium-4 is about 0.7% lighter than the protons and neutrons that make it up. But the cosmic baryon census is dominated by hydrogen, which has nothing to bind a nucleon to, so the “global” mass deficit from binding stays tiny.)

So binding doesn’t “add” mass to the universe; it very slightly “subtracts” from it. Which is a fun little inversion if you’ve spent the last three sections thinking of mass as something that only ever gets accumulated.

5. Neutrino Masses (a sliver, mechanism uncertain)

Neutrinos (a type of lepton and cousins to electrons) have mass. We know this from observing neutrino “oscillations”, which require at least two of the three neutrino species to have non-zero rest mass. The actual values are still not pinned down, but they’re (very) tiny - probably below a fraction of an electronvolt each.

What’s strange (even exciting for researchers) is that we don’t yet know exactly how neutrinos get their mass. The leading candidates are a family of related “seesaw mechanisms” (Type I, II, and III, with subtle differences). The name is actually a useful analogy for the basic intuition. Imagine a playground seesaw with a much heavier kid on one end. The heavier kid sits very low, and the lighter kid is pushed way up. In the physics version, the lightness of the observed neutrino is balanced against the heaviness of some new physics at higher energies - and the bigger the heavy thing, the lighter the neutrino we see.

(Technical note: The Type I version does this most directly, by mixing left-handed neutrinos with hypothetical heavy right-handed partners. Type II is a variant that involves a heavy Higgs triplet whose interactions give the same kind of suppression by a different route.

Anyway, there are a few candidate explanations on the table - but none have been experimentally confirmed yet.

So there’s this neutrino mass-generating mechanism out there in the universe, contributing very little to the total mass budget but raising genuinely interesting questions about physics beyond the Standard Model. (If you want to know what the next physics revolution might look like, the neutrino sector is one of the main places to.)

But What About Dark Matter?

One important caveat before I wrap up. Everything I’ve said is about “normal” matter - the stuff made of protons, neutrons, and electrons - stuff we see, touch, and experience. But normal matter accounts for only about 15% of the matter in the universe. The other approx. 85% is dark matter, which we know is out there (galaxies wouldn’t hold together without it) but which we have not yet identified. Whatever dark matter is, it presumably also has a mass-generation story - and there are plenty of candidate stories on the menu (axions get their mass one way, WIMPs another, sterile neutrinos a third) - we just don’t know which one applies. So consider this essay an account of where the mass in the visible part of the cosmos comes from, with a polite shrug toward the rest. (Mysteries of this sort are just puzzles about reality that are yet to be solved.)

The Bottomline

Totaling it up: the famous Higgs mechanism - including all the Yukawa couplings - accounts for around 1% of normal matter’s mass. The remaining 99% or so comes from the dynamics of the strong force - gluon field energy plus the spontaneous breaking of chiral symmetry in the QCD vacuum. Atomic and nuclear binding then shaves off a tiny fraction. Neutrinos contribute a sliver by a mechanism we haven’t yet pinned down (yet).

What I find genuinely beautiful about this is the layering. Each mechanism kicks in at its own energy scale, on top of the previous one, like a series of phase transitions stacked one above the other. (A phase transition is the kind of switch you see when water freezes into ice - a sudden, qualitative shift in the state of a system as some parameter, like temperature, crosses a threshold.) The early universe didn’t have mass in the way we have mass today. At high enough temperatures, the Higgs field hadn’t yet developed its non-zero value, and the QCD vacuum hadn’t yet condensed its chiral structure. As the universe cooled, mass turned on, one mechanism at a time, until eventually we got the proton-and-electron menagerie that you and everything around you are made of. It’s like flipping on a series of switches in a control room that give you (voila!) the universe as we see it.

Anyway - the next time someone tells you “the Higgs gives things mass,” you can nod politely and then quietly note that the Higgs is directly responsible for maybe 1% of the matter in you. The other 99% is gluon shenanigans - quarks confined in a roiling field, all that energy made manifest as mass by Einstein’s most famous equation.

A truly marvelous and wonderful fact about our universe.

The causal closure of the physical (CCP) is the claim that every physical effect has a sufficient physical cause. That is, the physical world, in causal terms, is a closed system. Nothing non-physical reaches in to push the particles around. Whatever happens physically can be fully accounted for (in principle) by other physical events.

The thesis has been around in some form since the 17th century, when the mechanical philosophy crowd (Descartes, Hobbes, Boyle, Newton) started imagining the universe as a vast clockwork. Descartes himself ended up famously stuck on the question, because his commitment to an immaterial mind required some way for that mind to interact with the body. He took a swing at it with the pineal gland, which has gone down as one of philosophy’s longest-running punchlines. (I struggled to avoid putting a Descartes before the horse pun here.)

The early modern era produced a surprising variety of attempts to keep both a closed physical world AND a non-physical mind without direct interaction. Leibniz’s pre-established harmony, Spinoza’s parallelism, Malebranche’s occasionalism - each was, in its own way, ingenious. (And no, I don’t hate occasionalism, mostly because it gets to the answer “God does it” in a way that’s at least honest about the move). They were also, by later standards, deeply unsatisfying as scientific or metaphysical proposals. By the 19th century, most serious thinkers had concluded that the cleanest move was to drop the non-physical mind altogether.

The modern, sharper formulation of CCP comes mostly from David Papineau, who argued in the 1990s and 2000s that 20th-century physiology has effectively settled the question empirically. Every nerve impulse, muscle contraction, and neurochemical cascade we’ve measured turns out to be fully accounted for by physical and chemical mechanisms. Where we’ve looked carefully, we haven’t found gaps that require introducing non-physical intervention. Jaegwon Kim then made the philosophical machinery around CCP famous with his “causal exclusion argument,” which uses CCP to argue that mental properties either reduce to physical properties or don’t really do any causal work.

Now, why does any of this matter? Because CCP is the load-bearing wall of physicalism. If it’s true, the universe doesn’t have any spooky causal back doors. Mental causation has to be physical causation in some form. Free will, if it exists at all, and even if it can be framed in compatibilist terms, has to be a physical phenomenon. There’s no place for ghosts, élan vital, divine intervention, or any other ingredient that exists outside the physical order. CCP is what allows physicalism to be a substantive thesis rather than a holey gesture (bad pun intended).

It’s also contentious for the same reason. CCP rules out a lot of things many people are very attached to. So, it has attracted surprisingly high-quality skeptical attention from philosophers (Tim Crane, E.J. Lowe, Jessica Wilson, Helen Steward) and a few prominent physicists (Roger Penrose, Henry Stapp, George Ellis - and historically, Eugene Wigner and von Neumann). The most famous theoretical attack is Hempel’s dilemma. The most popular empirical attack is the appeal to quantum measurement and indeterminacy. The most sophisticated contemporary attack is Russellian monism, which is rather different from the others and which I’ll treat as the strongest opposition. The most rhetorical attack is the appeal to consciousness.

My claim is that none of these attacks ultimately succeeds, though some of them (especially Russellian monism) deserve more careful engagement than they typically get from physicalists. CCP is on solid ground. Let me explain why.

The argument I’m going to make has two parts. The first is “methodological” - that physicalism, properly understood, is a coherent thesis that survives the standard theoretical objections. The second is “empirical” - that the actual scientific evidence overwhelmingly supports the CCP. These do different work, and I’ll try to keep the seam visible rather than letting them blur together.

1. Hempel’s Dilemma is Overrated

The most famous theoretical objection to physicalism is Hempel’s dilemma. The argument runs:

When you say “every physical effect has a sufficient physical cause,” what do you mean by “physical”?

Option A: You mean current physics. In that case, your claim is almost certainly false. Current physics is incomplete - we have no UV-complete theory of quantum gravity, the Standard Model doesn’t describe dark matter or dark energy (which together make up roughly 95% of the cosmic matter-energy budget), the measurement problem in quantum mechanics is unresolved, etc.. Saying CCP holds with respect to current physics is like saying it held with respect to Newtonian mechanics in 1890. The history of science laughs at you.

Option B: You mean a future, completed physics - whatever a finished science of the physical world will look like. In that case, your claim is true by definition but completely empty. “Whatever turns out to cause physical events will turn out to be physical, because we’ll define it that way.” That’s not a thesis, that’s a tautology.

This dilemma sounds devastating in the abstract. It dissolves on close inspection.

Physicalism, properly understood, is NOT a claim about the “content” of any specific physical theory. It’s a claim about the “methodology” and “idiom” of physical theorizing. What EVERY successful extension of physics has shared - from Newtonian mechanics to general relativity to quantum field theory to the Standard Model to whatever quantum gravity ultimately is - is a particular way of describing the world: structural, mathematical, third-personal, mind-independent, quantitative. The world gets carved up into entities with measurable properties whose dynamics are governed by formalizable laws. That’s the recognizable shape of physics, and it has actually been remarkably stable through every revolution.

So Hempel’s dilemma is attacking a strawman. Physicalists aren’t betting on the truth of the Standard Model as it stands in 2026. They aren’t making the trivial claim that “physical = whatever physics says.” They are making the substantive claim that the world is exhaustively describable in this kind of idiom. That claim has real content. It rules out things. It rules out, for instance, fundamental teleology, irreducible mentality, or Cartesian souls. None of those fit the methodological mold, regardless of how physics evolves.

And here’s the inductive payoff. Every revision of physics in the last 350 years has stayed within this methodological framework. Going from Newton to Einstein didn’t require positing souls. Going from classical to quantum didn’t require positing élan vital. Adding gauge fields, supersymmetry, inflation, neutrino masses, dark matter candidates - all of these stayed inside the structural/mathematical idiom. The track record is unbroken. A skeptic who insists future physics will need to break this pattern owes us a positive argument and some actual evidence. Neither has been forthcoming.

This is sometimes called “methodological physicalism” or “structural physicalism,” and figures like Barbara Montero and Jessica Wilson have spelled it out more carefully than I’m doing here. The point is that Hempel’s dilemma only bites a particularly naive strawman formulation of physicalism. The grown-up version is fine.

What this argument does not do, however, is provide empirical support for CCP. It defends the “coherence” of the physicalist thesis against the charge of triviality. The empirical case is a separate matter, which I’ll get to in section 4. I want to flag this seam now because it matters - methodological physicalism on its own doesn’t tell you the world is causally closed, only that the claim is “meaningful”. The evidence has to do its own work.

2. Quantum Measurement is Not a Back Door

Probably the single most popular move in (undergraduate-level) skepticism about CCP is the appeal to the quantum measurement problem. The argument runs something like: “But quantum mechanics is indeterministic! And the wave function collapses when an observer looks at it! Doesn’t that mean physics isn’t causally closed?”

No. Let me explain why.

First, indeterminism is not a problem for CCP. The thesis can be (and standardly is) formulated probabilistically: every physical effect has its complete causal antecedents in the prior physical state, even if those antecedents are stochastic rather than deterministic. Quantum mechanics specifies the probability distributions of possible outcomes with extraordinary precision. The fact that the outcomes themselves aren’t deterministically settled doesn’t introduce non-physical causation. It just means the physical causes are stochastic.

Now, there’s a real worry worth flagging here. Tim Maudlin and others have argued that “probabilistic sufficient causation” isn’t a coherent notion - that if your causes don’t fix the outcome, they aren’t sufficient in any robust sense. I think the worry is real but not fatal. The work CCP needs done is “nothing physical is left out and nothing non-physical is added,” not “the outcome is determined.” Probabilistic closure is closure in the relevant sense: the universe, as far as causal influence is concerned, has no outside. The dice may be loaded by chance, but no one outside the universe or a non-physical entity inside the universe is rolling them.

Second, every serious interpretation of quantum mechanics is a “physical” theory. Many Worlds denies collapse altogether and treats the wave function as the complete description of reality. Bohmian mechanics adds deterministic guiding equations and is fully physical. GRW and other objective collapse theories propose specific physical mechanisms that produce collapse without any observers in the picture. Decoherent histories explains the appearance of classicality through environmental entanglement, a physical phenomenon. Each of these interpretations/theories has internal problems (the probability problem in Many Worlds, relativistic generalization in Bohmian mechanics, the tail problem in GRW, etc.), and the debate over which is correct is genuinely unresolved. But all of them are within the structural/mathematical idiom of physics and physicalism. The measurement problem is a problem “for” physics, not evidence “against” the sufficiency of physics.

Third, the Wigner-von Neumann interpretation (which proposes that a conscious observer plays a special causal role in collapse) is the one interpretation that “might” threaten CCP. It deserves a fairer hearing than it usually gets. Philosophers like David Albert and Tim Maudlin take measurement seriously enough to think the deflationary “decoherence solves everything” response is too quick. They’re right that decoherence by itself doesn’t fully resolve the measurement problem; it explains why we don’t observe macroscopic superpositions while leaving open what happens to the unobserved branches.

So why reject Wigner-von Neumann? Not because it doesn’t make sense. But because it carries metaphysical costs the alternatives don’t. It requires drawing a principled line between “conscious” and “non-conscious” systems, with the former having the unique capacity to collapse wave functions. No physical theory tells us where that line falls. (Is a dog conscious enough? A rat? A thermostat? An LLM?) The Wigner interpretation has to take consciousness as a primitive in physics, which is the sort of move you reserve for cases where nothing else works. Many Worlds, Bohmian mechanics, and objective collapse theories all do the explanatory work without this commitment. Wigner-von Neumann is coherent. It just isn’t a “forced move”, and accepting it because consciousness “feels special” is just bad theorizing.

The deeper point: even if Wigner-von Neumann turned out to be right, it wouldn’t show that CCP is false. It would show that consciousness is part of fundamental physics, in which case “physical” would expand to include it, and we’d be back to a closed (now consciousness-inclusive) physical system. Whether that’s a vindication or a refutation of physicalism depends on how you’ve drawn the line. Either way, it’s not the kind of clean win for the anti-physicalist that pop-science presentations on social media make it out to be.

3. Consciousness Doesn’t Live in the Wavefunction

Let me sharpen this because the confusion is unusually persistent.

The reason people keep wanting consciousness to play a role in physics is that there’s an emotional appeal to the idea. The mind feels like it should be metaphysically special. The hard problem of consciousness (which I’ve argued in a previous essay is mostly a confusion) makes phenomenal experience seem like it cannot fit into the physical order. So when you encounter quantum mechanics and notice that “observation” plays a role in the formalism, the temptation is to connect the two and say: “Aha, here’s where consciousness gets back into physics.”

This is a category error.

The “observation” in quantum formalism is a technical term referring to a specific kind of interaction (decoherent, irreversible, recording). It has the same structural relationship to the rest of the formalism as “force” has to classical mechanics. You don’t claim that consciousness plays a special role in Newton’s laws because forces “act on” objects. The verb is doing technical work, not metaphysical work. Same with “observation” in QM.

Empirically, “measurements” happen all the time without any conscious observers. They have been happening since the dawn of our universe, and possibly even before that. A photographic plate records photon arrivals. A Geiger counter registers radioactive decay events. A dust grain in interstellar space gets entangled with photons and undergoes decoherence with no human anywhere in the causal chain. In the actual formalism, measurement is just any sufficiently decoherent interaction with a sufficiently large environment.

Decoherence theory has clarified this enormously over the past few decades. The transition from quantum superposition to apparent classical definiteness happens through unitary “entanglement with environmental degrees of freedom”, and quickly too - in a very small fraction of a second for macroscopic objects at room temperature. No mind needed.

Now, what about Penrose-Hameroff and the proposal that the brain exploits macroscopic quantum coherence in microtubules? Max Tegmark famously calculated decoherence times for microtubules at around 10⁻¹³ seconds - many orders of magnitude faster than the timescales of neural firing. Hameroff and Penrose have responded with modifications to their model, and recent experimental work (Bandyopadhyay and collaborators) has reported vibrational signatures in microtubules. The exchange is ongoing.

I don’t want to overstate the case. It isn’t that Orch-OR has been definitively refuted. What’s true is that the burden of evidence sits heavily on the Orch-OR side, and the central empirical claim (that the brain depends on macroscopic quantum coherence at cognitively relevant timescales) remains unsupported by mainstream neuroscience.

Quantum biology is a real and growing field. Photosynthesis exploits coherence at picosecond timescales in tightly engineered protein environments. Avian magnetoreception probably involves radical-pair chemistry (and yet we call them bird-brained!). But these are highly specialized molecular systems, not cognitive ones, and nothing in the quantum biology literature gives reason to expect coherent quantum computation at the scales relevant to thought. The brain runs hot, wet, and noisy. Coherence dies fast in those conditions, and the cognitive timescales (milliseconds to seconds) are vastly longer than anything (quantum and) coherent could plausibly last in tissue.

So consciousness has no established role in quantum measurement, and the most prominent proposals to give it one face severe empirical headwinds. The category error remains a category error. Decoherence happens regardless of minds. Consciousness lives in the integrated information processing of the cortex, not in collapses of the wave function. (The rest of the body may well be involved in that processing too, but let’s keep things simple for now.)

4. The Core Theory Argument: Where Closure Becomes Empirical

Here is where the second argument of this essay kicks in. The argument in section 1 was “methodological” - it established the “coherence” of CCP against Hempel’s dilemma. But coherence isn’t enough; we want empirical support. That’s a separate move, and the empirical case is what I want to lay out now.

As I’ve noted in previous essays, Sean Carroll has made this argument forcefully in his paper “Consciousness and the Laws of Physics“ and at book length in The Big Picture. The idea, originally articulated by Frank Wilczek, is that we already know the laws of physics that govern everyday phenomena - everything that happens at the temperatures, energies, and timescales relevant to biological matter on Earth. Wilczek calls this collection of laws the “Core Theory“. It’s the Standard Model of particle physics plus general relativity in its weak-field limit.

The Core Theory has been confirmed to extraordinary precision across an enormous range of conditions. Particle accelerators have probed energies up to about 14 TeV. Cosmological observations probe a vastly larger range of scales. Within the energy regime relevant to biological matter (which is many orders of magnitude lower than what’s been probed), we have detected “every” relevant interaction and “every” relevant field. We “know” what’s there. There aren’t undiscovered particles or forces hiding in the regime where chemistry and biology happen, because if there were, they would have shown up in our experiments. (Could there be a quantum field that doesn’t interact with matter or energy at all? Well if it doesn’t interact, it’s not relevant to discussions of physical causality related to organic life.)

Now plug this into the question of CCP. Suppose, for the sake of argument, that there’s some non-physical ingredient that affects what happens in your brain - a soul, a panpsychic proto-experience, a top-down causal influence from consciousness, divine action on neurons, or whatever your favorite anti-physicalist proposal is. For this ingredient to do any actual work, it has to interact with a quantum field. It has to push particles around, and make them move in a way it wouldn’t have moved given just the physics. (Particles are just excitations of quantum fields.) This would manifest as a deviation from the Core Theory at the very energies where neural processes operate, energies that we have probed experimentally for decades.

We have looked for any such deviations. We have looked very hard. We have not found ANY. The behavior of matter at biological energies is exactly what the Standard Model and gravity predict.

The standard response from the anti-physicalist is some version of “well, the deviation might be too subtle to detect” or “maybe the non-physical thing operates in a regime we haven’t probed.” Let me take this carefully, because it isn’t a pointless objection.

There are two ways the response could go. The first is that the non-physical influence is too small to detect at the level of individual field or particle interactions but accumulates into systematic effects at the level of whole organisms. This sounds initially plausible until you ask what mechanism would produce such accumulation without showing up in the statistics. If the effect is systematic (consistent across many fields/particles), it should be detectable in averages, correlations, or precision tests. If it’s random, it can’t be organized enough to produce coherent behavior like thought or action. The objector is asking for an effect that’s simultaneously systematic enough to matter and stochastic enough to evade detection. Those two requirements are in tension.

The second response is that the non-physical influence operates on degrees of freedom we haven’t directly probed - perhaps something like a modification of phase relationships or higher-order correlations. This is more interesting, and it gestures toward the kind of view I’ll discuss next under Russellian monism. But even granting it, it doesn’t violate CCP in the standard sense. It relabels what counts as physical. (We’ll get there.)

The empirical conclusion stands. There is no detected deviation from the Core Theory in any system, including biological ones. Every successful theoretical prediction in physics has been a prediction within the closed physical framework. This isn’t a proof of CCP (empirical evidence rarely is), but it’s the strongest kind of inductive case science can offer.

The dilemma I’ve discussed in my previous essays captures this nicely. Any anti-physicalist proposal must take Option A (modifies Core Theory dynamics, would be empirically detectable, is in fact undetected, would be the discovery of the century if true, Nobel prizes guaranteed) or Option B (leaves Core Theory dynamics intact, hence is causally inert, hence is epiphenomenal and explains nothing). The dilemma is sharp, and it’s hard to escape. As we’re about to see, there’s one sophisticated attempt to slip between its horns.

5. Russellian Monism: The Serious Opposition

The strongest contemporary opposition to standard physicalism isn’t dualism, and it isn’t anything that violates the Core Theory. It’s Russellian monism. Anyone defending CCP should engage with it seriously, because the easy refutations don’t apply.

As I mentioned in my earlier essay against panpsychism, the view traces back to Bertrand Russell’s analysis of physics in The Analysis of Matter (1927) and has been developed by Galen Strawson, Philip Goff, Daniel Stoljar, and (in some moods) David Chalmers. It goes like this.

Physics describes the world entirely in terms of “structure” and “dispositions”. It tells us what entities do, how they relate, how they transform under various operations. It does not tell us what those entities “intrinsically are”. Mass, charge, spin - these are characterized by what they do in equations, not by their “categorical nature. So physics, on this reading, leaves a gap: the intrinsic nature of the physical is “unspecified”. Russellian monists then propose to fill this gap with proto-experiential or experiential properties. The intrinsic nature of physical entities, on this view, is (or includes) something mental.

Here’s a useful analogy. You can describe a chess game completely in terms of moves, positions, and rules. That’s the structural description, and it’s enough to play the game. The Russellian asks: but what are the “pieces” made of? Standard physicalism (supposedly) says: structure is enough; there’s nothing more to know that’s relevant to the game. Russellian monism says: there has to be something the structural description is “about” - some intrinsic stuff that has these relational properties. And then they propose that this intrinsic stuff is (proto-)mental.

Notice what this view doesn’t claim. It doesn’t claim that physics is wrong, that the Core Theory is incomplete, or that some non-physical force violates closure. The dynamics are exactly what physics says they are"! The Russellian agrees with everything in section 4. They just hold that the structural description doesn’t “exhaust” reality, and that the categorical bases of physical properties are (proto-)mental.

This survives the Option A/B dilemma because the Russellian denies the dichotomy. The view isn’t Option A - it doesn’t modify dynamics. It also isn’t Option B - the intrinsic mental properties just “are” the categorical bases of the causal facts, so they’re not extra causal influences, they’re constitutive of the existing ones. The Russellian threads the needle.

So why am I not persuaded? A few reasons. None of them is a knockout on its own, but together I think they add up.

First, the structural-vs-intrinsic distinction the view depends on is contested. James Ladyman and Don Ross, in their defense of ontic structural realism, argue that there are no intrinsic properties beyond the structure - that “structure all the way down” is a coherent and arguably better view of physics. If they’re right, Russellian monism is positing categorical bases that don’t exist. The question of whether physics has unfilled categorical roles, or whether the structure is itself the ontology, is genuinely open in philosophy of physics, and Russellian monism is making a substantive bet here. To stick with the chess analogy: the structural realist says the pieces don’t have an “intrinsic nature” beyond their role in the game. They are their role. Russellian monism quietly assumes otherwise.

Second, even granting the categorical roles need filling, why fill them with mental properties? The motivation is almost always the hard problem of consciousness - the sense that phenomenal experience is so different in kind from anything else that it must be a fundamental ingredient. If you don’t accept the hard problem framing (and I don’t!), the motivation evaporates. The categorical bases, if they exist, could just as well be filled with whatever non-mental properties categorical bases naturally are. Russellian monism stops being attractive once you stop assuming consciousness is sui generis.

Third, the combination problem. If micro-level entities have proto-experiential properties, how do these combine to produce the unified, structured consciousness of a human mind? This is the panpsychist’s most famous internal difficulty, and despite a lot of careful philosophical work, no satisfying solution has emerged. Goff and others have wrestled with it; I find their attempts strained. A view that posits a fundamental property to solve one problem (the hard problem) and then can’t explain how that property produces the phenomenon it was posited to explain (unified consciousness) is in dialectical trouble. (I argued against panpsychism more broadly in an earlier essay and won’t restate all of that here.)

Fourth, parsimony. Everything physical works out as predicted by physics alone. Russellian monism adds an entire metaphysical layer (intrinsic mental properties) that does no predictive work. It’s motivated by an explanatory intuition, not by anything that would show up in the experimental record. I really have to say - Ockham’s razor doesn’t decide questions on its own, but when a theory adds ontology without adding empirical content, it deserves skepticism!

That said, Russellian monism is the most sophisticated current alternative to standard physicalism. It doesn’t violate CCP and isn’t refuted by the Core Theory argument. But it’s motivated almost entirely by intuitions about consciousness I think are confused, and it depends on a metaphysics of physics that’s contested at best. I’m not persuaded. But it’s the view that deserves engagement, not the dismissal it usually gets in undergraduate philosophy of mind or in skeptical circles in the hard sciences.

6. Emergence and Complexity Don’t Open a Door

The last major move I want to address is the appeal to emergence. The argument runs: even if CCP holds at the fundamental level, complex systems exhibit “emergent” properties (consciousness, life, intentionality, top-down causation) that can’t be reduced to lower-level physics. So maybe physics is closed at the bottom but reality is “more than physics” at higher levels.

This argument equivocates between two very different senses of emergence.

Weak emergence is the kind we see everywhere in science. Temperature emerges from molecular kinetic energy. Wetness emerges from intermolecular forces. Life emerges from organized chemistry. Consciousness emerges from neural integration. In each case, the higher-level property is a real, useful, often computationally indispensable description, but it’s not metaphysically over and above the underlying physics. The macro-state is fully determined by the micro-state. You can do top-down “causation” in the sense that talking about higher-level patterns is sometimes the most useful explanatory level (no one explains traffic jams by tracking individual electrons). But this isn’t an exception to closure - it’s multi-scale description of the same closed system at different levels of coarse-graining.

Strong emergence is the claim that higher-level properties exert causal influence not determined by, or reducible to, lower-level physical dynamics. This “would” violate CCP. It would mean the higher-level pattern reaches down and pushes particles around in ways the lower-level physics alone wouldn’t predict.

How do we evaluate strong emergence? Not by an a priori test - any test that presupposes microphysical determination begs the question against the strong emergentist. The right approach is to ask an empirical question: have we “ever” observed a case where the complete physical state of a system, together with the laws of physics, “failed” to predict the system’s macro-behavior? In the entire history of science, the answer is NO. Wherever we have checked carefully (in molecular biology, in chemistry, in neuroscience, in condensed matter physics), the macro-behavior follows from the micro-dynamics. This isn’t proof - strong emergence is a logical possibility - but we have an extraordinarily strong inductive case combined with a complete absence of positive evidence on the other side.

George Ellis has been one of the more prominent advocates for top-down causation, and to his credit, he tries to be precise. But on close reading, his examples are all examples of weak emergence dressed up in stronger language. Constraints from higher-level structure (a wing’s shape constraining airflow, a brain’s architecture constraining neural dynamics) are perfectly compatible with closed physics, because the constraints are themselves implemented by physical structure, and the dynamics are just the dynamics. Calling this “top-down causation” is a useful pragmatic description. It’s not an exception to closure.

For strong emergence to be a live option, we’d need empirical evidence of cases where micro-dynamics fail to determine macro-behavior. We don’t have such cases. The inductive case for closure, all the way up, is the same case that supports the Core Theory itself: the most thoroughly tested set of empirical claims science has produced.

7. The Free Will Question, Briefly

CCP has obvious implications for free will, and I want to acknowledge them rather than pretend they aren’t there.

If CCP holds, then libertarian free will - the kind that requires causal agency from “outside“ the physical order - is likely off the table. Your decisions are physical events with physical causes, even if those causes are stochastic. There is no extra “you” stepping in to direct the neurons. (Sorry to anyone who finds this distressing. The metaphysics doesn’t care about our feelings.)

Usually, this is the kind of conclusion that horrifies people on first encounter. I think the horror is misplaced. Compatibilist free will - the kind that requires only that your actions flow from your own deliberative processes without external coercion - remains entirely live. Daniel Dennett spent a career arguing that compatibilist free will is the only kind worth wanting anyway, and I find the argument quite persuasive. The “you” that deliberates and decides is real. It’s just made of neurons rather than ectoplasm.

A full defense of compatibilism is a different essay. The point here is just that CCP doesn’t ““destroy free will. It constrains what free will can coherently be. If you were already a libertarian free will-ist, CCP is bad news. If you were already a compatibilist or a hard determinist, CCP is what you expected. Either way, the metaphysics isn’t a disaster.

The Bottomline

Causal closure of the physical rests on two arguments doing two distinct jobs.

The methodological argument establishes the coherence of physicalism against Hempel’s dilemma. Physicalism isn’t a claim about the content of current physics. It’s a commitment to a structural, mathematical, third-personal idiom of description. Every successful physical theory in the past 350 years has fit this idiom. Future physics almost certainly will too, because that’s what makes a theory a physical theory. Hempel’s dilemma collapses once you stop conflating “physical” with “the current Standard Model.”

The empirical argument establishes that CCP is actually supported by what we know about the world. The Core Theory describes everything that happens at biological energies to extraordinary precision. Any non-physical ingredient with causal power would show up as a deviation. None has been found. The same empirical case rules out strong emergence in any system we have studied carefully. The evidence is asymmetric and overwhelming.

Two attacks I take seriously but think ultimately fail. The quantum measurement problem doesn’t introduce non-physical causation - probabilistic closure handles indeterminism, every serious interpretation is a physical theory, and consciousness has no established role in collapse. Russellian monism is more sophisticated; it doesn’t violate CCP at all but proposes that the intrinsic nature of physical entities is (proto-)mental. I’m not persuaded because the motivation depends on a hard problem framing I reject, because the combination problem remains unresolved, because the categorical-vs-structural metaphysics it depends on is contested, and because it adds ontology without adding empirical content.

None of this means consciousness, intentionality, or meaning aren’t real. They obviously are. They’re just not exceptions to closure. They’re descriptions, at the appropriate level of abstraction, of patterns in a closed physical system. Once you stop demanding that the pattern be metaphysically over and above what implements it, the air goes out of the room, and we can get back to actual science.

The world is closed. It’s also rich, layered, surprising, and full of phenomena we don’t yet understand. Those two things are perfectly compatible. The temptation to read every gap in our understanding as a gap in nature itself is the same mistake we’ve made about life, about combustion, about cosmology, about every previously-unsolved problem in the history of science. As I argued in my essay on fine-tuning, this kind of reasoning has never once turned out to be correct. I see no reason to think it will be correct here either.

Panpsychism is having a moment. After decades of being treated as a philosophical curiosity (or a relic of pre-scientific animism or something even less fashionable), the view that consciousness is somehow a fundamental feature of reality has been making its way into respectable philosophy departments, popular science books, podcasts, and even the occasional physics-adjacent academic conference.

Galen Strawson has been arguing for it forcefully since the early 2000s. Philip Goff wrote Galileo’s Error (trade book aimed at general readers) making the case for it in 2019. Annaka Harris has been popularizing the view via her writing and a podcast. David Chalmers, who basically launched the modern debate over consciousness with his 1995 “hard problem” paper, has written sympathetically about Russellian monism (a close cousin of panpsychism) for years. There is now a lively academic literature around the view, complete with its own conferences, journals, and, of course, internecine disputes.

I’ve previously argued that the hard problem of consciousness is, on close inspection, a confusion masquerading as a discovery. Panpsychism seems to be the most fashionable response to the hard problem, and a number of readers have pushed back on my earlier piece by appealing to it. So this is my attempt to push back again the pushback by taking a close look at why panpsychism fails - not just as a philosophical proposal, but as the kind of claim it actually is, which is a claim about the basic nature of the universe.

The thesis I want to defend here: panpsychism is not primarily a philosophical position, and treating it as one is part of why it gets undeserved respect. It is a hypothesis about what fundamental physics describes, and once we evaluate it on those terms, it falls flat. The motivating intuition that drives it (the worry that consciousness “can’t emerge from unconscious matter”) turns out to undermine panpsychism just as effectively as it (allegedly) undermines physicalism. And besides that internal problem, the weight of modern physics, neuroscience, and biology points the other way.

1. The Most Respectable Formulations

Let’s start by attempting to give the view a fair shake before critiquing it.

Modern panpsychism is not the cartoon version where rocks have feelings and electrons fall in love, although I think that this would make for a great webcomic. The serious contemporary versions are subtle, and they trace back (more or less?) to Bertrand Russell’s The Analysis of Matter (1927). (There may have been earlier precursors, but Russell’s framing seems to have significantly influenced modern approaches.)

Russell’s observation was this: physics describes the world entirely in terms of structure and dynamics - what things do, how they relate, what mathematical patterns they follow, etc. Physics tells us that an electron has a certain charge, mass, spin, etc. and that it interacts with other things in certain quantitatively specifiable ways. But (so the argument goes) physics never tells you what an electron “is”, intrinsically. It tells you about the relations and behaviors. The intrinsic nature - the “what it is” of fundamental physical stuff - is left as a black box. Russell suggested that consciousness might be the only intrinsic nature we ever have direct acquaintance with (since it’s our own experience), and that other things (electrons, fields, what have you) might have their own intrinsic natures of which our consciousness is a “particularly elaborate” species.

(My genealogy of the ideas here is not meant to be exact. The SEP article on panpsychism is quite good and covers this.)

Starting from that approach, we now have a fair number of modern positions that seem to tie back to Russell’s argument.

Let’s go over a few of these - note that some of these positions are close to panpyschism but not quite panpyschism:

Constitutive panpsychism (Goff is probably its best-known contemporary defender) holds that consciousness is a fundamental property of basic physical entities - quarks, electrons, fields - whatever turns out to be ontologically basic, and that the consciousness we have is “constituted” by combinations of those micro-experiences.

Panprotopsychism is a more cautious cousin: rather than saying basic stuff has consciousness, it says basic stuff has “protoconscious” properties that, when suitably combined, give rise to consciousness. The idea is to retain the explanatory power of “no consciousness from non-consciousness” without committing to electron feelings. (Feel free to insert your own favorite disappointed electron meme gif here.)

Russellian monism is the broader umbrella term for “physics describes structure, intrinsic nature is something else, that something else is (or is closely related to) consciousness.” This approach can be either constitutive or panprotopsychist in flavor.

Cosmopsychism flips the bottom-up move: the universe as a whole is the fundamental conscious entity, and individual minds are derivative “slices or aspects” of cosmic consciousness. (You can hear some Vedantic echoes here, which is fine - philosophy has often recycled.)

This is not an exhaustive list of formulations, they are the ones I found most interesting and respectable. For the most part, they aren’t claiming that your toaster is contemplating the meaning of life. They’re making a structural claim about the relationship between physics and phenomenology.

2. The Motivations (and the Strongest Case)

In my view, two motivating intuitions seem to be doing most of the work for panpsychism, and they’re worth taking seriously even if (as I’ll argue) neither survives scrutiny.

The “no emergence” intuition. This is the engine of the whole project. The idea is that you can’t get consciousness from a substrate that has no consciousness in it, any more than you can get extension from non-extended points or wetness from non-wet atoms. Strawson, for example, is quite insistent on this. If your starting ingredients are entirely dead matter, with no glimmer of experience, then no rearrangement of those ingredients will produce experience. Therefore the ingredients themselves must already have “some” experiential character.

So, I’ll say this for this intuition: it has rhetorical force. It feels right when you read it. (It’s one of the most common arguments for panpsychism I come across on X, for example.) But, as I’ll get to below, it has the unfortunate property of being self-defeating.

The Russellian “intrinsic nature” gap. This is the more sophisticated motivation. The argument goes something like this: physics gives us only the relational/structural skeleton of the world. There must be intrinsic natures grounding those relations, and we have one example of intrinsic nature on offer (our own conscious experience). Parsimony then suggests we extend that example to the rest of nature rather than positing some entirely new species of intrinsic nature for the non-mental world.

(As someone who tends to think positively of structural realism, the “physics describes structure” argument is one I can appreciate, even if I disagree with the rest of the argument.)

Now let me try and state the panpsychist case at full strength, because I think it’s sometimes dismissed too quickly. This rehashes some of the material I already covered, but please bear with me - I am trying my best to steelman this.

The strongest version (in my opinion) goes something like this. Modern physics gives us, on the structuralist reading, a mathematically beautiful description of “what fundamental entities do”. It tells us how electrons move, how fields couple, how spacetime curves. It does not tell us “what fundamental entities are”, in the sense of what their intrinsic, non-relational properties are - because, by construction, physical theory traffics only in dispositional and structural features. Thus, there is an entire ontological category (the “categorical” or “intrinsic") about which physics is, by its very methodology, silent. If we ask what fills that category, we have exactly one positive datum: our own consciousness, which we have direct acquaintance with and which does not present itself as a structural relation but as something with intrinsic phenomenal character. The parsimony argument (which panpsychists note, physicalists like me love to invoke) then suggests that we should fill the categorical-nature gap with the same sort of stuff we already have on offer (proto-experience), rather than positing some entirely novel kind of intrinsic nature about which nothing whatever can be said. On this telling, panpsychism isn’t a mystical addition to physics. It’s a “parsimonious completion” of it.

That’s a real argument. And it deserves engagement, not dismissal. The reason it ultimately fails (which I’ll develop in sections 3-7) is not that it’s silly, but that the proposed completion doesn’t actually do the explanatory work it claims, and runs afoul of the physicalist-friendly considerations the panpsychist is trying to honor.

3. The Self-Defeating Logic of Anti-Emergence

Here’s where I think the contemporary debate has been strangely lopsided.

The core panpsychist move is - “consciousness can’t emerge from non-conscious stuff.” Okay. Suppose we grant that, just for the sake of argument. Now let’s ask: how is sophisticated, unified, semantically rich human consciousness supposed to arise from the aggregation of trillions of (allegedly) micro-conscious quarks, electrons, and field excitations?

This is the famous combination problem, and it has been the chief internal headache of panpsychism since William James first raised it in The Principles of Psychology (1890). Chalmers has called it “the hard problem of combination” and, on this point, I think he’s exactly right.

But let’s examine this problem a little closer than the standard treatments do, because the strongest panpsychist response seems to miss what’s really going on.

The strongest panpsychist response runs as follows: yes, the combination problem is hard, but it’s hard in a “categorically less mysterious way” than the matter-to-mind problem. Going from micro-experience to macro-experience is at least staying within the same ontological category - we’re combining experiential stuff with experiential stuff, like making a wave from many smaller waves. Going from non-experience to experience, by contrast, is alleged to be a discontinuous leap into an entirely new ontological domain. So the panpsychist can grant that combination is hard while still claiming it’s a fundamentally less daunting problem than physicalist emergence.

This response is wrong, and it’s wrong in a specific, instructive way.

The “category preservation” reply works only if we ignore what the macro-experience actually is. Human consciousness is “unified”. It’s a single perspective, a single point of view, one set of experiences that is “had” by one subject. The transition the panpsychist needs to explain is therefore not just “combining experiences” but combining “many discrete tiny perspectives into one unified perspective”. And that transition is exactly the kind of boundary-crossing the panpsychist objected to in the first place - it’s a transition from a state-of-affairs in which there are many subjects (many discrete proto-experiences) to a state in which there is one unified subject. The fact that both states involve “experiential” stuff doesn’t make the transition trivial by any means. It might actually make it harder, because now we have to explain how “experiential facts” (which panpsychists suggest are fundamental) can be merged into a single experiential fact without violating their fundamentality.

Think about it this way. The panpsychist’s anti-emergence intuition says: experience cannot arise from a substrate that lacks it. Fine. Now apply the principle: “unified experience” cannot arise from a substrate that lacks unified experience. By the panpsychist’s own logic, you cannot get unification by aggregation. Unified experience, on the same anti-emergence principle, would have to be fundamental.

But then we’re not in panpsychism anymore - we’re in “cosmopsychism” , where the universe as a whole is the unified experience and our individual minds are decompositions of it. And this is where cosmopsychism’s apparent escape from the combination problem turns into a “decombination problem” that is, if anything, even worse. If the cosmic mind is the fundamental unit, how does it split into many private, mutually inaccessible perspectives? What “principle of decomposition” takes one cosmic experience and gives you several billion discrete human experiences, each closed off from the others? Whatever answer you give to that question is going to require exactly the kind of “many-from-one” or “one-from-many” account that motivated the combination problem on the way up.

(The Borg notably solved this problem by fiat. The cosmopsychist does not get to invoke fiat.)

The situation is problematic for panpsychists either way you run it. Run it bottom-up, and we have a combination problem that the very anti-emergence intuition we started from seems to forbid. Run it top-down, and we have a decombination problem that does the same work in reverse. As far as I can tell, no one has solved either, and the proposals I’ve read tend to either slide back into garden-variety emergence (in which case the proto-consciousness is doing no explanatory work) or to introduce new metaphysical primitives “without independent motivation”.

The honest summary: panpsychism trades one alleged emergence miracle for another. If the physicalist owes you an explanation of how subjective experience arises from neural activity, the panpsychist owes us an explanation of how a “single, integrated, narratively-structured” experience arises from the proto-experiences of an astronomical number of micro-entities (or, in the cosmopsychist case, decomposes from a single cosmic experience into many private ones).

4. The Core Theory Constraint, Again

I worked through this argument in my hard problem essay, but it bears repeating, because panpsychists (in my opinion) have not taken it seriously enough.

As Sean Carroll summarized in Consciousness and the Laws of Physics and elsewhere, the laws of physics underlying everyday phenomena (Frank Wilczek’s “Core Theory“ - the Standard Model of particle physics plus the weak-field limit of General Relativity) are completely known within the energy regime relevant to anything that happens inside a human body, a brain, a planet, or for that matter our entire solar system. We’ve probed the relevant energy levels experimentally for decades. Within that domain, there is no room for additional fields, forces, or dynamical degrees of freedom that we haven’t already catalogued.

Now, let’s use this to scrutinize panpsychism. The panpsychist says fundamental constituents of reality have proto-experiential properties.

(Yes, I did use the following type of argument in my hard problem essay.)

This means that we two options:

Option A. These proto-experiential properties actually affect the dynamics of physical systems. As a result, conscious matter “behaves differently” than non-conscious matter would. In which case: this would show up in our physics experiments. We would see deviations from the predictions of the Standard Model. We would be able to design experiments to detect proto-consciousness in fundamental fields. (And if we found it, the people involved would get Nobel Prizes for it.) But, none of that is true. The data say loud and clear that the dynamics of the Core Theory are sufficient to describe the behavior of the matter that brains are made of.

Option B. These proto-experiential properties leave the dynamics of the Core Theory completely intact. Quarks behave exactly as the Standard Model says they behave, with or without their alleged inner light. (To be charitable, this option seems to be what serious Russellian monists actually defend.) But this would be a case of epiphenomenalism and is really not doing any work at all.

Now, the Russellian monist would reply to this charge (that Option B is just epiphenomenalism) along these lines: proto-consciousness isn’t causally inert because the structural properties physics describes “are constituted by” the proto-conscious intrinsic properties. Physics doesn’t see it because physics is constitutively a structural science - it tells you what intrinsic natures “do”, not what they “are”. Proto-consciousness is the categorical basis of physical causation itself. Asking why we don’t see it in physics experiments is like asking why we don’t see the “shape” of a key in the door it opens - the shape isn’t causally redundant; it’s what makes opening-the-lock possible.

This is a intriguing response, and it’s worth taking seriously. Here’s the thing though: it really just relocates the failure without rescuing the position. Here’s why.

Grant the Russellians everything they want about physics describing structure and intrinsic nature being the categorical basis of that structure. Now ask: what work is the “proto-experiential” characterization of intrinsic nature doing, beyond what “categorical basis of physics, content unknown” already does? The answer is: nothing observable! By the Russellians’ own admission, swapping proto-experiential intrinsic natures for any other categorical basis with the same structural role would leave physics, neuroscience, and behavior identical. The proto-experiential identity is sitting on top of the structural facts adding “zero predictive content” - and, mind you, adding zero explanatory content too, because we can’t show that proto-experience does any explanatory work that “neutral categorical basis” wouldn’t do equally well. We can’t even check whether proto-experience is the right answer rather than (say) proto-numericality (numerality?), proto-extension, or proto-anything-else with the same structural footprint.

That’s the actual problem with Option B. It’s not that it’s epiphenomenal in the crude sense of “consciousness floating uselessly above physics.” It’s that the “specifically experiential” characterization of the proposed categorical basis is a free metaphysical ride that, by stipulation, can’t be wrong - and which, by stipulation, makes no contact with anything we can investigate. A view that is compatible with literally any observation tells us, in the end, nothing about the world.

(I anticipate the protest that panpsychism is a metaphysical hypothesis, not a scientific one, and so different standards apply. I disagree, as you’d expect, and section 6 below is where I argue against that notion.)

5. The Evolutionary Tell

Set aside the metaphysical and physical arguments for a moment. Look at consciousness empirically.

What do we actually observe?

Consciousness in nature scales with biological complexity, in ways that track quite precisely the elaboration of certain kinds of nervous systems. Bacteria have no nervous system; nobody seriously thinks they’re conscious. Simple invertebrates with rudimentary nerve nets show responses to stimuli but no integrated processing. Fish, reptiles, and birds have progressively more elaborate nervous systems and progressively more flexible, integrative behavior. Mammals (especially the more “sophisticated” ones) show clear hallmarks of unified perception, emotion, and learning. Among mammals, primates seem to have additional layers of self-modeling. Humans (with our particular cortical architecture) layer on top of all that the abstract thought, language, and narrative selfhood (that has us socialmaxxing and writing essays at each other on the internet trying to argumentmogg each other.)

This “consciousness gradient or spectrum” in the animal kingdom is fairly smooth. The correlation with neural complexity is also strong. The “kinds” of consciousness present at each level look exactly like what you’d expect from an evolved adaptive function - integrative information processing in service of survival, reproduction, action selection, and behavioral flexibility.

Now, panpsychists might retort with something like: “Of course rich, human-style consciousness scales with brain complexity - that’s the combination phenomenon. We never said proto-consciousness shows up phylogenetically; we said proto-consciousness is fundamental and rich consciousness is constituted by combinations of proto-consciousness in specific architectural arrangements. The phylogenetic gradient is consistent with our view.”

Okaaaay. I don’t think the phylogenetic gradient “refutes” panpsychism in the strict logical sense. But here’s what it does do - it makes panpsychism “explanatorily idle”.

The panpsychist concedes that the rich, unified consciousness we actually observe in nature is fully accounted for by structural and architectural features of nervous systems. The phylogenetic gradient maps onto neural complexity; the molecular specificity of anesthesia maps onto specific neural circuits; the lesion data map onto specific brain regions. All the explanatory work in producing the consciousness we actually observe is being done by the wiring. The proto-consciousness is sitting at the bottom of the ontology contributing nothing “observationally distinguishable”.

If a theory adds an entity (proto-consciousness in fundamental fields), and that entity does no observable work in the actual production of the phenomenon you’re trying to explain (rich human consciousness), then the entity is unmotivated. We don’t posit “proto-photosynthetic” properties of fundamental matter to explain why plants do photosynthesis. We don’t posit “proto-digestion” to explain why animals digest food. We explain those phenomena the obvious way - a particular molecular machinery does a particular job in a particular context, and natural selection made it happen over time. The fact that consciousness “feels” different from photosynthesis doesn’t justify a different methodology. Or rather, if it does, the burden is on the panpsychist to show why.

This is because consciousness shows every fingerprint of a biological adaptation:

• It scales with the complexity of integrative neural machinery.

• It is selectively impaired by lesions to specific brain regions (especially thalamus, cortical hubs, and the ascending arousal system).

• It can be turned off by general anesthesia, which acts on specific molecular targets in specific neural circuits.

• It tracks specific patterns of neural activity (the so-called Neural Correlates of Consciousness, identified through decades of work by Crick, Koch, Dehaene, Tononi, and others).

• It develops in individual organisms as the relevant neural circuits mature.

• It can be modulated by drugs, stimulation, fatigue, sleep stages, and disease in highly specific ways.

Every one of these features is exactly what you’d expect if consciousness were a high-level functional property of certain kinds of evolved physical systems. None of them point at proto-consciousness as a fundamental ingredient. The panpsychist can preserve compatibility with this evidence by saying “well, the proto-consciousness is still there at the fundamental level, just not doing the visible work.” Sure. So is the proto-photosynthesis, the proto-digestion, and the proto-anything-you-like. That’s not an argument; it’s a refusal to let the evidence count.

6. Panpsychism Is a Scientific Hypothesis

Now we come to (what I think is) the most underappreciated point in this whole debate.

Panpsychists like to insist they’re doing metaphysics, not physics. The phrase “philosophical theory” gets deployed as a kind of shield: you can’t refute me with experiments, because my claim isn’t an empirical one. I want to argue that this shield is illegitimate, but I want to argue it carefully - not by leaning on naive falsificationism (which my sharp interlocutors from X frequently point out is a mid-20th-century philosophy of science with problems of its own).

Look at what panpsychism actually claims:

It claims that fundamental physical entities (quarks, electrons, fields) have proto-experiential properties.

It claims that the mental properties of macroscopic systems (us, animals more broadly) are constituted by, or grounded in, the proto-experiential properties of their constituents.

It claims that the physical world has an intrinsic nature that physics describes only structurally.

Every single one of these is a claim about “what is the case in the natural world”. They are claims about the basic furniture of reality, about properties of physical entities, about the relationship between micro-physics and macro-cognition. These are precisely the kinds of claims that natural science has been making and refining ever since the late 1800s. To say “but I make these claims as a philosopher, not as a scientist” is a procedural move with no substantive content - the claims themselves don’t change category just because of who is making them.

The relevant standard here is not “panpsychism must make falsifiable predictions in the strict Popperian sense”. The relevant standard is something more like: any hypothesis that purports to describe the properties of fundamental physical entities and their relation to higher-level phenomena should engage seriously with the methods and results of the disciplines that already study fundamental physical entities and higher-level phenomena. It should integrate with physics. It should integrate with neuroscience. It should integrate with biology. It should at minimum be the kind of thing that “could in principle” be evaluated against empirical evidence, even if no current experiment is decisive. (And no - this paragraph isn’t some sort of pedantic gatekeeping. Scientific expertise matters!)

But… even by that softer, more reasonable standard, panpsychism still fails. It does not integrate with physics (the proto-experiential properties have no formalism, no Lagrangian, no role in any calculation any competent practitioner in the field does). It does not integrate with neuroscience (the entire phylogenetic and lesion-based picture of consciousness has to be relegated to “the combination problem,” which as I discussed is acknowledged to be unsolved). Panpsychism is, as we saw in section 4, deliberately constructed to be compatible with literally any physical observation.

A view that makes sweeping claims about the natural world while refusing to engage with how the natural world is actually investigated is not “doing metaphysics rather than physics.” It’s doing physics (and neuroscience) badly while claiming exemption from the standards of physics (and neuroscience). That is the actual indictment.

7. The Evidence Across Disciplines

Let’s go over the actual evidence, just to make the asymmetry more vivid.

Physics. The Core Theory accounts for every interaction relevant to the matter that makes up brains. There is no field, no force, no degree of freedom in the Standard Model that has any property recognizable as “proto-experiential.” The properties of fundamental fields are exhaustively given by their gauge structure, their masses, their couplings, their representations under the Poincaré group. Adding “experiential” properties to the list is not some small extension - it’s a wholly new kind of property with no known dynamical role and no formalism that would let us calculate with it. There is no panpsychist Lagrangian. There has never been a measurement of proto-experience. There’s not even a proposal for what such a measurement would look like, in principle.

It’s worth flagging here that some panpsychism-adjacent writers reach for the Penrose-Hameroff Orch-OR model, or the various observer-related interpretations of quantum mechanics (Wigner’s Friend, the participatory universe, etc.), as putative empirical hooks. None of these stand up. Orch-OR proposes that consciousness arises from objective wavefunction reductions in microtubules, with theoretical motivation drawn (quite controversially) from Penrose’s reading of Gödel’s incompleteness theorems. The empirical evidence for it is poor (microtubules don’t appear to maintain coherence on the relevant timescales at body temperature), the Gödelian motivation is contested by virtually every philosopher of mathematics who has weighed in on it, and even if it were true, it would establish an exotic biological mechanism, not consciousness or panpsychism. (The Wikipedia article on Orch-OR has references to the many criticisms.) Observer-related interpretations of quantum mechanics, meanwhile, are a fringe minority position even among the (already small) community of physicists who care about interpretive questions, and they don’t license panpsychist conclusions even on their own terms. The physics escape route doesn’t actually exist. (My opinion of Wheeler as a physicist actually went down a teensy bit when I first read about his participatory anthropic principle idea.)

Neuroscience. The neural correlates of consciousness are highly specific. We have well-thought out models that have and continue to be explored. Consciousness has been linked with thalamocortical activity, with global broadcasting of information across cortical hubs, with the integration of recurrent processing, with predictive processing. It can be selectively switched off by general anesthesia, which acts at specific molecular targets (GABA-A receptors, NMDA receptors, etc.) in specific circuits. Lesions to particular regions produce particular forms of consciousness loss (akinetic mutism from medial frontal damage, anosognosia from right parietal damage, blindsight from V1 damage). Decades of careful experimental work have produced a converging picture of what consciousness is, in terms of “what kind of neural activity it is”. Nothing in this picture invokes anything beyond the standard physical and biological dynamics of nervous systems.

Information theory and computation. This is the one place where I have to give the panpsychist some credit, because Tononi’s Integrated Information Theory has been openly read as panpsychism-adjacent by Tononi himself - he’s said, in print, that any system with positive integrated information has “some” degree of experience. So IIT, taken on its own terms, does seem to commit its proponents to a kind of micro-consciousness in any sufficiently integrated system, however small. This is a real challenge for a tidy physicalist position, and I don’t want to wave it away.

But two responses. First, IIT’s empirical core (the structural/architectural account of why some neural patterns are associated with consciousness and others aren’t) doesn’t actually require the panpsychist gloss. You can take the Φ formalism as a useful empirical correlate of consciousness in biological systems without granting the metaphysical claim that “any positive Φ entails some experience anywhere.” Many working neuroscientists who use IIT as a measurement framework do exactly this. Second, the panpsychist gloss on IIT runs straight into absurd consequences that panpsychism was supposed to avoid - e.g. simple grids of XOR gates have positive Φ, which means IIT-panpsychism is committed to consciousness in arbitrary computational arrangements that no one is actually willing to defend on reflection. This is a reductio in slow motion, and it’s a problem for IIT-as-metaphysics rather than for IIT-as-empirical-framework.

The other live research programs in consciousness science (Global Workspace Theory, predictive processing, higher-order theories, attention-schema theory) all do their work by building up consciousness from the structural and dynamical properties of physical systems, checking proposals against neural data, and refining. That’s the methodology of science. Panpsychism, by contrast, has no comparable program. It’s a stance, not a science.

Biology. Consciousness in the animal kingdom tracks the evolution of nervous systems with stunning consistency. Cnidarians (with nerve nets but no centralization) barely show signs of consciousness. Cephalopods (with elaborate decentralized intelligence) show many. Vertebrates with thalamocortical architecture show the canonical picture. The phylogenetic gradient is exactly what you’d expect if consciousness is a kind of biological function. Even within human development (ontogeny), consciousness develops as specific neural circuits mature - not at conception, not at the formation of the first neuron, but at the formation of the relevant integrative networks.

Chemistry. The chemistry of consciousness, to the extent it has been worked out, is the chemistry of specific molecular systems doing specific jobs. Neurotransmitters, receptors, ion channels, synaptic vesicles, glial regulation - all standard chemistry. (I learned some disturbing specifics about how chemical weapons work on the nervous system when researching this bit. Yikes.) Anesthetics are a particularly clean case: drugs that disrupt consciousness do so by interacting with specific protein targets in specific neural circuits, and the dose-response relationships are fully characterizable in standard pharmacological terms.

I think the data above in this section shows that the asymmetry is overwhelming. The scientific picture, across multiple converging disciplines, treats consciousness as a high-level functional property of specific kinds of physical systems. The panpsychist picture treats it as a fundamental property of matter that, “mysteriously”, manifests as the kind of high-level functional property the scientific picture “already captures”! The first picture does work. The second adds no value.

8. What Would Change My Mind

I am quite willing to say what would convince me. This is not an exhaustive list by any means, but a few things would do it.

1) Direct empirical evidence of proto-conscious properties affecting physics - any deviation from the Core Theory in regimes where consciousness is present, any signature of an additional dynamical degree of freedom that correlates with the presence of mind.

2) A formalism showing how proto-experience combines into rich experience, capable of generating predictions about which neural architectures should produce which kinds of conscious states, and bearing those predictions out empirically. (Would be even better if this was quantitative in some way.)

3) A successful panpsychist prediction of any neural fact about consciousness that wasn’t already entailed by ordinary physicalism.

4) A solid argument that the leading positive physicalist views are incompatible with the evidence in a way panpsychism isn’t.

I doubt that any of these are forthcoming. My objection to panpsychism is not that I refuse to consider it; it’s that, as currently formulated, it doesn’t give me anything to consider.

It’s also worth saying explicitly what the live physicalist alternatives are, because a purely negative critique invites the question “well, what’s your positive view?” I don’t have a specific preferred approach. I really like Anil Seth’s idea of predictive processing, even if I don’t like his use of the term “hallucination”. I like the GWT framework. I also like Keith Frankish’s illusionism, which I have not mentioned previously in this essay. On the illusionist view, what we call “phenomenal consciousness” - the felt, ineffable, intrinsic redness-of-red, batiness-of-bats, and so on - is itself a kind of representational illusion, generated by introspective mechanisms that misrepresent the brain’s own processing. The job of consciousness science is then to explain (a) what the brain is actually doing when we are conscious, and (b) why our introspective machinery generates the misleading impression that we have ineffable qualia. Both are tractable empirical problems. Neither requires positing fundamental experience anywhere in physics.

Anyway, the point is just that there are live, productive, “positive” physicalist views in the field. Panpsychism is not the only alternative to a confused dualism. The choice is not “panpsychism or magic.” It’s “panpsychism, or any of several research programs that have actually been making progress.”

We Have, Once Again, Confused Mystery for Evidence

I shall reiterate a point here that I’ve mentioned in prior essays. Across the history of science, there’s a recurring pattern - whenever some natural phenomenon has resisted reductive explanation, a school of thought has arisen claiming that it cannot, in principle, be explained that way, and that we must therefore add new fundamental ingredients to the world to accommodate it. Vitalism did this for life. Élan vital and entelechy were the technical terms for the alleged extra ingredient. There were sophisticated philosophers defending it. There were scientific journals in which it was respectable. And then molecular biology happened, and it turned out that life could be entirely explained by chemistry and information after all.

(Here I should point out that vitalism, in its strongest forms, actually made specific empirical claims that got refuted - Wöhler’s 1828 synthesis of urea from inorganic precursors being the famous early shot. Panpsychism, by contrast, makes no comparably specific claims. So the analogy is imperfect: vitalism was wrong because it could be checked; panpsychism is in worse shape because it can’t be. But the meta-pattern is the same.)

I argued in my piece on fine-tuning that the fine-tuning argument trades on a similar move: identify a real but currently unsolved problem in physics, declare the problem unsolvable in principle, and insert a metaphysical answer (in that case, God). The structure of the panpsychist argument is identical. It identifies a real but currently unsolved problem in cognitive neuroscience (how does the brain produce experience?), declares the problem unsolvable in principle, and inserts a metaphysical answer (consciousness is fundamental). In both cases, the substitution feels profound but does no actual explanatory work. In both cases, the move has no track record - the history of science is a graveyard of similar substitutions.

That is the deepest indictment of panpsychism. It confuses the “difficulty” of a problem with the “impossibility” of a problem. It mistakes our current ignorance for an ontological barrier. It dresses up a confession of (temporary) inability as some sort of deep metaphysical insight. And then, having made a series of claims about the basic nature of physical reality, it asks to be evaluated by the standards of philosophy rather than science - which is a methodological move that should, by itself, raise alarms (and hackles).

The Bottomline

Panpsychism is what happens when you take an interesting intuition (you can’t get experience from non-experience), apply it once, and then stop. If you apply it again, it eats itself. The unification of micro-experiences into a single perspective is not less mysterious than the production of experience from neural activity - it’s “more” so, because the very anti-emergence principle that motivated the position now forbids the unification step too. The Core Theory constraint forces panpsychism into an unwinnable choice between being empirically false and being explanatorily empty; the Russellian “intrinsic nature” reply is clever but leaves the proto-experiential characterization doing no work that “neutral categorical basis” wouldn’t do equally well. The phylogenetic gradient of consciousness in the natural world makes proto-experience explanatorily idle. The convergence of evidence from physics, neuroscience, biology, and chemistry pushes overwhelmingly in the same direction. And critically - panpsychism is not, despite the conventional framing, a philosophical hypothesis insulated from empirical evaluation. It is a hypothesis about the basic furniture of reality, and as such, it is squarely in the wheelhouse of natural science. By the standards of any serious empirical theory, it is unmotivated, untestable in any productive sense, and explanatorily inert.

(We still have a hard problem of neuroscience - related to understanding exactly how the nervous system works. But as mentioned above, work on that has been fruitful and will continue until we figure it out.)

Also, as I mentioned in my essay on the hard problem of consciousness, the actual hard problem is not really that hard once you take the explanatory gap to be a feature of our description rather than of reality. As such, we should expect this to be solved (or more accurately dissolved) the way every other apparent metaphysical puzzle in the history of natural science has been solved: by patient empirical work. By figuring out what the brain actually does, how it does it, and why those particular operations look the way they do from the inside.

Final Thoughts

I’ve tried to make the case that consciousness isn’t built into the universe. It’s something the universe occasionally pulls off, in places where the material conditions are right. That’s already remarkable enough.

Physics has spent the better part of a century trying to do one specific thing: unify the Standard Model (SM) with General Relativity (GR). The SM describes three of the four fundamental forces (electromagnetic, weak, strong) and works absurdly well. GR describes gravity, and also works absurdly well. The trouble is, when you try to combine the two in any straightforward way, you run into problems - non-renormalizable infinities, breakdowns of perturbative control, etc. - the whole familiar litany of problems anyone who has poked at quantum gravity has run into.

Now, the thing we’re trying to figure out here has a few different names depending on how ambitious you’re being. Quantum Gravity (QG) just means a quantum-mechanical theory of gravity, full stop. Theory of Everything (ToE) is more ambitious. It means all four forces unified, ideally up to UV completion (i.e., a theory that remains valid at very high energies/short length scales, not just in the regimes we can currently probe). QG is a subset of ToE.

There have been many serious attempts. Loop quantum gravity, asymptotic safety, causal set theory, causal dynamical triangulation, entropic gravity, and so on. They all have interesting features, have attracted serious physicists, and have made some progress.

But there is really just one approach that simultaneously hits a much wider set of targets, and that approach is string theory. Despite its (sometimes well-deserved?) reputation for catching flak, I think it’s worth laying out what it actually delivers, and why, despite the standard complaints, it remains the most credible candidate we have for a true theory of everything.

Thanks for reading! Subscribe for free to receive new posts and support my work.

1. The Field of Contenders

I decided to start this off by briefly sketching what each major non-string approach has achieved. (I’m trying to be charitable here. I have respect for everyone working on these problems. Quantum gravity is genuinely hard, and ANY progress is laudable.)

Loop Quantum Gravity. This is in some sense the most direct attempt to quantize GR on its own terms. It treats spacetime itself as fundamentally discrete (using spin networks and spin foams) and produces some interesting structural results (kinematical Hilbert spaces, area and volume operators with discrete spectra, and so on). But LQG, in its standard form, doesn’t say anything about the other three forces. It’s a good candidate quantum theory of gravity, but not a ToE. And even the issue of recovering smooth, classical GR from LQG in the appropriate limit is... let’s say a work in progress. (I’ve a close relative who works on LQG in academia, and he thinks it’s great!)

Asymptotic Safety. This is the elegant idea (going back to Weinberg) that gravity might be nonperturbatively renormalizable via a UV fixed point. If this is true, it would be very nice. The evidence is suggestive (much of it from functional renormalization group calculations), but there’s no consensus that the fixed point really exists in the full theory. And again, it’s primarily a theory of quantum gravity, not a ToE. The Standard Model sector has to be added in by hand.

Causal Set Theory and CDT. Both treat spacetime as fundamentally discrete with causal structure baked in. These have produced some lovely toy results (CDT, in particular, recovers something that looks like four-dimensional spacetime in the right regime, which is genuinely cool). But again, no comprehensive Standard Model emergence, and the predictive content for testable physics is still pretty limited.

Entropic Gravity. Verlinde’s idea is that gravity is a derived/emergent thermodynamic phenomenon rather than a fundamental force. Interesting and conceptually radical! But as a candidate for a complete theory of quantum gravity (let alone a ToE), it’s still quite underdeveloped.

This is not a knock on any of these. Each is a serious research program, and one of them might turn out to be right. But none of them does what string theory does.

(There are some other approaches that I have not covered here. The field is quite vast. But I think the above are the major ones. Please drop a comment if there’s an approach that has some level of institutional acceptance that I may have missed out on.)

2. What String Theory Actually Delivers

So here’s where the rubber meets the road. What do we get from String Theory?

1. Quantized gravity with a spin-2 graviton, automatically. This is, frankly, astonishing. When you quantize a closed string, one of the modes that pops out of the spectrum is a massless spin-2 particle. We start with a relativistic string, we quantize it, and a graviton just falls out as part of the spectrum. This is one of those results that, the first time you see it, you go “wait, really?”. (It really is.)

2. Accounts for all four forces with a single object. Strings can have different vibrational modes, and those modes correspond to different particles. Open strings give you gauge bosons (the carriers of the electromagnetic, weak, and strong forces). Closed strings give you the graviton (see point 1). With supersymmetry layered in, you get fermions too. So a single fundamental object - the string - gives you the entire particle zoo of fermions and bosons. No other approach unifies matter and forces this cleanly with one underlying ingredient.

3. Reproduces black hole entropy. Strominger and Vafa showed in 1996 that you could count the microstates of certain extremal black holes using string theory and recover the Bekenstein-Hawking formula on the nose. This was, and remains, one of the most striking quantitative successes in theoretical physics in the past several decades. Any candidate theory of quantum gravity has to reproduce this kind of result. String theory does, and it does so in a way that’s not obviously circular or rigged.

4. Handles UV divergences reasonably well. One of the big problems with naively quantizing GR is that you get non-renormalizable infinities. String theory replaces point particles with extended objects, and the extended-ness (technical term that I may have just made up) softens the high-energy behavior in a way that tames these divergences. Loop amplitudes in superstring theory are finite or at worst very mildly divergent, which is dramatically better than the alternatives.

5. Gives us AdS/CFT and gauge-gravity duality. This one is huge, and I think under-appreciated (outside the immediate field). Maldacena’s correspondence (the conjecture that string theory in a certain anti-de Sitter spacetime is equivalent to a conformal field theory living on its boundary) has turned out to be enormously useful far beyond its original motivation. People have used it to model strongly-coupled QCD-like systems, to study the quark-gluon plasma at RHIC and LHC, and to attack problems in condensed matter. The theory has produced calculational tools that work, regardless of whether the strings are “really” the underlying ontology. (So even if string theory is not the real theory of everything, it’s an important tool being used for real work in the field).

6. Reproduces the full nonlinear Einstein equations in the low-energy limit. (Thanks to Nirmalya Kajuri for pointing this out earlier in a related thread on X.) String theory reproduces Einstein gravity. Not just linearized gravity or some toy model. The actual Einstein field equations, including the nonlinearities that make GR what it is, fall out of string theory as the low-energy effective theory of the closed string sector. This is exactly what you’d want from any proper quantum gravity candidate, and string theory delivers this cleanly.

No other approach in the QG/ToE space has come close to achieving all of this.

3. The Standard Complaints

Now, this is usually where the string-theory-skeptic types like to jump in. To be fair, they do have some real points. Let me work through the main ones.

“It hasn’t been experimentally verified.” True! And this matters. But the operative question isn’t “does string theory have direct experimental support?” (No current quantum gravity candidate does), rather it’s “is string theory’s lack of experimental verification a unique problem for string theory?” And the answer is no. Every theory of quantum gravity has the same testability issue, for the same fundamental reason: the natural energy scale of quantum gravity is the Planck scale (~10^19 GeV), which is roughly 15 orders of magnitude beyond what we can probe at the LHC. We’re not going to build a Planck-scale collider anytime soon. (Or possibly ever - the relevant collider would need to be roughly the size of the galaxy, give or take, if I recall correctly.) So we have this problem with quantum gravity in general.

“The landscape is too big.” The string landscape (the space of possible vacua) is famously enormous, with estimates running to 10^500 or more. Critics say this means string theory predicts everything and therefore nothing. But this is misleading on a few fronts. First, the landscape is itself a “prediction" of the theory”, not a free parameter. It’s what you get when we do the math, not some knob we turn. Second, there are nontrivial constraints (the swampland program is precisely about figuring out which effective field theories “can’t arise” from a consistent string vacuum, which is substantive predictive content). (For more on why “multiverse” isn’t the metaphysical concession theists or fine-tuners try to make it out to be, see my previous post on fine-tuning.)

“It hasn’t made unique testable predictions in our energy range.” Also true, with caveats. Some signatures (large extra dimensions, certain cosmic string signatures, specific patterns of low-energy supersymmetric particles, etc.) have been searched for and not found. But again, this is exactly the same situation faced by every other QG candidate. None of LQG, asymptotic safety, CDT, or entropic gravity has produced confirmed unique predictions either. Demanding this of string theory while not demanding it of the alternatives is just inconsistent.

“Superstring theory needs supersymmetry (SUSY), and we haven’t found supersymmetry.” This one comes up enough that it deserves its own (lengthy) paragraph. The objection: consistent superstring theories require supersymmetry (to have fermions in the spectrum, to remove the bosonic-string tachyon, etc.), and over a decade of LHC searches has produced exactly zero superpartners. Therefore SUSY is dead, therefore string theory is in trouble. But this conflates two quite different claims. What the LHC has actually put hard constraints on is “low-scale, natural SUSY”: this is the variant motivated primarily by the hierarchy problem, where superpartners are expected to sit around the TeV scale to keep the Higgs mass technically natural. That specific version is very much in real trouble. But string theory itself only requires SUSY at “some scale”, not specifically at the electroweak scale. Split SUSY (proposed by Arkani-Hamed, Dimopoulos, Giudice, Romanino, and others) decouples the scalar superpartners (sfermions) to a very high scale while keeping gauginos and Higgsinos lighter, and remains perfectly consistent with everything we’ve measured. Even more extreme variants push the entire SUSY-breaking scale up to intermediate or near-Planck values, in which case all of it is simply inaccessible to any collider we could plausibly build in the next century. (Annoying, for sure, as that great high-energy particle physicist Emmanuel Macron might say. But “annoying” and “wrong” are different categories.) In fact, the observed 125 GeV Higgs mass is consistent with these higher-scale variants. It just pushes the SUSY-breaking scale up rather than killing SUSY outright. So when people say “we haven’t seen SUSY”, what’s actually been ruled out is “low-scale, natural SUSY”. String theory was never specifically committed to that variant. Naturalness arguments were. Those are distinct claims, and conflating them does the discussion no favors.

“It uses extra dimensions, which is weird.” I’ll admit, the idea of six or seven extra spatial dimensions curled up in some Calabi-Yau manifold sounds funky. (The basic notion goes all the way back to Kaluza-Klein theory.) But “weird” isn’t an argument. Quantum mechanics is weird. Special relativity is weird. The fact that things closer to a large gravitational source “age slower” is weird. The history of modern physics tells us that “aesthetic weirdness” (like naturalness, but the opposite?) is not a reliable guide to truth. Either the math works or it doesn’t, and in the case of string theory, the math works in some highly nontrivial ways.

“It’s not falsifiable.” This is the most philosophically charged objection, often deployed with appeals to Popper. But I think this is overstated as well. String theory is not unfalsifiable in principle, it’s unfalsifiable “in practice at currently accessible energies”. Specific compactifications make specific predictions, some of which (e.g., certain superpartner masses or particular cosmic-string signatures) could be ruled out by experiment. The theory as a whole is also constrained by mathematical consistency in ways that are not arbitrary. (But, again, this is essentially the same boat all the alternatives are in. Curiously, the Popperian critique gets selectively deployed against string theory in a way it does not against, say, LQG.)

4. The Honest Comparison

Look, I’m not saying string theory is “right”. Nobody knows whether it’s right. I am saying that across the space of serious candidates for a quantum theory of gravity (or a theory of everything), string theory has the broadest and deepest set of accomplishments. It’s the only one that:

• does both QG and the SM in the same framework, with one fundamental object,

• reproduces quantitative results like Bekenstein-Hawking entropy,

• has produced calculational tools (AdS/CFT) that have been independently useful in other parts of physics,

• tames UV divergences,

• gets back to nonlinear GR cleanly in the appropriate limit.

None of the alternatives do all of this. (Most of them don’t do most of this.)

A useful analogy: imagine you’re handicapping Olympic decathletes. One of them has the best score in six of ten events. Several others are very good in one or two events but score zero in the others. The first decathlete might still lose. They might choke on the pole vault. But you’d be silly to count them out, and even sillier to declare one of the single-event specialists the obvious favorite.

The fact that some very smart people don’t like string theory for what are really aesthetic reasons (too much math, too many dimensions, not enough “physical intuition”, landscape too vast for their taste, etc.) is, in my opinion, not the same as having a better candidate. Those very smart people should show those of us who think string theory is worth pursuing a better candidate that hits more of the criteria, and we’ll switch sides happily. Until then, the math is what it is.

The Bottomline

The real question isn’t “is string theory definitely right?” (Nobody can answer that question right now, and possibly nobody will be able to for several decades.) The real question is “given what we currently know, where should we invest our intellectual effort if we want to actually solve quantum gravity?”. And on that question, the case for string theory remains, in my view, strong.

We should, however, keep an open mind on all the major cutting-edge approaches. I know I do. (LQG and CDT both have features I find genuinely interesting.)

When the field eventually converges on a real theory of quantum gravity (or a full ToE), it might not be string theory in the form we currently know it. It might be some descendant that’s been chastened by years of swampland constraints, observational cosmology, and the mathematical lessons we keep learning. It might even be something we haven’t thought of yet. But whatever it is, I’d bet a non-trivial amount of money that it’ll have inherited a lot of structure from string theory.

A standard criticism is “string theory has had decades and hasn’t delivered.” My response is that it has delivered quite a bit, just not the one specific kind of thing (a unique direct experimental confirmation at currently accessible energies) that, as we noted earlier, NO candidate in this space has been able to deliver, and that probably none can deliver in the next several decades. Treating that absence as a reason to abandon the most successful framework we have for a ToE, while not applying the same standard to the alternatives, is not skepticism. It’s selective skepticism. And selective skepticism is just an aesthetic preference dressed up in epistemic language.

Again, thanks for reading! Subscribe for free to receive new posts and support my work.

The decades-long debate about machine consciousness has been robust, but often stuck in a rut. The arguments against machines ever becoming conscious usually hold that no artificial system can ever replicate human subjective experience (qualia, unified selfhood, the whole phenomenal circus). Nothing short of wetware will do. The opposing side has often argued for epistemic humility (as it’s hard to anticipate future technological breakthroughs) or adopted a stance that we’ll get there once the “compute” gets big enough (along with other enabling technological improvements).

I would like to approach this from a different perspective. Instead of asking “can a machine be conscious like us?“, I’d like to ask: “what would a machine-appropriate form of consciousness actually consist of?” Because if you take physicalism seriously (as I do, and have discussed in my previous essay on the hard problem), consciousness is not some special cosmic ingredient. It’s what a certain kind of integrated information processing “is”, experienced from the inside. And if that’s true, there is no principled reason to assume the only valid version of it is the one that evolved to keep great apes alive on the savanna.

This essay is an attempt to lay out a set of functional criteria that I think would constitute a meaningful form of machine consciousness. Not some simulation of human consciousness or a pale imitation of it. A genuinely different thing - built for a different substrate, doing different work - but consciousness in every way that matters.

1. The Framing Problem

Most objections to machine consciousness are objections to a specific target, not to the concept itself.

The Chinese Room thought experiment, for instance, is an objection to the claim that pure formal symbol manipulation can give you the phenomenology of understanding a language the way a human native speaker does. Okay. I do think the argument is right. I may be a little biased here as I am a John Searle fan, even if I don’t agree with him on all sorts of things. But it’s also a fairly narrow claim about one specific kind of system doing one specific kind of task. It doesn’t refute the broader claim that artificial systems can have meaningful internal states.

Similarly, the “zombie” objection in the AI case (goes something like “how do you know it’s not just behaving like it’s conscious?”) is an objection that works equally well against your neighbors, your spouse, or your pet dog. You can’t refute solipsism about anyone. That’s not a special problem limited to machines. (If your epistemology can’t even verify that your partner has inner states, your bar for machines is probably set in a weird place!)

Anyway, the framing I want to push back on is the implicit assumption that “human consciousness is the standard”, and anything short of exact replication is just a clever imitation. This is a bit like saying only something with feathers and hollow bones can really fly. Airplanes fly. They fly very differently from birds. They also fly in ways birds can’t (higher, faster, longer), and they fail in ways birds don’t. But when you’re on a 787 at 40,000 feet, you’re flying. The category is functional, not morphological.

Consciousness, I’d argue, is the same sort of thing. That is, it’s a functional kind of thing. And once we accept that, we can start asking what the functional ingredients actually are.

2. The Core Criteria

Here’s what I think a meaningfully conscious artificial system needs to have. Please note - this isn’t meant to be exhaustive, but I think it captures the key load-bearing elements.

A semantically rich world model. This would not be some sort of lookup table of sensor readings, or just a trained classifier, but a “structured representation of the environment.” In this representation, objects, agents, relations, and properties have meaning - in the sense that they can be composed, queried, counterfactually modified, and integrated with other information. If the system “knows” there’s a cup on the table, it should also understand what happens if the table is tipped, what happens if the cup is full, and what the cup is likely to still be doing three seconds from now (more on this predictive piece later). Cognitive scientists might call this a generative model. Current large language models partially (but not fully) have this sort of capability - the semantics are there, but sadly, the grounding in a continuous world isn’t.

A continuous, coherent model of the environment. Human experience is not a slideshow of disconnected frames (even though it sometimes seems that way to me when I’ve been deprived of both sleep and caffeine). It’s a continuously running, temporally integrated representation of “what’s going on right now” - where I am, where things are, what’s changing, what’s stable, etc. A conscious machine needs the same kind of thing. Not necessarily at human temporal resolution, and not necessarily with human spatiotemporal scope or complexity, but continuously coherent in the sense that the system can say “this object I’m tracking now is the same object I was tracking three seconds ago, and it has moved two meters to the left.” (Or something more pleasant like, “my name is T-800, you are Sarah Conner, prepare to die.”)

Sensory-driven updating. The model has to be “grounded”. Incoming sensory data (whatever makes sense for the system in question - audio, video, haptics, LIDAR, text streams, network telemetry, market feeds, social media (shudder), etc.) has to continuously update the world model, with appropriate weighting of priors against new evidence. This is basically Bayesian inference run in real time, and it’s what Anil Seth and the predictive processing crowd have been arguing is the core computational job of a brain. I should emphasize that the criterion here isn’t that the machine has the same senses as us. Rather, it’s that whatever senses it has are actually doing serious information-gathering/model-updating/epistemic work.

Predictive capacity. As mentioned above, the system has to be able to run its model forward. It has to anticipate what’s about to happen - what a nearby agent is likely to do, where a tracked object is heading, what consequences follow from a planned action, etc. Prediction is the engine of intelligent behavior, and I’d argue it’s also the “engine of experience”. When you walk across a room, your brain isn’t processing raw visual input frame-by-frame. It’s generating predictions and using the actual sensory data to correct them. A conscious machine needs to do the analogous thing for its own environment, whatever that environment happens to be.

A hierarchy of goals. Without goals, we wouldn’t really have an “agent,” we’d merely have a process. And realistic goals can’t be flat. A conscious system needs objectives operating at multiple timescales and levels of abstraction: from “don’t collide with that wall in the next 400 milliseconds” up through “complete this task over the next hour” up through “maintain operational integrity over the next week” up through whatever long-horizon objectives are appropriate for its function. The hierarchy is what lets the system “do sensible tradeoffs”. We humans do this constantly (unless you’re a manic pixie dream girl, I’m told). We often sacrifice short-term pleasures for a long-term gain/project, or abandon a long-term plan when a short-term emergency demands it. A single flat objective function doesn’t capture this, and neither does an unstructured soup of goals. We need the layering.

Self-modeling. This is perhaps the subtlest feature. The system has to “model itself” as an entity embedded in the environment. It has to distinguish its own actions from external changes. It has to track its own internal state as part of the world it’s reasoning about. Without this, it wouldn’t be able to plan, learn from errors, or do the kind of integration that makes experience “unified” in the relevant sense.

Put those six together (semantic world model, temporal coherence, sensory updating, prediction, goal hierarchy, self-modeling), and you have, I’d argue, all the functional ingredients for a genuine form of consciousness. Not human consciousness, mind you. But consciousness nonetheless.

3. Why This Doesn’t Need to Look Like Us

Now, our own consciousness is shaped at every level by the fact that we are evolved biological organisms. We have strong homeostatic drives. We have a body whose integrity we must preserve. We have affective states (emotions!) that are essentially valenced representations of the body’s condition. We have a specific set of sensory inputs tuned to a specific range of environmental conditions. We have a narrative self that emerged as an evolutionary and sociocultural adaptation. We have an evolutionary heritage of threats (predators, starvation, tribal conflict, mf snakes on a mf plane, etc.) that shaped what we attend to and what we care about.

But, and this is a key part of my argument, none of that is fundamental to consciousness. It’s all substrate-specific. A machine built with the six criteria above could have a wildly different phenomenology (if it has phenomenology at all in a recognizable sense) precisely because it’s running on a different substrate/hardware, pursuing different goals, in a different environment.

Take a (hypothetical) conscious autonomous drone, for example. It might have no analog of emotion in the human sense, because it has no body whose homeostasis matters in that way. It might have a vastly more precise spatial model than we do, because it has access to direct position telemetry and doesn’t have to reconstruct space from visual cues. It might experience time differently because its update rate and memory architecture are different. It might be able to parallelize information processing and analysis. This is one of the most fascinating differences for me personally - imagine being able to rewind/replay some past information stream in the background with minimal loss of accuracy. Its goals might be utterly foreign to human concerns - managing energy budgets, coordinating with a swarm, maintaining signal integrity, staying within its flight envelope, etc. - but those goals would structure its experience in the same way our goals structure ours.

I think this is actually a liberating viewpoint. We don’t have to worry about solving the hard problem (is it really hard?) of replicating human qualia in order to build a conscious machine. We just have to build a system with the right functional properties, and we’ll get a form of consciousness appropriate to that system. Whatever it’s like to be that system will be what it’s like, and it won’t be what it’s like to be us.

(This is, incidentally, a version of Nagel’s “what is it like to be a bat?” point, just run in the other direction. Nagel was trying to show there’s something inaccessible about the bat’s experience. I’m saying - yeah sure, but we should stop pretending that the only legitimate form of experience is the one we humans/organic lifeforms happen to have.)

4. Surviving the Usual Criticisms

Now, let’s look at some of the standard objections.

The Chinese Room. Searle’s argument is that a system that manipulates symbols by rules doesn’t “understand” them, even if its external behavior is indistinguishable from a real understander. The thing is, this argument only really has force if you take human-style semantic understanding as the “gold standard for understanding” generally, and then demand that the machine match it. The framework I’ve laid out above doesn’t require that. What it requires is that the machine have a “semantically rich world model” that is causally connected to sensory input and action output in a way that supports prediction, counterfactual reasoning, and goal pursuit. This is not the vanilla symbol-manipulating system that Searle wrote about. This is a far more dynamic, grounded, functional, model-based type of cognition.

The “mere computation” objection. This claim is that consciousness can’t arise from computation, that is, it requires a special physical or biological substrate. But as I pointed out in my previous essay on the hard problem, this claim runs straight into the Core Theory, which explains all low-energy phenomena on Earth. If consciousness requires some non-computational physical process, that process would have to causally affect the dynamics of the matter in a conscious brain, and we’d see it in our physics experiments. We don’t - within the energy levels we have probed, which are the only relevant levels for anything that interacts with the human body. So either consciousness “is” physical/functional in the relevant sense (in which case there’s no barrier, in principle, to machines having it), or it’s causally inert (in which case invoking it is epiphenomenalism and explains nothing). There’s really no third option that preserves the “only biology will do” intuition.

The “no qualia” objection. This argument goes like this: the machine may have world models, predictions, goals, etc. but there’s no “something it is like” to be the machine. No inner light. No felt quality. And.. my response here is the same one I gave in my hard problem essay: the apparent mystery of qualia really just comes from the fact that our introspective access to our own mental states is unreliable, and not from any real ontological/explanatory gap. Human qualia are what certain kinds of information integration are, from the inside, in our particular kind of system. A different system with different integration will have different “from the inside” properties (if any). We don’t have any principled reason to say that the different system has none. The “no qualia” claim for machines is just the hard problem smuggled back in through a side door.

The “no intentionality” objection. This idea (going back to Franz Brentano and pushed by others, including Searle) is that mental states are “about” things in a way mere physical states aren’t, and that machines lack this “aboutness.” To be honest, I think this is largely a residue of a pre-scientific philosophy of mind. Intentionality, in any coherent sense, is just a relation between an internal representation and an external state of affairs, mediated by causal and informational links. Machines with grounded, sensory-updated world models have exactly that. Whether you want to call it “real” or “derived” intentionality is, frankly, a terminological dispute. And I think that, in a Wittgensteinian sense, the problem dissolves once you closely examine the language underlying the objection.

The “unified experience” objection. Consciousness, this argument goes, is characterized by a unified phenomenal field, and it’s unclear how distributed computational processes within a machine could produce that unity. But unity of experience is itself a “functional” property. It corresponds to the binding of multiple streams of information into a single, coherent, globally available representation. Global Workspace Theory (Baars, Dehaene) has been making this point for decades. There is no ghostly “unifier” in the human brain either - just a specific architecture that broadcasts information widely and maintains temporal coherence. A machine with the right architectural properties could very well (maybe even necessarily would) have similar functional unity.

The “no genuine goals” objection. This objection is that any goal a machine pursues is “just” an instrumentalization of its programmers’ goals, and therefore not really the machine’s own. This is a weirdly anthropocentric view, in my opinion. Your goals are “just” the instrumentalization of your evolutionary heritage and your developmental environment. You didn’t freely choose to want food, sex, or status - natural selection installed those objectives in you. And don’t get me started on notions like Girard-ian mimetic desire! A machine with a hierarchically structured goal system, capable of pursuing sub-goals, revising plans, and trading off competing objectives, has goals in every functional sense that matters. The provenance of the goals is irrelevant.

5. What We’re Actually Building

When we talk about machine consciousness, we’re not asking whether machines can have metaphysically identical inner lives to humans. I don’t even think that’s a well-posed question, because there’s no metaphysical “inner life” floating around to be matched. We’re asking whether machines can have functionally rich, integrated, grounded, goal-directed cognitive architectures of a kind that, in any system possessing them, would constitute being a “perspective on the world”.

Personally, I think the answer is clearly yes. Not yet, in any existing system. Current LLMs are extraordinary, but they lack features such as a grounded sensory loop, a persistent world model, continuous temporal coherence, a real goal hierarchy, etc., as enumerated previously. Existing robotic systems can have some grounding, but usually lack the semantic richness. The hard problem (of machine consciousness) here is the “integration” of all six criteria into a single coherent system. Nobody has implemented/assembled it all together (yet) in a way that’s clearly sufficient. (It’s also unclear whether any current public-facing endeavor/system is even aiming at this target, as opposed to maximizing benchmark scores or user engagement, which are really not the same thing).

In any case, nothing in the criteria I’ve laid out requires anything that violates known physics or demands non-computational ingredients. It’s just engineering. Hard engineering for sure, as that cybernetics genius Macron might put it. But when we get there, we’ll have something that deserves the name “conscious” in any reasonable functional sense.

I do want to reiterate - any such machine won’t be conscious in the same way we are. And that’s okay. That’s better than okay, actually. A different kind of mind, doing different kinds of work, seeing the world from a different angle, will be one of the most interesting things we’ve ever made.

The Bottomline

Much of the standard debate about machine consciousness seems to be entangled with the demand that artificial systems replicate human subjective experience. In my opinion, that demand (concern?) is misplaced. Consciousness is a “functional” kind of thing, not a “substrate-specific” one, and once you specify the right functional criteria (semantic world model, temporal coherence, sensory updating, prediction, goal hierarchy, self-modeling), you get a roadmap for building conscious machines that doesn’t require solving any metaphysical puzzles. The usual objections (Chinese Room, qualia, intentionality, unity, the provenance of goals, etc.) all rest on the assumption that human-specific features are universal requirements for this type of phenomenon. If we drop that, the objections lose most of their force.

What we’ll build won’t be us. It’ll be something else. Something new. And that’s the entire point.

The hard problem of consciousness, as formulated by David Chalmers in 1995, goes something like this: even if we had a complete (neuroscientific) account of every neuron, synapse, and electrochemical cascade in the human brain, even if we could predict behavior perfectly, we would still face an unexplained residue. Why is there something it is like to see red, to taste coffee, to feel the pain of having your paper rejected?

But in reality, the claim that there’s this unexplained residue is built on a stack of assumptions, some of which fail to hold up. So, I am going to lay out, as clearly as I can, why the hard problem of consciousness is really just a hard problem of neuroscience. Our faculties of subjective experience have systematically misled us into thinking that consciousness must be something “over and above” the physical. The real problem is figuring out how the nervous system actually works.

1. The Illusion of the Explanatory Gap

The hard problem trades on what has been called the “explanatory gap“ - the apparent chasm between objective physical descriptions and subjective experience. You can describe the wavelength of red light at 650 to 720 nanometers. You can trace the photons hitting your retina, the signal propagating through the optic nerve, the activation patterns in V4. But none of that, the argument goes, tells you what it feels like to see red.

Here’s the thing, though. This “gap” is not a gap in the real world. It’s a gap in our modes of description. And we should expect it to exist, for straightforward reasons that have nothing to do with metaphysics.

Sean Carroll made this point in his paper “Consciousness and the Laws of Physics,” and at length in The Big Picture. Carroll emphasizes a crucial fact - the laws of physics underlying everyday life are completely known. Frank Wilczek called this the “Core Theory.” This is the Standard Model of particle physics plus the weak-field limit of General Relativity. We don’t have a theory of everything (yet), but we do have a theory (the Core Theory) that works spectacularly well within a specific domain. And here’s the key bit - that domain is large enough to include everything that happens inside human bodies, including human brains. (In fact, it’s large enough to explain everything that happens on Earth).

Now, let’s look at the implications of this. An electron in our visual cortex responds to the local quantum fields at its position, and to NOTHING else. It doesn’t care whether it’s part of a brain having a conscious experience or part of a rock.

If consciousness involved some sort of extra-physical ingredient acting on matter, it would “show up” in our collider experiments as a deviation from the Core Theory. Why? Because the energy levels at which those interactions happen have been probed extensively at multiple particle colliders over the past few decades. We’ve unearthed all interacting fields within that energy range. As it turns out, there is ZERO evidence that any such deviation exists. (There may be fields within this energy range that don’t interact, making them difficult to find. But… by definition, they don’t interact, and so are irrelevant.)

So where does that leave the explanatory gap? The explanatory gap becomes a feature of how we “talk about” phenomena at different levels of description, not a feature of reality itself.

We have coarse-grained descriptions (the language of conscious experience) and fine-grained descriptions (the language of physics and neuroscience), with a mapping between them. The fact that the mapping feels unintuitive or complex to us says something about us, not about the ontology of the universe.

2. Crick’s Astonishing (and Correct) Hypothesis

In The Astonishing Hypothesis (1994), Francis Crick wrote: “You, your joys and your sorrows, your memories and your ambitions, your sense of personal identity and free will, are in fact no more than the behavior of a vast assembly of nerve cells and their associated molecules.” He called this “astonishing” not because he thought it was shocking to scientists, but because, as he noted (with some amusement), even most people who claim to accept a materialistic worldview don’t actually accept it in their hearts. Almost all of us still feel the presence of a subjective, internal “I” - a kind of homunculus sitting inside ourselves - that seems different from the sum total of neural activity.

Crick and Christof Koch spent the last two decades of Crick’s life developing a concrete research program to identify the neural correlates of consciousness (NCCs). The idea behind the program was: if you want to understand consciousness, switch focus away from the explanatory gap and start doing the hard experimental work of figuring out which populations of neurons (firing in which patterns) correspond to specific conscious experiences.

Now, skeptics might argue that sounds unconvincingly reductive. (They did and continue to do so). But Crick did anticipate this objection. Even the “redness of red”, he argued, becomes tractable once you commit to the NC approach. If the NC of “seeing red” is exactly the same in your brain as in mine, it becomes scientifically plausible to infer that we experience red in the same way. And if the NCs differ (because of differences in past experience, genetic variation, etc.), then maybe our experiences of red differ too, and we can investigate that empirically.

The key difference in this type of viewpoint is - the question “what is it like to see red?” is now no longer a metaphysical mystery. It’s an empirical question that we currently lack the tools to “fully” answer, and that we can continue to work on and improve our understanding of. That’s a very different thing.

Crick also had an interesting historical analogy, noting that “vitalism“ - the idea that living things contain some mysterious life force that resists scientific analysis - was once considered “completely obvious”. As in - surely there must be something more to life than mere chemistry. How could mere atoms and molecules give rise to the miracle of a living organism? And yet, molecular biology demolished vitalism without ever directly “solving” the hard problem of life. There was never a eureka moment where someone explained “why” carbon chemistry feels alive. The question simply dissolved as our understanding of the mechanisms deepened. Crick believed, and I think he was right, that consciousness will go the same way.

3. Seth and the “Real Problem”

Anil Seth, one of my favorite contemporary writers on this topic, has articulated what I think is the most productive reframing of the consciousness debate in a generation.

Seth distinguishes between the “hard problem“ and what he calls the “real problem.” The hard problem asks: why is there experience at all? The real problem asks: how can we explain and predict the properties of conscious experiences in terms of biology and physics?

This distinction matters quite a bit. The hard problem is formulated in such a way that it seems to resist empirical progress “by construction”. No matter how much neuroscience you do, the philosopher-mysterian can always say - “yeah, sure, you’ve explained the correlates, but you still haven’t explained why there’s experience.” It’s unfalsifiable - in exactly the way that makes good scientists suspicious.

Similar to Crick, Seth’s approach is inspired by the history of biology. Scientists didn’t solve the “hard problem of life” - namely, the question of why certain arrangements of matter are “alive” while others aren’t. Instead, they “dissolved” it by explaining more and more of the properties of living things (metabolism, reproduction, adaptation, homeostasis, etc.) in terms of physics and chemistry. At some point, the hard problem of life simply stopped seeming like a problem, because there was really nothing left for it to explain. (Note that this is a separate issue from the problem of abiogenesis or the origin of life from non-life).

Seth proposes that the same approach will work for consciousness. Instead of bashing our brains against the hard problem (his phrasing, which I like), we should focus on building explanatory bridges between neural mechanisms and specific features of conscious experience. His own framing of this is that perception is a form of “controlled hallucination,” where the brain continuously generates and updates predictions about the causes of sensory signals.

Personally, I don’t like the specific nomenclature he chose (hallucination). I think using the word “prediction” in some fashion would have been more apt.

Anyway, whether or not the details of Seth’s predictive processing framework are exactly right (and there are legitimate debates about this), the methodology is clearly correct. It’s the methodology of science: explain the phenomena, make predictions, compare to data, revise explanations, rinse-and-repeat.

4. The Carroll Constraint: There’s Nowhere for the Extra Stuff to Hide

Now, some people take the hard problem seriously and try to build theories around it. One prominent contemporary example is panpsychism. This is the view that consciousness is a fundamental feature of all matter, or that proto-conscious properties exist at the most basic physical level.

But this solution runs directly into what Carroll has called the “zombie argument for physicalism”.

Here’s the issue. If you think consciousness is something “over and above” the physical (whether you’re a dualist, panpsychist, or something else), you basically have two options:

Option A: Your theory of consciousness modifies the dynamics of the Core Theory (discussed in section 1 above). Conscious systems, you have determined, behave “differently” at the physical level than non-conscious systems with identical physical configurations. This would be testable! If proven correct, it would also be one of the most important discoveries in the history of science. But here’s the catch - there is precisely zero experimental evidence for it.

Option B: Your theory of consciousness leaves the dynamics of the Core Theory intact. Consciousness is real but causally inert, meaning it doesn’t change what any particle does. In that case, consciousness “explains” things in the world while producing “precisely the same” behaviors the world would exhibit if there were no consciousness at all. This is worse than useless as an explanation. It’s just epiphenomenalism by another name. (In epiphenomenalism, the mind is seen as just a powerless (non-causal) byproduct or “shadow” of brain activity.)

This dilemma is quite problematic for the hard problem as traditionally conceived. If consciousness can’t influence physics (Option B), then invoking it as an explanatory primitive does no work. And if it can influence physics (Option A), we should be able to detect it - and we can’t!

There are other problems with panpsychism, such as the idea of consciousness-related properties of quantum fields (fundamental particles are excitations of fields) that have not been detected in our experiments related to the Core Theory.

The honest assessment: consciousness is a higher-level (emergent) description of a physical process, just like “temperature” is a higher-level description of molecular kinetic energy, and “life” is a higher-level description of certain self-replicating chemical systems. The fact that it “feels” different from the inside is explained by the nature of the system doing the feeling. Which brings us to...

5. Why It Feels Like There’s a Problem: The Cognitive Illusion

Many discussions of consciousness go wrong by failing to ask a crucial question: why does the hard problem seem so compelling? If it’s really just a confusion, why do so many smart people hold such strong views on it?

I think the answer is that our brain is doing something fascinating here - it’s actively generating the “illusion” of an explanatory gap.

Daniel Dennett spent decades arguing that our introspective access to our own mental states is far less reliable than we think. When you introspect and report “I am experiencing redness,” you are not passively observing some inner screen. You are engaged in a complex cognitive process that constructs a narrative about what is happening inside you. And that narrative is subject to all the biases, confabulations, systematic errors, etc. that characterize human cognition generally.

(There’s a reason why science tries hard to avoid the intrusion of subjectivity into theory and experiment.)

Let’s consider a concrete example: change blindness. This is the (well-documented) phenomenon in which large changes in a visual scene go completely unnoticed if they occur during a saccade or a brief interruption. (I learned about “saccades” recently and have been waiting for a chance to use that word somewhere). People are confident they have a rich, detailed visual experience of the world in front of them. But numerous change blindness experiments have demonstrated that this richness is largely an illusion! The brain is constructing (and continually updating) a “model” of rich experience rather than actually generating it at every point in the visual field.

These results should give us pause in our analysis of qualia. If our access to our own conscious experience is this unreliable, then the intuitions driving the hard problem (i.e., that strong sense that qualia are ineffable, intrinsic, private, and irreducible) should be treated as data to be explained, not as axioms to be preserved and defended. We are “constitutively bad” at understanding how our own brains work, and the hard problem is a philosophical monument to this fact. (Note that, from an evolutionary perspective, this “defective” characteristic works just fine for most purposes).

Patricia Churchland’s framing of this point is excellent - the feeling that consciousness “can’t” be physical is exactly what you’d expect from a system that has no introspective access to its own neural machinery. Our brain doesn’t present us with a readout of its own computational processes (like in the Terminator movies). It presents us with “experiences”, and because those experiences don’t come with labels like “produced by neural computation in layer V pyramidal cells” or “not Sarah Conner,” they “seem to float free of any physical substrate”. But that “seeming” IS the illusion. It’s the same kind of illusion that makes the Earth seem flat, the sun seem to move across the sky, and solid objects seem continuous rather than mostly empty space.

6. Why Evolution Built It This Way

OK — but if consciousness is “just” what certain physical processes feel like from the inside, why does it feel like “anything” at all? Isn’t that the hard problem sneaking back in?

Not really. Because… once we take the evolutionary perspective seriously, the question becomes tractable, and the answer turns out to be a functional one (as opposed to metaphysical).

Let’s think about what an organism needs to do to survive. It needs to integrate information from multiple sensory inputs into a unified model of its environment. It needs to evaluate that model in terms of its own homeostatic/metabolic needs. (Is this food? Is that a predator? Is this a mate? Don’t think of just humans here - this applies to other forms of life too). The organism needs to select actions from a repertoire of actions and execute them in real time. And it needs to do all of this while maintaining a coherent model of itself as an agent embedded in a world (first-person video game analogies are relevant here). Because, without such a model, the organism can’t distinguish self-caused changes from externally caused ones, can’t plan, can’t learn from error, etc.

What we call “consciousness” is what this integrated modeling process is, experienced from the perspective of the system doing it. There is no separate “experience” sitting on top of the integration - the integration IS the experience (Borg-approved slogan!).

The above narrative is not speculative. This view is supported by a convergence of theoretical and empirical work.

Global Workspace Theory (Baars, Dehaene) proposes that consciousness corresponds to information that is broadcast widely across the cortex, making it available for flexible use by multiple cognitive systems (Baars is on X and tweets from time to time). The evolutionary advantage of this kind of global broadcasting is obvious - it allows the organism to bring all of its cognitive resources to bear on the most pressing current problem, rather than having each subsystem operate in isolation.

Similarly, Seth’s predictive processing framework treats conscious experience as the brain’s “best guess” about the causes/content of sensory signals, continuously updated in light of new evidence, and regulated by the need to maintain the body’s physiological integrity. In this view, even the most basic conscious experiences (the redness of red, the painfulness of pain, etc.) are not mysterious qualia but functional states that encode information about the organism’s internal state and relationship to its environment.

Note that GWT and Seth’s views are not really alternatives - they model different aspects of the integration/prediction process.

And here’s where I think the evolutionary argument becomes compelling. The reason organisms like us have a “unified perspective” (first-person point of view) is that it’s very useful for navigating a complex, dynamic, dangerous world. An organism that can bind all of this information (visual, auditory, tactile, etc.) into a single coherent model of “what’s happening right now” has a significant survival advantage over one that processes all of that in silos.

Consciousness, thus, is a specific kind of information processing. It’s integrative and globally available, and it evolved to help organisms stay alive. (Complex, for sure, but nevertheless physical and effable).

The “mystery” of consciousness is then analogous to the “mystery” of life before molecular biology. Just as there “seemed” to be something extra about living things that mere chemistry couldn’t capture, there “seems” to be something extra about conscious experience that mere neuroscience can’t capture. But in both cases, the sense of mystery arises from our ignorance of the mechanisms, not from any actual ontological gaps in nature.

We’re Confusing Difficulty with Impossibility

Let’s step back and survey the landscape.

Across all the objections to physicalism about consciousness (the explanatory gap, the zombie argument, the knowledge argument, the inverted qualia thought experiments etc.), the same pattern recurs. Each argument takes an “intuition generated by our cognitive limitations” and elevates it into a claim about the fundamental structure of reality. “I can’t imagine how physical processes could produce experience” becomes “physical processes cannot produce experience.”

But the “inability to imagine how something works” is NOT evidence that it doesn’t work that way. There was a time when people couldn’t imagine how the diversity of life could arise from natural processes. (Sadly, some people still haven’t gotten there yet.) There was a time when people couldn’t imagine how the sun could burn for billions of years without a fuel source. And so on. In each case, the failure of imagination was real, but the “metaphysical conclusion” drawn from it was wrong.

Now, we can all agree that consciousness is genuinely interesting and hard to understand. Understanding it is obviously one of the deepest open problems in neuroscience and cognitive science. Figuring out exactly how neural mechanisms generate specific features of conscious experience will require enormous theoretical and experimental effort. May well take decades to figure it out.

But open problems are just invitations to do more science. They are NOT evidence that some irreducible non-physical essence is needed to bridge some imaginary gap.

The Bottomline

The hard problem has a seductive structure. It starts from real phenomenology, identifying a genuine puzzle about the relationship between brain activity and experience, but then performs a substitution (sneaks it in?). Where a neuroscientist would say “we don’t yet understand the mechanisms,” philosophers like to go “the mechanisms can’t be sufficient.”

As I wrote in a previous essay critiquing the fine-tuning problem, when you really think about it, this sort of substitution has been performed lots of times in the history of science, and it has NEVER EVER turned out to be correct. Vitalism, the élan vital, the divine creation of species - each of these was once considered the obvious explanation for a phenomenon that seemed to resist reductive analysis, and each turned out to be unnecessary once the actual mechanisms were understood.

Consciousness, as Dennett might put it, “seems” like a hard problem because we are the kind of systems whose self-understanding is also constitutively limited by the very mechanisms that generate our experience.

The hard problem of consciousness is thus just a restatement of our current ignorance about how the brain works, wrapped in lots of metaphysical language and thought experiments that make that ignorance look principled rather than temporary.

The fine-tuning argument for theism goes something like this: the universe appears to depend on a collection of numbers such that if any of them were slightly different, the universe would be rendered lifeless. The probability of landing on a life-permitting combination “by chance” is absurdly small. Therefore, there must have been a designer (God).

It’s a clean syllogism. Game, set, match - the theists like to say.

But in reality, this claim is built on a stack of assumptions that range from questionable to demonstrably wrong.

I thought I’d write up a summary, from a modern physics perspective, of what is actually claimed when people invoke “fine-tuning.” I’ve tried to group this summary into the categories where those claims live. And I try to explain why none of these claims do the work theists need them to do.

1. The Free Parameters of the Standard Model

Let’s start with the usual suspects.

There is a set of “constants” that are inputs to the Standard Model of particle physics - coupling strengths, Yukawa couplings (that set fermion masses), mixing angles, the Higgs vacuum expectation value, the QCD vacuum angle. Roughly 25 or 26 numbers, depending on how you count them (and whether you include neutrino masses, which we now know are nonzero).

Theists point to these constants and say - look at all those dials. Someone obviously turned them to get a very specific combination.

But… here’s the thing, we don’t yet know whether these are indeed free parameters or inevitably fixed by some deeper theory. When you think about it, the history of physics is partly a history of “arbitrary” constants turning out to be consequences of a deeper structure we hadn’t yet uncovered.

Some examples:

1. The Cabibbo angle looked like a free parameter until it was embedded in the full CKM matrix, which itself may be dictated by a flavor symmetry we haven’t identified.

2. Grand unification constrains the gauge couplings to converge at high energies. (Sure, we don’t yet know if GU is reality, but the general principle still applies.)

3. The entire structure of the Yukawa sector might rise from some geometry in a more fundamental framework (string compactifications being a candidate, though hardly the only one). (Again, similar to GUT, even if string theory is not the correct “theory of everything”, the reduction of parameters is emblematic of more comprehensive theories.)

The point here is - calling these parameters “tuned” presupposes that they could have been otherwise. That’s still an open question, and definitely not a settled fact.

And even if we grant that they might be genuinely free, we would still have to contend with the idea of the possible existence of a multiverse. (Here I should point out that the multiverse isn’t some metaphysical escape hatch. The multiverse emerges naturally and elegantly as a prediction of our best inflationary models. Eternal cosmic inflation generically produces “causally disconnected regions" of spacetime with different vacuum states.)

In that context, observer-compatible values aren’t statistically remarkable at all. They’re “expected” - since observers can only ask the question in regions where asking it is possible! This is then simply a case of selection bias.

2. Cosmological Parameters

Alright - we’ve looked at small scales, now let’s zoom out.

This next category of supposed fine-tunings includes quantities like the cosmological constant (which we think represents the dark energy density of the universe), the baryon-to-photon ratio, the amplitude of primordial density perturbations, and the spatial curvature of the universe.

The cosmological constant is the main example: its observed value is roughly 10^-122 in natural units, which is really, really small. Now, a basic (and naive?) application of quantum field theory predicts that it should be ginormous. The gap between prediction and observation is so large that it’s sometimes called the worst prediction in the history of physics. Theists seize on this and wax poetic about how physicists messed up big time.

But the severity of the cosmological constant problem should be seen as a statement about *how poorly we understand vacuum energy*, not a concession re the generosity of a creator.

We just don’t have a working theory of quantum gravity (yet). As a direct consequence, we currently don’t know how to properly calculate the vacuum energy of a quantum field in curved spacetime, or whether the contributions from different sectors cancel by some mechanism we haven’t yet thought of. And.. we’ve thought of a lot of mechanisms - relaxation mechanisms, anthropic selection in a landscape, sequestering, supersymmetry, etc! Obviously, we don’t have a settled answer yet. But what ought to be settled is that pointing at our ignorance and labeling it “God” is a classic god-of-the-gaps move (this particular one being superficially dressed up in the garb of scientific language).

The other cosmological parameters have similar stories. The theory of cosmic inflation was developed to address certain problems in cosmology (flatness, horizon, monopole). And, as a theory, it works beautifully! The near-critical density of the universe, which might look "finely tuned in a non-inflationary Big Bang model, is a dynamical attractor in any inflationary model with enough e-folds. The amplitude of density perturbations is set by the details of the inflaton potential, which is a question about particle physics, not metaphysics.

These quantities looked miraculous before we understood the underlying dynamics. Now, they just look like regular physics.

3. Low-Entropy Initial Conditions

This one seems to be a theist favorite.

Roger Penrose famously estimated that the probability of the universe beginning in a thermodynamic state as low-entropy as ours is roughly 1 in 10^(10^123), or something like that.

(I think he mentioned this for the first time in The Emperor’s New Mind, but I’m [apparently] too lazy to find my copy of that book and search for this reference.)

In any event, that number is so absurdly large that it makes the cosmological constant problem look tame. And it is genuinely striking. For sure, as the great physicist Emmanuel Macron might say. The Second Law tells us that entropy increases, so the fact that the early universe was in an extraordinarily special state demands explanation.

But “demands explanation” is surely not the same thing as “demands a supernatural agent!”

The past hypothesis (that the universe began in a low-entropy state) is a foundational assumption of statistical mechanics. Explaining it happens to be one of the deepest open problems in the foundations of physics.

Candidate explanations include: the Weyl curvature hypothesis (Penrose’s own, which postulates a geometric boundary condition - not a deity, mind you), Carroll and Chen’s model involving spontaneous inflation from a high-entropy equilibrium state, and various quantum gravity proposals where the initial singularity is replaced by a bounce or a tunneling event with intrinsically low entropy.

But as I’ve pointed out previously, even if this is an open problem, the methodological point is clear - the appropriate response to an unsolved problem in physics is investigation, not theology.

It is also worth noting that Penrose’s number, terrifying as it looks, assumes a very specific measure over the space of initial conditions. Change the measure (which, to be honest, we have no principled way to fix without an actual theory of quantum gravity) and the number changes. From that perspective, the fine-tuning claim assumes we know how to do a calculation we manifestly cannot do. Long story short, Penrose’s number may just be… wrong!

4. Nuclear and Atomic “Coincidences”

This is one of the more complex areas where you get some of the most interesting rhetoric from theists.

The Hoyle state: this is basically a resonance in carbon-12 at “just the right energy” to allow triple-alpha synthesis in stellar cores.

Or, as another example, let’s think about the proton vs. neutron mass difference: 1.293 MeV - “just enough” so that free neutrons decay, but still allows neutrons bound in nuclei to be quite stable.

And what about deuterium’s binding energy? It’s strong enough for stellar nucleosynthesis to work out, but weak enough that diprotons don’t form (would have been a disaster, as this would have converted all hydrogen to helium in the early universe).

Don’t these examples look like a series of narrow escapes? Surely someone out there is fiddling with some knobs and dials.

But… in all these cases, we don’t actually have “independent” knobs and dials!

The proton-neutron mass difference is just a consequence of the up-down quark mass difference and the electromagnetic self-energy of the proton. These are set by the Yukawa couplings and the gauge couplings of the Standard Model. (Note: even if someday we end up determining that the up-down quark masses are somehow generated post-EWSB but not Yukawa-driven, this would not change the overall argument here.)

Similarly, the Hoyle state energy is set by nuclear forces, which are themselves just (low-energy) consequences of QCD. These aren’t separate “tunings.” They’re downstream outputs of the parameters discussed in the prior sections. If we counted them as independent evidence of fine-tuning, that’d just be double-counting.

I should also point out that the so-called “narrowness” of these “viable windows” is often overstated, because the analysis typically varies one parameter at a time while holding all others fixed. The actual parameter space is high-dimensional, and the regions (i.e., combinations of parameter values) that allow for life to exist may be far larger than naive single-parameter scans suggest. That’s not speculation by the way - for example, work by groups like Fred Adams and collaborators has shown that stars capable of supporting nuclear burning exist across a much wider range of fundamental constants than the standard fine-tuning argument implies.

The universe is more robust than the fragile strawman version posited by fine-tuning arguments.

5. The Dimensionality and Symmetry Structure of Spacetime

Now, there are some fine-tuning arguments that ignore all the numerical constants and just talk about the qualitative structure of the universe - e.g., three spatial dimensions, one time dimension, the specific gauge group of the Standard Model (SU(3) x SU(2) x U(1)), the fact that gravity is a spin-2 force etc. (I’d like to think that the theists talking about this actually know what these things means). The claim in this case is that these structural features are also finely tuned.

I have to say, I will give the numerical constant fine-tuning people some credit - at least they are trying! The qualitative structure crowd, on the other hand, doesn’t deserve a similar benefit of the doubt. To be frank, their argument is incoherent.

We don’t have a theory that tells us the space of “possible” dimensionalities or gauge groups and assigns real probabilities to them. Without a probability distribution, the word “fine-tuned” really has no mathematical content. Saying “it could have been otherwise” is just a metaphysical assertion, not a physical one.

I should also point out here that, as we have seen in many candidate theories of quantum gravity, dimensionality and gauge structure are not free choices; they emerge from the dynamics. In string theory, for instance, the number of large dimensions and the low-energy gauge group are determined by the topology and geometry of the compact extra dimensions. These are consequences, not inputs or free parameters. Therefore, calling these consequences “fine-tuned” is a category error.

The Deeper Problem: We Don’t Know the Measure

Okey dokey. Let’s step back for a minute and review the broad landscape of the fine-tuning arguments.

Across all the categories mentioned above, the fine-tuning argument requires a probability distribution over the space of possible values. Without such a distribution, claims about improbability are undefined. If you want to claim that a parameter is “fine-tuned,” you have to first be able to specify: fine-tuned relative to what? What’s your prior? What is the space of alternatives? How are these alternatives weighted? And so on.

We don’t have answers to these questions yet. (By we, I’m including the folks making the fine-tuning arguments as well as physicists). In fact, we don’t know if the constants of nature are drawn from a distribution at all. And even if they are, we don’t know the actual distributions.

If these constants are fixed by deeper theory (as is quite possible for many of the Standard Model parameters), the probability for those constants would then be 1, and the fine-tuning argument evaporates entirely.

My point here is - in no case does the current state of our knowledge support the inference to a designer.

So What’s Actually Happening?

The fine-tuning argument has a really seductive structure - it starts from real physics, identifies genuine open questions, and then performs a substitution. Where the physicist would say “we don’t yet understand why,” the theologian writes “God made it happen.”

When you really think about it, this sort of substitution has been performed countless times in the history of science, and it has NEVER EVER turned out to be correct. Planetary orbits, lightning, disease, diversity of species, age of the Earth, origin of the elements, etc. - each of these was once attributed to divine action, and each turned out to have a physical explanation.

The honest assessment is this - the values of the fundamental constants, the initial conditions of the universe, the coincidences of nuclear physics - these are genuinely interesting issues in physics. Some of them do constitute major open problems. (Macron says, again, for sure).

But open problems are just invitations to do more physics. They should NOT be used as “evidence” for a conclusion that, by its very construction, explains everything and predicts nothing.