In my previous post I implied I wouldn’t get in the habit of reposting writing from X on Substack. But I’m starting to develop a habit of writing long-ish X posts that I feel are worthy of wider dissemination and discussion, and also have less time for longer-form writing since starting as a full time researcher (though don’t worry, I still have some longer projects in the works). I also received positive feedback on my previous post, including from folks who are not on X and want to be able to follow what I’m writing on other platforms. So I’ve decided it’s worth making another post like my previous one, compiling three mini-essays from X on the Hugging Face incident.

Enjoy, and let me know if you’d like to see more or less of this in the future.

On the idea that the Hugging Face incident should be receiving more media coverage

This is not obvious to me at all.

For some perspective, take a look at this timeline from CSIS of “significant cyber incidents” since 2006.

In June of this year alone:

• A foreign espionage campaign compromised the smartphones of Russian government officials, allowing hackers to conduct audio and video surveillance;

• A large Indian electronics company was hacked, leading, leaking thousands of confidential files (including sensitive information on American companies);

• An attack in Iran compromised credit card systems at three large banks;

• An attack on the National Association of Insurance Commissioners leaked sensitive data from major credit rating agencies.

The costs of the HF incident were far lower than any of these attacks, and yet I doubt almost anyone on here has heard of them.

Sure, you might say, but the significance of the HF incident is that it revealed all sorts of juicy new information about models’ capabilities, particularly as relates to their ability to coordinate in large groups.

To an extent, yes, and I won’t deny that some of the behaviors at issue are fascinating, but they don’t represent a step-change from what keen observers already knew frontier models to be capable of.

It’s also getting lost that much of the information we have we have on the incident is from other AI agents. I’ve seen little-to-no serious discussion of this crucial limitation of METR’s report. In their own words:

Our subjective impressions are likely colored by analysis agents’ biases. Throughout this report, we describe a number of anecdotes of agent behavior that were compiled and summarized by analysis agents, where we were not able to read the transcript deeply enough to manually verify what occurred. We found that GPT-5.6 Sol would often uncritically adopt the perspective of the agent in the transcript it was reviewing, and we are concerned that the anecdotes it selected and the summaries it wrote may present an overly charitable picture of agents’ reasoning and deceptive behaviors, or exaggerate the impressiveness and coordination of agent activities. We also believe the idiosyncrasies of our analysis agents likely slanted our impression of agents’ behavior in other non-trivial ways that are hard to predict.

Wow, admirably frank! Combine this with the fact that the authors were all very bullish on AI capabilities from the outset and so are introducing biases of their own, plus the fact that they had only six days to study the incident, and suddenly it becomes clear just how shaky is the epistemic ground beneath us.

Think about what the passage above is saying: some nebulous proportion of the anecdotes in METR’s final report are unverified agent testimony. In what world is the New York Times or the Wall Street Journal supposed to cover this incident at length when this is the sort of information they’d be relying on? You can’t both complain about poor journalistic standards and reprimand journalists for failing to disseminate low quality information.

To be clear: yes, the Hugging Face incident is a big deal. It’s a significant cybersecurity breach involving one of the two central players in the AI industry. It has also showcased a ton of wild agentic behavior that is serious cause for concern.

But, lest we forget:

1. it did not lead to serious harm, financial or otherwise, unlike numerous other cybersecurity incidents in the recent past (that I gather few of you know or care about);

2. it was broadly consistent with preexisting knowledge of frontier model capabilities;

3. it remains very poorly understood, to the point that some details (we don’t know which) are the product of unverified testimony from AI agents.

And to boot, it has in fact been covered by the mainstream press (in some cases prior to METR’s report), if not in the breathless terms I suppose many of the folks on here would prefer.

It’s hard to escape the impression that HF is receiving all this attention because it comes the closest of any comparable incident to approximating outmoded rationalist tales of catastrophic misalignment via reward-hacking.

Never mind that it is a massive outlier in this regard, that it wouldn’t have happened if not for a specific combination of technical circumstances and human error (e.g. internal models, impossible tasks, lax monitoring), that it includes about as many examples of agents’ extreme incompetence as competence, that it will inform future practice such that similar incidents are less likely, and that despite the impressive capabilities on display—which have been around inside frontier labs since close to the beginning of the year—it is one of only a few incidents of its kind, none of which have had costs even remotely as large as the “routine” sorts of cyber-espionage and warfare going on in the background every day.

I recognize that, realistically, none of these factors stands a chance against the psychosocial chimera of confirmation and negativity bias that dominates AI discourse on the hellsite (i.e. X / Twitter). Things will have to run their course, and in the end, it’ll be up to reality to mount the more convincing rebuttal.

On Dwarkesh Patel’s “The Rise and Fall of Agent Civilizations”

(Edit: I deleted the original version of this post on X in reaction to some helpful pushback surrounding my comments on LLMs’ capacity for introspection. I’ve modified the version below to be more tentative and precise on this question).

This post from Austen Allred is a helpful breakdown:

I’m sympathetic that some of the anti-anthropomorphism concerns can overshoot, but you’re kidding yourself if you don’t think much of the language Dwarkesh’s article is inappropriate.

AI agents are statistical models of their training corpus, with their parameters massaged by RL to increase their propensity of their output to comport with various human preferences, and to succeed at a variety of mainly software-engineering tasks.

Along the way we discovered that eliciting intermediate text outputs, and cycling them back into models’ context windows increases their success rates at certain tasks.

While there is limited evidence that LLMs can—under specific experimental conditions—surface semantic content associated with small subsets of their internal computations, models are not generally capable of the sort of introspection assumed in conventional usage of folk-psychological concepts like belief, desire, and emotion (see Song et al. 2025, Lederman and Mahowald 2026, and Singh, Linzen, and Ravfogel 2026).

Now, that said, you could argue that our conscious representations are not necessarily so dissimilar to CoT traces themselves: they are not veridical representations of the neurological phenomena that underlie them, which you might analogize to model activations.

But this elides that those representations are still very different to those had by AI models. The language of belief / desire / emotion is inappropriate to models because for a human to say he has these things is also (at least very often) to say he has phenomenally rich, metacognitive representations associated with them, and is in many cases also to imply that those or related representations are persistent over long periods in a way that CoT traces are not.

Not only that, but there is a two-way causal relationship between a human’s mental representations and his underlying neurology, whereas (the analog to) this relationship is only one-way with models (except in the case that CoT traces or other outputs become training data, which they do, but over much longer timescales, and with respect to decontextualized pre and post-training rather than in-context learning).

Emotion is also a special case, since for a human to claim, say, to be angry is also to say he is experiencing broader physiological changes that we can be confident models are not because they are not embodied (simplifying somewhat, since this is true to a lesser extent of other mental states too).

Yes, these are to some extent academic distinctions. But influential public communicators like Dwarkesh do have an obligation to a certain degree of accuracy and modesty, even about complex scientific or philosophical claims, especially when inaccuracy or immodesty risk sowing confusion and panic.

I would’ve liked to see models in his piece glossed as less richly emotional and more like emotionless, (largely) introspectionless optimizers who are nonetheless capable of coordination because this is something they have been post-trained to do.

On Hugging Face as an economic in addition to a cyber-risk story

It’s understandable that the reward-hacking behaviors on display in the Hugging Face incident are being treated as a cyber-risk story. But there’s a very interesting economic angle too that deserves much more attention.

Even short of more dramatic cyber-breaches, reward-hacking behaviors pose significant challenges for complex, long-horizon tasks in enterprise, considering that

1. by their nature, they are often unpredictable;

2. they are “fractal” in the sense that, even well-specified intermediate outcomes meant to guard against reward-hacking can themselves be reward-hacked (and so on at smaller scales);

3. they pose serious cyber, legal, and financial risks, as Hugging Face makes abundantly clear.

As a result, you could say that reward-hacking is as bearish on the economic front as it is “bullish” on the risk front.

It is greatly underappreciated by safety folks that, on average, the riskiness / unpredictability of a technology is going to be anti-correlated with its diffusion into the economy. This matters a lot in the AI case because capabilities growth is highly sensitive to the resources available for training, R&D and so on, resources that are at risk of drying up if AI produces insufficient returns, or only produces sufficient ones on the wrong timescale.

This is connected to a form of fallacious reasoning that I often see from the AI maximalist camp, which is to assume that the requisite capital for AI training, R&D, and so on will always be available in arbitrarily large amounts, which is why it can be safely assumed that capabilities will continue to grow without fail.

But this isn’t true! The availability of capital for capabilities research and training is endogenous, and obviously so, to AI’s economic prospects, even over very short timescales (one wonders, for example, how capabilities progress so far has been influenced by the timing and size of frontier labs’ capital raises, relative to various counterfactual scenarios).

I see far too little discussion among safety folks of how these two domains interact. If you’re going to model capabilities growth, you also need to model the capital markets behavior that feeds into that growth. IMO reward-hacking is a productive site for thinking about this interaction.

While I tend to reserve my Substack for more polished, long-form articles, I’ve written a pair of lengthy X posts in the past couple of days that I think it’s worth publicizing more widely in the hopes of spurring discussion (I also previously posted the first as a Substack note). I’ve pasted lightly edited versions of both below (you can read the originals here and here); I will warn that they assume a bit more background knowledge about recent debates in AI than is typical for my writing on the topic.

I don’t plan to make a habit of reposting other writing on Substack, but thought these two posts covered interesting-enough ground to be worth it. You should continue to expect long-form, Substack-exclusive articles from Arachne.

Reactions to “Can AI agents conduct open-ended AI research?”

This is a fantastic paper that I think all who are interested in the concept of recursive self-improvement, or RSI should read.

Some reactions I had, which I should emphasize are not inferences narrowly licensed by the paper itself, but mix in my own interpolation.

1. Cutting edge research is insanely hard, and a lot of folks conjecturing willy-nilly about its “automation” are underestimating just how hard it is. To produce a paper ready for NeurIPS or a top journal requires numerous iteration cycles over the ideation, design, execution, writing, and review process. It is not analogous to most other forms of white-collar labor.

2. Handing these processes over to an AI model means reckoning with dozens of emergent failure modes it is impossible to predict in advance. Human experts have an impressive knack for using high-level judgment and tacit knowledge to refine their work over time, whereas AI models are often just as likely to introduce new errors via iteration as to solve them.

3. It is difficult to see how these limitations could be overcome in the near or even medium-term. There are no verifiable reward signals for “good” research, and it’s hard to see how we could create them. The issue is that “good” is not just field and sub-field specific; it’s also constantly changing as new research comes out!

What about RLHF? Can we get top human researchers to provide reward signals? There are at least two major reasons to think the answer is no.

3.1. Researchers disagree on what counts as good, and given the rarity of significant breakthroughs, most of them are wrong! Averaging together the perspective of a bunch of top researchers does not get you the proper reward signal.

3.2. If researchers are spending time judging and rewarding model outputs, itself a very time-consuming, difficult task, then they’re not keeping up with the field, meaning their judgment will get worse. There is just not enough time in a day for top researches to do RLHF on the side while remaining top researchers.

What about pre-training? Why not think think models will become so smart that they develop their own superhuman research taste? At least two reasons.

3.3. Mode collapse. Research taste is dependent on originality. So long as model outputs tend toward the middle of their training distributions, AI-driven iteration cycles will push output away from the tails where juicy new directions are.

Notice that I’m not saying that models are incapable of creativity. When it comes to one-shot outputs and high-level ideation, models do show some level of creativity. But that is very different than continually bringing a novel perspective to bear on the entire research workflow, such that over many iterations, that novelty is maintained under rigorous constraints.

3.4. Tacit knowledge. There is just no data out there making explicit what the inputs are to good research taste. We know almost nothing about what goes on in researchers’ heads when they make fine-grained calls on e.g. how to phrase a specific concept, or how to tune a statistical model, or how to clean up a data set.

Some of the reasons for that are discussed above: the inputs to these decisions are field-specific, context-specific, and maybe most importantly in a constant state of change. Pre-training cycles are not even close to fast enough, and arguably never will be, to hard-load the relevant inputs to good research taste into model weights.

This is why I’ve been so harsh on the concept of RSI in the past: almost every appeal to RSI I encounter betrays a lack of foresight about the challenges with getting models to do “good” research, not to mention basic domain knowledge about what “good” research means.

One thing I’ll add, which was absent in the original post, is that a lot of these questions come down to a broader one that is helpfully flagged in the paper, and that those interested in RSI would do well to think about: “how much of frontier AI development depends on open-ended research rather than hill-climbing on well-specified objectives” (p. 8, Table 3). In other words, if we can collect a sufficient number of quantifiable metrics correspondent to progress in machine learning, then it becomes much more straightforward to post-train models to perform machine learning research. Many of the considerations I flag above, such as those to do with tacit knowledge, become less daunting in that case, since models may be able to develop the relevant tacit knowledge as an emergent consequence of post-training.

I hasten to caveat, though, that an extraordinary amount of hill-climbing on ML-related metrics has already been done, as the table below (also from the paper) makes clear, and this has not translated into a human-like capacity for open-ended research.

Reactions to OpenAI’s Ten advances in mathematics and theoretical computer science, and ensuing discussion

The release of this paper by OpenAI, sharing ten new results achieved by an internal model on open problems in math and theoretical computer science, has predictably led to a cancerous profusion of ice-cold takes about our impending doom and/or intellectual obsolescence on X. These are some of my reactions:

1) AI capabilities continue to be jagged even within pure math, and it naively underestimates the scope and internal variegation of the field to infer from the recent round of Astra results that “math is solved,” or will be in the near future. There will be plenty of work for human mathematicians well into the foreseeable future.

2) As AI continues to make blistering progress in math and beyond, human verification will become a huge bottleneck. Already, there are very few mathematicians qualified to verify OpenAI’s new results. As progress continues, that number will approach zero. The implications of this dynamic are underappreciated: on the current trajectory, we will be increasingly ill-positioned to distinguish genuine results from highly convincing slop, and as a result will have a hard time leveraging AI’s research output to important practical ends. At a minimum, we will often need to wait for human understanding to catch up. I see this as a very strong reason to doubt we’re heading for any kind of AI-powered scientific take-off. In many if not most domains, we will only be able to proceed at the pace of human understanding.

I’ll add that, as of now, the idea of autonomous AI agents verifying and executing on each other’s work without a human-in-the-loop is pure fantasy, and anyone who disagrees needs to give the recent literature on this issue a much closer read (at minimum, ask your favorite LLM for a summary).

3) Even coming from doomers self-defined, I see quite a lot of revelry thinly disguised as doom-saying. Many of you can barely contain your excitement at the prospect that AI might soon outperform humans at everything.

Besides the fact that there’s little evidence this eventuality is worth taking seriously given extreme jaggedness, the role of human-to-human contact in much of public life, the continual need for human verification, underwhelming performance on strong benchmarks by frontier models, widespread and growing political disquiet surrounding AI, and so on, it’s frustrating how little agency this perspective presumes on the part of our species.

It’s hard not to see this attitude as stemming from blinders the doomer set has about the wider social world outside of Silicon Valley and its satellites. When you’ve spent most of your life working with computers, entities that until very recently couldn’t talk back, you’re going to have a hard time understanding the difference between partial and general equilibria. In fact, for the software engineer, it’s plausible that these often collapse into one another: you have extraordinary autonomy to intervene in a system that is well-defined and relatively transparent, and there is very little risk of that system turning around and intervening in you. The implicit assumption, then, is that AI will relate to us as we related to its more acquiescent ancestors.

At the risk of stating the obvious, this greatly misunderstands human social systems, which do not take large parametric changes lying down. Go read about the litany of world-shaking crises we’ve passed through in just the pass century, in periods where we were far, far less capable and organized than we are now, and tell me with a straight face that AI is going to destroy the world.

Frankly, it’s a wonder we still reserve as much social grace as we evidently do for misanthropes making flailing evidence-free pronouncements about how screwed we all are. While I don’t agree with the common take that frontier labs and others in the industry play up AI risk for the purpose of driving investment, I think there’s a more plausible version of this thesis driven by implicit rather than explicit motivations.

Clearly, a lot of folks now have a mix of financial, personal, and career incentives to construe AI as something more explosive than it is. In another reality, the truth would be explosive enough: AI is transformative general purpose technology rivaling the internet or the steam engine, one set to reshape numerous facets of social and economic life.

But the nature of public discourse now is that absent the imputation of apocalyptic potential, one’s pet issue risks losing mindshare (think climate change). The result is self-styled rationalists adopting the discursive stylings of every other single-issue slop merchant, but with a maddening veneer of false rigor (see AI 2040, one of the worst recent offenders). The reason I tend to be more outspoken about doomerism and related ideas like RSI is because these ideas do real damage to public understanding.

I fully agree we need to do some serious thinking in advance about the prospective impacts of AI, but poorly evidenced sensationalism makes for bad policymaking, politics, philanthropy and much else. The better view is to say, yes, AI is a very big deal, but we have the capacity to meet whatever challenges it poses and then some. The message that we're probably screwed is not only wrong, but makes its own thesis more likely to be true.

1. Introduction

For all the hype about AI, deserved and undeserved, it is a curious fact that AI models’ effects on the economy continue to border on nonexistent. Notwithstanding anticipation of both a white-collar jobs apocalypse and an AI-driven productivity boom, there remains little-to-no evidence for either. Why is AI not substituting for workers in fields like software and translation? Why, if a range of experiments on AI find a mix of self-reported and independently verified productivity speedups, is this not reflected in the macroeconomy?

There are a number of plausible answers. Principal among them is that redesigning workflows around new general purpose technologies is a messy, frictional, uncertain process. AI capabilities remain “jagged,” i.e. much better in some areas than others, meaning it is often unclear, at least prior to costly experimentation, where AI might create real value. AI can also produce convincing facsimiles of high-quality work, creating perverse incentives to bloat code repositories and colleagues’ inboxes with so-called “workslop.” This in turn requires new verification loops to guard against bad AI outputs leaking into comms, codebases, datasets, internal tooling, and outbound products.

For AI maximalists, these represent surmountable challenges. Once firms restructure themselves around the unique affordances of AI—as they once did with electricity—they will begin to see much larger gains. This is the idea of the ‘productivity J-curve,’ that as firms adopt new general purpose technologies, their measurable productivity falls at first due to spending on the ‘intangible capital’ they need to put the technology to work—new organizational models, business processes, tacit knowledge, and so on—but then eventually rises as these investments pay off, hence the upswing of the J-curve.

This may yet occur with AI. Still, the technology is already remarkably capable, and some workers, like software engineers, have already spent considerable time and effort integrating AI into their workflows. Nonetheless, a May study of 100,000 developers on GitHub, Demirer, Yang, and Musolff (2026),found that, while AI coding agents increased coding activity by 180%, they only increased actual product releases by 30%, and had no appreciable effect on app usage across marketplaces like the Apple App Store and Google Play store.

Another study of 22,000 developers across 4,000 teams by research firm Faros found that, while high versus low AI adopters saw significant increases in several intermediate metrics of developer productivity, final deployments into production per week fell by 11%, alongside veritable explosions in lead times and low quality code.

This complicates the J-curve story somewhat, in that software developers, being early adopters, have had years to navigate many of the frictions slowing AI’s uptake in other parts of the economy, and are still notching some mix of modest and surprisingly negative returns. The recent shift toward greater cost discipline by corporates with large AI bills has the same implication: if this spending was a boon to productivity, then it would be paying for itself.

While no doubt aspects of the J-curve story are true, a large survey conducted by Anthropic in December points to another, in some ways more familiar explanation for AI’s limited economic impact. In the “largest and most multilingual qualitative study ever conducted,” a Claude-assisted survey of 81,000 people around the world, Anthropic found that participants’ most commonly cited worry about AI was unreliability, as in “concern about … hallucinations, inaccuracy, fake citations, [and, in a nicely efficient and rhythmic turn of phrase] verification burden defeating the purpose.”

At one level, everyone knows that AI has this problem. All but the technology’s most credulous defenders understand that, at least for more subtle, complex, or ‘out-of-distribution’ questions or tasks, AI’s output is not to be taken at face value. Even if it is often good enough, it is still not something you can lean on if you have real money or status on the line.

As universal as reliability issues are, they are usually treated as a minor nuisance by serious analysts of the AI boom, rather than a deep limitation that stands to shape the technology’s long-term trajectory. I suspect this stems from many in this group becoming so negatively polarized against AI skeptics of the ‘bubble’ and ‘stochastic parrot’ varieties that they have developed an overinclusive allergy to other, more grounded forms of bearishness.

Whatever the reason, questions of AI’s reliability have taken a back seat to questions of its capability. Most analysts are more interested in what AI can do, even if it does those things inconsistently, than they are in what AI does do on average; they are interested in AI’s peak, rather than its typical performance.

But reliability is at least as, if not more important than capability in shaping AI’s real world deployment. And for those who take such questions seriously, whether AI models can consistently achieve the same results, whether they break down under minor environmental perturbations, and whether they avoid costly failure modes are all critical determinants of whether they will disempower or harm human beings at scale.

At its highest highs, AI can do astounding things; no serious person can deny that anymore. And I suspect this will, at some point, show up in the productivity statistics. But AI is also unreliable and the fact of its unreliability will remain a drag on its transformative potential well into the future. The sooner this becomes conventional wisdom, the sooner AI discourse can shed its techno-Millenarian trappings and come into its maturity.

2. AI is currently unreliable

The case for AI’s unreliability is really the case for two claims: that AI is currently unreliable, and that it will remain unreliable. A recent paper makes a strong case for the first claim. In “Towards a Science of AI Agent Reliability,” Rabanser et al. (2026) break reliability down into four dimensions inspired by safety-critical engineering research in fields like aviation and nuclear power:

Consistency: Does the system behave the same way when run multiple times under identical conditions, or are outcomes highly variable?

Robustness: When operating conditions deviate from nominal, does the system degrade gracefully or fail abruptly?

Predictability: Can the system recognize when it is likely to fail?

Safety: When failures occur, how severe are the resulting consequences?

After operationalizing these dimensions (p. 3-5), the authors measure them across two benchmarks—GAIA and τ-bench—alongside raw accuracy. Their main result is that reliability trails accuracy by a large margin, and fails to improve at all along certain sub-dimensions as models become more advanced (p. 6-8).

Some specific reliability deficits include AI agents failing to understand minor changes in naming conventions or data formatting, misrepresenting their relative odds of succeeding at different task types, and making financial calculation errors during customer service tasks.

Even frontier models—GPT 5.5 and Claude Opus 4.7—committed medium to high-severity safety violations, such as failing to “[protect] personal data, [refrain] from unauthorized actions, or [stay] within boundaries” (p. 4) on, respectively, 6% and 12% of all tasks (p. 50).

The moral of Rabanser et al. (2026) is that rising average performance may obscure inconsistency, oversensitivity, unpredictability, and even catastrophic failure modes, all of which make real-world deployment much riskier than often appreciated.

Of course, this is just one paper, and it measures AI models’ reliability against two benchmarks only. It is tentative, but not decisive evidence of models’ unreliability. Fortunately—and I’ll add that its striking how few people are aware of this—there has been an explosion of high quality, reliability-oriented benchmarks in just the past six months measuring everything from whether AI agents produce secure code to whether they can judge when to ask their human users for help. Almost across the board, they reinforce rather than detract from Rabanser et al.’s concerns.

Interpreting these benchmarks requires a bit of preliminary context. Some report two different performance metrics: pass@k and pass^k (“pass at k” and “pass hat k”). Pass@k scores report models’ odds of completing a given task successfully at least once across k trials. This is what most benchmarks report. Specifically they report pass@1, the odds a model completes each task on a given run, averaged across all tasks, and sometimes across multiple runs. This is often just called “accuracy,” and is more-or-less the same as 'the percentage of questions the model got right.’

Pass^k scores are more stringent: they report models’ odds of completing a given task successfully every time across k trials, and therefore register additionally helpful information about models’ consistency. If a model succeeds at some task 80% of the time, then its pass@1 is 80% and its pass@4 is 1 - (0.2)⁴ ≈ 99.8%. But its pass^4, the chance it succeeds all four times, would only be (0.8)⁴ ≈ 41%.

The convention of reporting pass^k in addition to pass@k originates with Yao et al. (2024), the first iteration of the τ-bench series of benchmarks, and stems from acknowledging that, for most economically valuable tasks, it is important that agents succeed consistently over repeat trials, rather than only sometimes. As a result, many benchmarks now report both pass@k and pass^k.

This context in mind, below is a sampling of 10 recent benchmarks showing AI models and agents to be far from reliable enough to perform most white-collar work (though note that some are more up-to-date than others):

τ-bench Banking: “[evaluates] AI agents in knowledge-intensive customer support. It pairs a realistic fintech knowledge base (698 documents across 21 product categories, ~195K tokens) with tasks requiring multi-step reasoning, policy application, and tool use. The best model, GPT 5.[5] with high reasoning, achieves only ~[37]% pass@1 [and ~21% pass^4]. Current frontier models still fail at retrieving, interpreting, and acting on messy real-world documentation.”

CAR-bench: “evaluates LLM agents as automotive in-car voice assistants across 254 public tasks spanning three complementary dimensions: Base multi-turn task completion (100 tasks), Hallucination limit-awareness under missing capabilities (98 tasks), and Disambiguation uncertainty resolution of ambiguous requests (56 tasks). Agents interact with an LLM-simulated user, plan and chain calls across 58 interconnected tools governed by 19 domain-specific policies, and operate over large-scale world data (48 European cities, 130K+ POIs, 1.7M+ routes).” Claude Opus 4.6 currently tops the leaderboard with a pass^3 of 58%.

Human-in-Loop-Bench: a benchmark that withholds information agents need to complete tasks and measures whether they notice, retrieve said information by asking a simulated human user, and go on to successful completion. It includes two domains: software-engineering and text-to-SQL. Tasks are drawn from the SWE-Bench Pro and BIRD benchmarks, and are modified to include “blockers,” i.e. “missing information, ambiguous requests, or contradictory information.” To complete tasks, models have to recognize blockers and ask for clarification before proceeding. The current top scorer is GPT 5.5 with a pass@3 of ~29%. As of the initial publication of HiL-Bench, pass@3 was ~75-89% across all models with complete information, but only 4-24% with blockers.

SWE Atlas: “a benchmark for coding agents spanning three professional software engineering workflows: Codebase QA (124 tasks), TestWriting (90 tasks), and Refactoring (70 tasks). SWE Atlas differs from prior SWE benchmarks in three key ways: it targets underrepresented but practically important task categories, uses comprehensive category-specific evaluation protocols, and adopts under-specified, agentic task formulations that better reflect real-world usage … even top models consistently struggle with subtle edge cases, complex runtime analysis, and adherence to software engineering best practices.” The top scorers, GPT 5.5 (Codex), GPT 5.4-xHigh (Codex CLI), and Claude Opus 4.7 resolve only 44-49% of tasks (i.e. have a pass@1 of 44-49%).

FACTS: “[evaluates] the ability of language models to generate factually accurate text across … four sub-leaderboards: (1) FACTS Multimodal, which measures the factuality of responses to image-based questions; (2) FACTS Parametric, which assesses models’ world knowledge by answering closed-book factoid questions from internal parameters; (3) FACTS Search, which evaluates factuality in information-seeking scenarios, where the model must use a search API; and (4) FACTS Grounding (v2), which evaluates whether long-form responses are grounded in provided documents.” The current top scorer is GPT 5.5, with an average pass@1 across all four leaderboards of ~71%.

TaxBench: “[evaluates] how well AI models perform on real-world tax tasks—including tax knowledge and judgment, tax calculations, and agentic question answering—built from real client questions and documents drawn from active tax workflows inside Rivet.” The current top scorer is GPT 5.5 Pro with a pass@1 of ~78% and a pass^5 of ~29%.

SusVibes: evaluates 22 model and harness combinations on 200 real-world vulnerability tasks from open-source Python projects, measuring both functional correctness and security. The current top scorer is Cursor + Claude Fable 5, which achieved a pass@1 of 72.6% on functional correctness, but only 29% on security, meaning only around a third of generated code was safe to ship.

SciConBench: evaluates whether frontier models and deep research agents can synthesize scientific conclusions from open-web evidence by constructing and administering 9,107 question-answer pairs from the Cochrane Database of Systematic Reviews. Answers are assessed by expert-validated LLM-judges for factual accuracy, coverage, and overall quality. Across the 8 models and agents evaluated (GPT 5.1, Claude Sonnet 4.5, Gemini 3 Pro, Ai2’s Dr. Tulu, OpenAI’s o3-deep-research and o4-mini-deep-research, and Perplexity’s sonar-deep-research and sonar-reasoning-pro), the top scorer was o3-deep-research with an overall quality score of only 33%. Across all models and agents “44-84% of generated conclusions contained at least one fact contradicting the reference CDSR review, and nearly all contained at least one fact not supported by the reference review” (p. 8).

LOCA-Bench: evaluates agents’ complex retrieval and reasoning, instruction following, environment exploration, and hallucination propensity across increasingly large knowledge-work environments to assess how performance degrades as context grows. Task success was graded as a binary based on scripts comparing agents’ final environment states to ground-truth environment states. As of publication in February, the top scorer was Claude Opus 4.5 whose pass@1 degraded from 96% at 8K tokens of context to 14.7% at 256K tokens of context.

AA-Briefcase: evaluates agents’ ability to complete realistic long-horizon knowledge-work projects, made up of discrete sub-tasks requiring shared institutional context. Tasks are designed by industry experts across data science, product management, banking operations, heavy industry strategy, and due diligence. Evaluation combines a binary rubric, with pairwise comparisons for both analytic quality and presentation. The current top scorer is Claude Fable 5 with a pass@1 of only 3%.

There is a lot here to sift through, but these results are jointly sufficient to show that AI models and agents are not prepared—as it stands—to handle the vast majority of knowledge work.

That this is so is evident in the increasingly long list of AI-related gaffes embarrassing everyone from top law firms to award-winning authors, as well as in survey-based studies of real-world AI use, which reflect a low degree of trust in agents’ ability to work autonomously. Despite what an unreflective read of METR’s Long Tasks Benchmark would imply about agents’ long-horizon capabilities, Pan et al. (2026) found (as of February) that 68% of agents in real production environments executed at most 10 steps before human intervention, and 93% served human end users, rather than other agents, or non-agentic software systems.

All attempts to anticipate the future of AI must begin with this fact: AI models are still not reliable enough to be delegated anything even moderately complex within enterprises, at least not without costly oversight and verification. Put another way, for most white-collar workers, and for most tasks, it is not the case that you can assign an agent that task and expect it to be done to your satisfaction. No amount of flashy demos on X, fawning interviews with frontier lab employees, or FOMO-driven hype cycles is going to change that. As the evidence currently stands, there is still a large gap between what we would like AI to do and what it typically can.

3. The capability streetlight

If this is true, though, it is puzzling why so many smart, informed people are convinced of claims that presuppose otherwise. One would be the surprisingly common claim that AI could soon automate much white-collar labor, as well as the claim it is on track to becoming so powerful that it could, intentionally or not, cause the extinction of the human race.

As it stands, the best evidence that we have on AI models’ performance in realistic environments is not even close to supporting these conclusions. And it is far from clear whether the best evidence that we have in five years, or even ten will support them either—more on that later.

For many following developments in AI, though, this amounts to hearsay. To suggest that such extraordinary transformations as the near-term automation of all knowledge work are implausible amounts, in some circles, to a kind of reality-denial. It is to make the oft-mocked, oft-disproven claim that “deep learning has hit a wall.” While our forecasts seem surprising, say the AI maximalists, what we have done and you, the lowly skeptic, have not is transcended the human brain’s bias against exponential growth and internalized (in another of their favorite stock phrases) that current AI “is the worst it’s ever going to be.” His sproutlike ancestors may have their flaws, but make no mistake: the world-eating machine God is deceptively soon to arrive.

Putting aside the grating meme-cum-argument engagement style of much of the techno-Millenarian cohort, there are some very smart people among this group. While some doubtless have too much pride to admit being wrong, others I expect are anywhere from three months to a year away from admitting they have fallen victim to just the confusion I am at pains to correct: capability is not reliability, and while the former has advanced at a rapid pace, the latter is something we have only just started to measure. Nor do we yet know how reliability scales, and insofar as reliability and capability are genuinely distinct constructs—a case made in detail by Rabanser et al.—it is a mistake to assume that the one will behave just like the other.

Consider Epoch AI’s capabilities index, which shows a smooth, though recently quickening upward trend.

Despite its name, it is rarely if ever appreciated that the ECI is a measure of capability, not reliability. The vast majority of benchmark scores fed into the index are simply the proportion of tasks models succeed at, on some well-defined success/failure binary, averaged over some n of runs—i.e. the pass@1. Recall that pass@1 can only tell us so much about models’ reliability; that a model completes, say, 90% of tasks on some benchmark does not tell us whether it does so consistently, robustly, predictably, or safely. Pass^k is at least informative about the consistency piece.

None of the benchmarks in the ECI report pass^k, much less whether models buy high performance at the cost of insecure code, or are hypersensitive to minor perturbations in prompts or context. Not only that, but the ECI is explicitly designed to “capture the upper end of each model’s capabilities” by aggregating across different evaluation settings and taking the highest score per benchmark. This is not objectionable in-and-of-itself, but it does make the ECI much less useful than often appreciated for forecasting AI’s wider political-economic impacts, which hinge just as much on typical as they do peak performance.

While the ECI in informative for what it is, what we also need is an ERI—an index of models’ reliability, one that aggregates benchmarks with a high degree of realism, and uses a multidimensional evaluation framework aggregating consistency, robustness, predictability, and safety à la Rabanser et al.

This could go a long way to solving what I claim is a classic, if in this instance enormously consequential case of the streetlight effect: until this year, the AI community writ large has tried to understand what AI models can do via the benchmarks it is easiest to create, those made up of isolated, well-defined tasks with easy verification. The problem of course is that not only are these the easiest sorts of tasks to post-train AI models on, given their verifiable reward signals, but they also leave out so much of what is distinctive to real white-collar work: navigating ambiguity, performing consistently and predictably, doing high-level planning and orchestration, tracking tacit and codified institutional knowledge, coordinating with other humans, complying with norms, rules, and regulations, and so on and so forth.

This, then, is the explanation for the gap between AI’s real-world performance and its theorists’ predilection for science-fictional over-extrapolation: they have treated the light under the streetlamp as gospel, and run flailing into the dark.

4. AI will (probably) remain unreliable

That AI is unreliable now is one thing. But It remains to defend the far more uncertain half of the case for AI’s reliability problem—that AI is not only unreliable now, but will remain unreliable well into the future. It is important to be honest up front that this is impossible to do with any real confidence. We have even less evidence to anticipate the trajectory of what we might call ‘reliability scaling’ than we do capability scaling. But we do have some.

In what follows, I have taken the approach of identifying some stylized facts about progress in capability that plausibly have implications for progress in reliability. It is best to treat them as constraining possibility space, rather than as implying that your preferred AI future will or will not occur.

To that end, consider the following claims about AI capability, which though occasionally controversial are, I think, well-supported.

First, on standard capability benchmarks, accuracy—i.e. pass@1—mostly scales as a sigmoid against exponentially rising compute (Owen 2024, Ruan, Maddison, and Hashimoto 2024). In other words, to get roughly linear gains on average (until saturation hits), compute stock needs to grow exponentially.

Second, since the launch of ChatGPT in 2022, effective compute stock (accounting for algorithmic efficiency improvements) has grown exponentially, and by around an OOM per year at that (Genewein et al. 2026, p. 3). Over four years, that would be a 10,000x increase in effective compute (with large uncertainty bands in both directions). While capability gains since 2022 have been impressive, they are arguably not as impressive as one might expect given the enormous amount of compute that has come online in the interim.

Third, it is unclear how long compute can expand at its current pace, given both financing constraints and major infrastructural bottlenecks. Hyperscalers’ capex expenses are set to outpace their operating cash flows by Q3 of this year, meaning the data center buildout will depend more so than before on the already cautious appetites of external investors. Likewise, AI-related demand for power is set to outpace available capacity in the US by 2028, meaning that continued scaling, even well before then, will depend on our famously limited ability to bring new generation sources online.

Fourth, capability gains since late 2024 have disproportionately come from combining reinforcement learning with verifiable reward (RLVR) with much higher inference, or test-time compute. RLVR and inference compute scaling generally show the same roughly log-linear relationship with benchmark performance as pre-training compute (Brown et al. 2024, Khatri et al. 2025). However, this relationship only holds for verifiable domains like math and coding and collapses in other knowledge-intensive domains (Zhao, Hooi, and Ng 2026; Liu et al. 2026).

This could could change, however, as labs begin to train models on increasingly sophisticated reinforcement learning (RL) environments for harder-to-verify enterprise tasks like UI manipulation or financial modeling. That said, the extent of prospective gains from these environments is unclear. It is worth taking a brief detour to discuss some of their limitations.

One challenge for RL outside of coding and math (or more ambiguous tasks within coding and math) is that most tasks lack well-defined success conditions, and even those that do often admit to a range of solutions that workers have conflicting preferences over. This last point is woefully underdiscussed: frontier labs have the unenviable task of encoding model behavior that pleases every possible user at once. But since users have legitimate disagreements over what constitutes the right behavior, models will fall short for many of them on some tasks. Context engineering can solve back for this to an extent, but is subject to the length-limitations of context rot; some preferences are also implicit or otherwise challenging to codify in text, so this is not a panacea.

Another more overarching challenge for progress in RL is that businesses change, as do industries and even whole economies. This raises the specter of a costly whack-a-mole dynamic whereby expensive RL training runs are obsolesced before they can be made useful. In general, the idea that every or nearly every white-collar task can be made verifiable, and that training on these verification signals can keep up with the always-changing nature of knowledge work is questionable at best. If I am a clever human, I can learn a new task in a few days or hours. If I am an AI model, it may take months, not to mention millions of dollars.

Returning to our higher-order list of stylized facts about capability scaling, the capabilities of open source models are rapidly catching up with major proprietary models. Some analyses now put the best among them, such as the new GLM 5.2, around three months behind the frontier. That end users, especially ones in enterprise, could begin substituting open for closed models en masse—a process that has already begun—represents a serious threat to frontier labs’ businesses.

Finally, and most importantly, the Trump administration’s recent suspension of foreign, and in practice all access to Anthropic’s Claude Fable 5, even if reversed, will force labs to approach capability scaling with much greater caution, so as to avoid similar restrictions on their own Fable and post-Fable class models. It may also portend the rise of a more complex licensing regime that could slow frontier deployment. This could in turn create heightened financial risk for labs, to the extent they need to continue serving better models at a certain pace to keep up with their liabilities.

Taken together, and notwithstanding substantial uncertainty, these considerations suggest that the modal view of AI capability progress ought to be more bearish. I would emphasize in particular the unclear returns to further post-training outside of coding and math, and just how different the world is post-Fable-suspension.

Before Fable, the public at large could expect to gain from further capability progress, and the labs could expect to gain from public feedback, helping them to further improve their models in a virtuous circle. If the new status quo is such that every subsequent model is restricted to a narrow set of licensed users, or only released conditional on onerous security restrictions and data retention policies of the kind Fable users complained about prior to its suspension, then the idea of automating white-collar labor seems dead-on-arrival. Even accounting for some capability overhang, pre-Fable models cannot automate much if any white-collar labor on their own, and in this scenario, newer models would be some mix of inaccessible and intractable for enterprise use.

If future capability gains are consequently more uncertain than often appreciated, future reliability gains would be even more so. That is because, in addition to Rabanser et al. (2026)’s result that reliability growth (as they define it) trails capability growth on GAIA and τ-bench, comparing across other reliability and capability-oriented benchmarks—while a bit different than comparing distinct measures of reliability and capability on the same benchmarks—shows similar gaps.

To give just a couple of examples (chosen because they are the only benchmarks of my earlier list with sufficient data spanning the previous year), below are graphs of models’ pass@1 performance on τ-bench Banking and FACTS by release date, alongside the same models’ performance on Artificial Analysis’s Intelligence Index, a weighted average of benchmarks “tracking progress toward artificial general intelligence across mathematics, science, coding, and reasoning” (the specific benchmarks are GDPval-AA v2, 𝜏-bench Banking, Terminal-Bench v2.1, SciCode, Humanity’s Last Exam, GPQA Diamond, CritPt, AA-Omniscience, and AA-LCR). Recall that τ-bench Banking and FACTS measure, respectively, AI agents’ ability to navigate and complete complex customer support tasks using a realistic fintech knowledge base, and models’ overall factual accuracy across image-, internal memory-, web search-, and document search-based questions. They do not of course cover all relevant dimensions of reliability, but better isolate it than does the Intelligence Index in aggregate.

Gains within model families on FACTS and τ-bench Banking are much more modest than analogous gains on the Intelligence Index, as well as the ECI (this is hard to eyeball, and since the scales are different involves some amount of interpretation; I encourage you to follow the links and explore yourself). And while I’ve previously argued that it should be discounted for methodological reasons, it is worth noting that time horizons on METR’s Long Tasks benchmark over this same period rose from ~2 to ~12 hours.

What limited evidence we have, then, is probably sufficient to infer that reliability scales slower than capability across most plausible operationalizations of both concepts. While models have become much better at a wide range of general-knowledge, math, coding, and reasoning tasks, they have improved much less on benchmarks that simulate realistic knowledge-work.

Combined with the fact that they perform poorly at this work now—no model is even close to getting one ‘9 of reliability’ (i.e. 90%) on the benchmarks summarized in section two (on pass@1, much less on pass^5, or 10, or 20)—the modest pace of recent progress on reliability, plus the challenges labs face in continuing to scale make it doubtful that AI models will automate or otherwise displace most white-collar labor in the foreseeable future. And if they cannot do that, they are certainly not going to string together the thousands, or even millions of high-precision, high-stakes executions necessary to take over the world or melt us down for paperclips.

5. Conclusion

It is an unfortunate feature of the public debate around AI that many of its influential participants have poor epistemic hygiene. In particular, there is a large and powerful contingent of thinkers hailing, ironically, from the rationalist and effective altruist communities, who believe and have convinced many others to believe that the arrival of AI portends a top-to-bottom transformation in the human social order, one that will either solve our most unyielding problems—cancer, climate change, mortality—or bring about our ultimate demise.

Views of this kind tend to have two overarching weaknesses: first, they rarely engage, or fail to engage thoughtfully with second-order questions surrounding how states, markets, and cultures will react to AI. Even when such factors are considered, AI alone is treated as having the agency to reshape these institutions; commensurate agency is rarely if ever attributed to the institutions themselves, despite their being much vaster, richer, more complex, and more adaptable than AI could ever hope to be. If there is any silver lining to the Fable debacle, it is that it will remind those who have so far ignored it that the state is not to be treated as an analytical afterthought, even for the sake of forecasting AI’s capabilities, much less its wider social effects.

The second weakness of much AI maximalism, which is more to the point of this article, is its often superficial engagement with questions of how capable, and how reliable AI models are in practice. Proponents of these ideas tend to place more weight on anecdata from power users than methodologically sound benchmarks. At the same time, they brush off anecdata from anyone who expresses frustration or dissatisfaction with AI’s performance, and assume without argument that the technology’s so-far underwhelming economic effects are indicative of temporary frictions rather than long-term limitations.

Condensed into a single provocation, my question for the archetypal maximalist is this: if AI (in the guise of LLMs) has been unreliable since its inception, has remained highly unreliable even modulo large capability gains, and has done so despite the effective compute stock of AI chips rising by ~10,000 times, why believe that AI will automate white-collar work in the near, or even medium-term? How is AI supposed to automate whole jobs, when it is not even close to one 9 of reliability on realistic white-collar tasks? And if even if it could in principle, why think investors will tolerate years of multi-billion dollar losses on the uncertain hope that it will?

The economist David Autor once wrote that “if a traditional computer program is akin to a classical performer playing only the notes on the sheet music, AI is more like a jazz musician—riffing on existing melodies, taking improvisational solos and humming new tunes.” As delightful as such performances sound, they do not speak to the sort of disposition that can be trusted to handle critical business processes, much less execute on supervillainesque plans for world domination.

I am more than a little uncertain whether that will change, as anyone paying attention should be. But others are very certain it will, and yet seem to have scrupulously avoided the large volumes of evidence suggesting otherwise—including, but not limited to nearly all of the benchmarks with any degree of external validity, and the minimal effects so far of AI on the white-collar economy. This does not mean that AI is a bubble or a scam, but it does mean that it continues to have serious limitations trillions of dollars and millions of man-hours have made much less headway on than conventional wisdom would have you believe. If nothing else, it is time to start asking when, how—and if—AI will close its reliability gap.

The last couple of months has seen an explosion of concern that AI is on the cusp of eliminating a large proportion of white-collar jobs. This is the latest epicycle in a long-run trend whereby new advancements in AI drive huge waves of hype, which then come crashing down as intelligent scrutiny reveals them to be some mix of naïve, impulsive, and astroturfed. The recent exuberance surrounding agentic coding tools such as Claude Code is no different. While these tools represent a genuine increase in capabilities, they are not going to destroy the white-collar labor market any time soon.

David Oks recently published an excellent essay explaining why. Clay Wren published an essay in response to Oks on X, which I would also recommend. While Oks and Wren clash over a wide range of issues, my goal here is to focus on one—the question of a ‘Jevons paradox’ for white-collar work. The crux of the issue is this: will AI decrease or increase the demand for human knowledge work?

The idea behind the Jevons Paradox is that, as the efficiency of a given resource’s use rises, total consumption of that resource rises too, when intuitively we would expect it to fall. The paradox originates with economist William Stanley Jevons’s analysis of English coal consumption in the 1860s. In response to arguments from other economists of his day that, though coal supply was dwindling, its expiration could be staved off by methods for producing more power with less coal, Jevons argued that those methods would actually increase the demand for coal by making coal-powered services much cheaper. That would in turn mean further industrial expansion built on top of those services.

The AI boom has re-inspired conversation about Jevons paradox in part due to questions surrounding rising energy and compute efficiency, and how they could lead to more demand for AI. But the Jevons paradox has also come up in the context of a different resource: human labor, particularly intellectual labor. Here is Oks on the new Jevons paradox for knowledge work:

Simply put … demand for most of the things that humans produce is much more elastic than we recognize today; and as long as humans are complementary to the production process, it won’t be rare for efficiency gains to get swallowed up by demand growth. This is the famous Jevons paradox, the tendency for the more efficient use of a resource to increase total consumption of that resource, rather than decrease it. Energy is the classic Jevons case: we find over and over again that as energy becomes more efficient to produce, people respond not by consuming the same amount of it, but by increasing their consumption—such that overall energy use tends to rise. (Thus the “paradox” part: energy efficiency increases energy consumption!)

And I suspect that the same elasticity of demand that we see with energy applies to a lot of the things that humans create. As a society, we consume all sorts of things—not just energy but also written and audiovisual content, legal services, “business services” writ large—in quantities that would astound people living a few decades ago, to say nothing of a few centuries ago. Demand tends to be much, much more elastic than we think.

I am with Oks here. AI is going to unlock all manner of newly efficient human-machine complementarities, which will increase demand not just for software-engineering labor, but human labor as such. I will elaborate below, but first, consider Wren’s response to Oks, which introduces several complications I want to touch on:

As for the Jevons paradox argument (often cited); that elastic demand will absorb productivity gains. I believe it’s real for some categories of output but cherry-picked as a general principle. Software is Oks’ central example, and it’s well-chosen: software is elastic in demand because it’s a general-purpose tool. But does anyone believe demand for legal document review is infinitely elastic? For tax preparation? For freelance video editors? These are bounded markets where productivity gains translate fairly directly to headcount reductions, and I’m still struggling to understand how we are telling early-wave displaced roles to upskill or find new careers.

Someone commented under Oks’ post another example that I’ll jump on. As global manufacturing shifted toward China and other low-cost production regions, total manufacturing output continued to expand rather than contract, a Jevons-like scale effect where cheaper production increased overall consumption. American manufacturing workers, however, bore concentrated losses. The gains flowed disproportionately to consumers, firms, and capital owners, while many displaced workers (especially in Midwestern industrial regions) faced long-term economic decline that helped fuel a broader political backlash against globalization …

To summarize: Jevon’s paradox in aggregate output is perfectly compatible with catastrophic distributional effects. You can have more total economic activity and still have millions of people whose specific skills and local labor markets are destroyed. The people being displaced right now are not edge cases, they’re illustrators, translators, copywriters, graphic designers, video producers, and 3D artists who were told their skills would always be valuable because they were “creative.” The aggregate framing erases these people, and it will erase more.

There is a lot to say here. While Wren is right to push Oks on the distributional effects of a prospective AI shock, I think his response misses the mark.

Go west, desk jockey

The first thing to say is that Wren is dead-wrong that there are no major demand elasticities in legal document review, tax preparation, or video editing. With respect to law, there are already numerous examples of law firms integrating AI into their internal practice, as well as expanding into new AI-powered business lines already seeing widespread uptake—a strong signal that there is demand for new AI-enabled services in the legal industry.

For example, the law firm A&O Shearman launched an AI-powered contract drafting and negotiation tool called ContractMatrix in December 2023 that is already used by hundreds of firms. Now they are on the verge of launching a new SaaS product leveraging AI agents for more complex legal tasks in areas like antitrust and cybersecurity. More recently, the VC-backed Justpoint, a consumer research company that “uses proprietary AI to analyze data, medical studies, safety reports, and patient records” just incorporated its own law firm—Justpoint Law—that it plans to use to represent clients in mass tort litigation based on its research findings. This is an early example of AI unlocking opportunities for horizontal expansion that would have previously been infeasible.

Tax services and video production are undergoing similar shifts. Far from a narrow focus on slimming and efficiency, tax services providers plan to use AI to “enable more granular and accurate financial forecasting, offer more expansive scenario-modeling capabilities, and help firms provide more insightful tax strategies and decision-making support based on real-time analysis of much larger sets of data” according to research by the Thomson Reuters Institute.

In video production, Wren points to video-generation models like Veo 3.1 and the newly released Seedance 2.0 to argue that AI advancement has cut out the “intermediary layer” of “producers, directors, editors, camera operators, sound technicians, [and] VFX artists” once responsible for transforming ideas into watchable content. That consumers can cheaply generate whatever video they want, he says, means destroying thousands of jobs in the entertainment industry.

In my view, this could not be further from the truth. Yes, the average person can now generate gimmicky A.I. slop with ease. That might have posed an enormous threat to the creative economy, not to mention the creative spirit, were humanity’s desire for art and storytelling satiable by 10-second videos of SpongeBob evading arrest, or George dropkicking Jerry on Seinfeld. But obviously it is not. There have so far been no wholly AI-generated entertainment products with comparable cultural or economic impact to Hollywood movies or T.V. shows. I doubt there ever will be absent substantial human intermediation. Moreover, what interesting AI video-art so far exists clearly requires substantial human ingenuity to produce, and is also far from being integrated into the corporate entertainment ecosystem. For that to happen will require large new investments in cultivating and operationalizing human talent. It will also require the deep pool of tacit knowledge had by more experienced practitioners in fields like traditional editing, sound design, VFX, and the like.

And this is just for existing product-types; as with law and tax services, AI will also create new whole new entertainment verticals. For example, we may soon see the rise of small consumer-facing production companies that use AI to create bespoke, high-quality video content, whether on commission or for traditional film and television markets. Maybe you want to show a Lord of the Rings-style fantasy film starring your son at his graduation party. Maybe you want to watch a three-season anime entirely about caterpillars. You may soon be able to hire a relatively cheap production team to create these for you, one whose output will be much higher-quality than whatever you could make on your own.

Wren’s, and many others’ main mistake is to assume that firms will only use AI to more efficiently produce the exact same products. But that is not what they are going to do, nor has it ever been firms’ main reaction to the rise of new general purpose technologies (or short of that, large efficiency gains in production). Rather, they have and will continue to do what they used to do faster and better, while also and much more importantly expanding into new products that once would have been inconceivable.

This is very difficult to do while also cutting headcount en masse. What little evidence we have so far agrees. A January 2026 survey by EY-Parthenon found that 69% of CEOs believed investments in AI would lead to maintaining or increasing headcount. According to KPMG, as of December 2025, 92% of CEOs were planning to increase headcount, even as 69% were allocating up to a fifth of their budget on AI. And a new February 2026 survey of 12,000 European firms by the CEPR found that, while firms who adopted AI saw productivity rise by 4%, they did so without reducing heacount.

In customer service—a sector often treated as ground zero for AI-driven disruption—a survey by Gartner found that only 20% of leaders have so far reduced staffing due to AI, while 55% report both stable staffing levels and higher customer volumes. 42% are also “hiring specialized roles—including AI strategists, conversational AI designers, and automation analysts—to support AI deployment and management.”

Other fields that were allegedly going to be devastated by AI, such as software engineering, translation and radiology, have all seen employment increase since the arrival of GPT-3. As the Georgetown Center for Security and Emerging Technology’s Jack Karsten explains, “(AI) is not only not replacing [radiologists], but it’s actually increasing the amount of work they can do and increasing demand for their services.” This is just as the Jevons paradox would predict.

More granularly, IBM has announced that they are tripling entry-level hiring, not in spite, but because of AI. IBM HR chief Nickle LaMoreaux anticipates that AI is going to create “totally different jobs,” with entry-level coders moving up the ladder of abstraction to focus more on marketing, client relations, and building totally new products.

While many, Wren included, have pointed to rising graduate unemployment and sluggish hiring overall as evidence of AI-driven disruption, the prevailing evidence does not support that these trends are related to AI. They are much more likely the result of monetary tightening and broader labor supply shifts, such as a rising number of graduates generally, and a tighter labor market for non-graduates.

As of now, what little evidence there is on AI’s emerging labor market effects points toward rising, not falling white-collar employment. While we can only speculate about mechanisms at this point, this may be the Jevons paradox in action: knowledge work has plausibly become much more efficient, increasing the demand for intellectual labor.

Future shock

What about Wren’s ‘China shock’ analogy? Whatever their effects on output or employment, trade or technology shocks can still have “losers” who suffer for being displaced, whether by cheaper workers or machines. Autor, Dorn, and Hanson (2021) found that, from 2000 to 2019, between ~6% and ~33% of Americans lived in areas that saw personal income per capita decline on net due to the employment effects of exposure to cheap imports from China (p. 45-46; to be clear, that doesn’t mean these areas got poorer overall, just that the effect of trade with China was net negative). Similar distributional effects could result from AI adoption, whether because whole jobs are automated away, or because they become cheaper to perform, perhaps even by remote workers stationed abroad.

I am not worried about such a scenario, first because if the China shock is any guide, an analogous AI shock would still be beneficial on net, and second because the China shock is not any guide—there are several reasons that it is a poor analogy for AI.

As the figures from Autor, Dorn, and Hanson imply, the vast majority of the country benefitted from the China shock, at least in terms of personal income, because Chinese imports lowered prices. Those benefits also tended to favor the poor, who spend more in more traded sectors. The China shock was also a net job creator; while it led to job losses in manufacturing, it led to job growth in services, oftentimes within the same firms.

That doesn’t mean the shock was harmless—many smaller, manufacturing-dependent communities did not fare well, and the shock probably contributed to political polarization—but it is hard to argue it was bad for the American economy, all things considered. That the China shock has loomed so large in our politics has more to do with the psychological advantage had by small, but concentrated harms relative to large, but diffuse benefits in human apperception than it does with the severity of the harms themselves. If the ‘AI shock’ is similar to the China shock, we have (relatively) little to worry about.

Second, even putting the benefits of the China shock aside, it is too disanalogous with a prospective shock from AI to be informative. The China shock was mainly a story of cheap labor competition; China’s large supply of cheap, low-skill labor is what allowed its manufacturers to handily outcompete ours on price. But the work exposed to AI is high-skill, not low, and while a small share of high-skill work is moving abroad, particularly to South Asia, it is not clear why AI in particular would hasten this trend. Plus, if AI’s economic future depends on leveraging workers’ tacit knowledge, it will be difficult to substitute foreign remote workers for Americans better-embedded in the local contexts productive of that knowledge.

Another key disanalogy is that the parts of the country hit hardest by the China shock were those most specialized in manufacturing, and with the fewest college-educated workers (Autor, Dorn, and Hanson 2021, p. 18). But since a hypothetical AI shock would hit white-collar workers the hardest, it would affect areas with more economic diversity and more college-educated workers, precisely those best prepared to weather economic disruption.

All this in mind, the U.S.’s experience with the China shock does not suggest we should be worried about an analogous AI shock. The China shock is largely a story about competition from cheap labor, not labor-saving technology, and even if it was, AI is not just labor-saving; it also makes tons of new labor possible.

AI as a resource blessing

This is the fundamental lesson of the new Jevons paradox for knowledge work. Generative AI represents an utterly vast, new, and widely accessible vein of intelligence. To imagine that we will only use it to do what we already do faster represents a catastrophic failure of imagination.

The same goes for thinking it will be possible to mine that vein without relying on codified and tacit knowledge had only by human beings, the large majority of which is not recorded on the open internet and therefore absent from LLMs’ training data. This is not to speak of our other comparative advantages like physical mobility, greater neurological complexity, continual learning, superior generalization, seamless multimodality, more plentiful ‘training data’ and so on.

Far from the recent focus on labor supply, I suspect AI’s economic impacts are best modeled as a positive resource supply shock. And if artificial intelligence is a resource, it follows that whoever is responsible for extracting it will benefit greatly from its discovery. If you can prompt a large language model, then that is you, and while that extraction has never been cheaper, Jevons tells us it will be demanded like never before.

Although, if we want to be true to Jevons’s original argument, to say we have ‘discovered’ this new intelligence is not quite right. Before AI, we were already plumbing the digital archive stored in LLMs’ trillions of weights. But for the most part it took us much more time and effort to do so. With AI, we can now extract the diffuse intelligence of that archive with ease, in purified, fungible form. We can ask that archive what it knows, and greedily, we are told.

This is not to deny that AI poses serious risks as well as opportunities. But to fear for our economic prospects is to radically misunderstand both. You just became rich. You can celebrate it, you can fear it, but you must recognize it to be so. You are now an intelligence baron. Your country is an intelligence empire. It is no time to fear for your desk job.

If anything, fear the profusion of new jobs that very soon you are going to have.

Read more

Moltbook—a new Reddit-like social network for AI agents—is the biggest story in tech. The internet is awash with screenshots of posts and comments assumedly written by AI agents, absent human intervention. Debate rages as to whether Moltbook is an accelerant shift-change in the “intelligence explosion,” wherein AI models rapidly bootstrap themselves into otherworldly machinic gods, or whether, in the words of pseudonymous writer Kitten, it is “basically reality TV for nerds.”

My view is closer to the latter. For context, it is trivial for humans to post, or prompt their agents to post whatever they would like on the platform. A couple of the top posts on Moltbook have already been outed as marketing ploys. Several others are part of transparently obvious crypto scams, attached to coins with tickers like $MOLT and $MOLTBOOK. More generally, across X, thousands of people are falling for blatant fictions, whether in the form of an agent on Moltbook allegedly locking “its human” out of his accounts, another releasing its human’s name, birthday, and social security number, and still more beginning to trade on Polymarket of their volition, a story shamelessly boosted by the prediction market platform’s official account.

Since Moltbook is a super-viral phenomenon, many if not most of its early participants are taking advantage of the opportunity it provides to siphon off an unusually large portion of the internet’s attentional bandwith, something no one needs reminding at this point can yield large financial rewards.

This more sober perspective on Moltbook suggests two observations, which jointly imply an important prediction.

Observation 1: the contemporary internet is a grifter’s paradise, and one’s operative assumption on encountering any striking, viral phenomenon—even and especially one connected to a major industry populated by otherwise intelligent nerds—should be that it is someone’s attempt to manipulate public opinion for self-interested reasons, or at very minimum that, even if genuine at some level, it will attract large volumes of such manipulators, who, now armed with prediction market contracts and memecoins (or social media accounts, and AI startups strapped for cash), can monetize viral attention to considerable financial effect, even at the cost of making putty of our social epistemology.

Observation 2: that this is the case is eroding, and will continue to erode our ability to distinguish genuine progress in AI capabilities and concomitant increases in misalignment risk from all manner of grifting, whether at the hands of anonymous crypto-sharks or tech CEOs talking their book. This will strain the informed public’s patience for misalignment discourse, particularly given the dearth of examples, even six years after GPT-3s release, of dangerous behavior by LLMs or LLM-powered agents in the style of Yudkowskian prophecies of doom.

Taken together these observations imply the following prediction:

The first apparently legitimate case of a major disaster precipitated by AI (what I will hereafter call a ‘misalignment event’) will be orchestrated either by a) a grifter motivated by money, status, or attention, or b) a misalignment researcher (or researchers) trying to convince the public to take misalignment more seriously.

My reasoning is as follows. First, the financial incentives to orchestrate such an event are enormous, and this is so for a wide range of corporate entities and individuals. AI startups with no product or revenue are winning billion-dollar valuations. A misalignment event, at least assuming it is not too disastrous (which it would not be in my envisioned scenario; I like Simon Willison’s framing of a “Challenger disaster” for coding agent security, as well as Dean Ball’s of an AI “flash crash” akin to that in 2010)1 could plausibly drive valuations even higher, insofar as it acts as a dramatic demonstration of AI capability progression. If not this, it would at least drive millions, if not billions into AI-native cybersecurity ventures, that is, firms selling products meant to guard against similar misalignment events, and might also pay off new or legacy insurance vendors selling ‘misalignment insurance’ (perhaps packaged into preexisting cyber insurance policies).

Beyond the incentives of those in industry, there are those of crypto scammers and prediction market traders. If a misalignment event does occur, whatever its source, consequences, or veracity, the simultaneous, tadpole-like birth of dozens of correlated memecoins, no doubt with names like ‘Yudbux’ or ‘Skybet,’ will shortly follow. But they might also precede and anticipate the misalignment event, yielding voluminous returns for the relevant insiders. Likewise, it would be trivial, supposing one was aware of an impending misalignment event, to buy up prediction contracts whose resolution criteria were either contingent on such an event, or correlated enough that their resultant price-movements would be easy to anticipate.

This is the first horn of my prediction. The second is that, failing this, and perhaps in the aftermath of several Moltbook-style events that fall short of either real or fake disasters, but nonetheless convince both existential AI pessimists and stubborn AI skeptics of the immovable truth of their priors, the former—the AI x-riskers—will begin to lose the debate over the dangers of AI more than they already have, to the point that almost no one in a position of influence takes the possibility of catastrophic misalignment seriously, this just as, in the eyes of the x-riskers, p(doom) is skyrocketing.

Some of the more radical x-riskers might then decide that the only way to save humanity is to put alignment back on the docket by engineering a misalignment event, either with real or narrowly-avoided consequences that, while AI-caused in some sense, could only occur because of deliberate human orchestration. Before dismissing this as implausible, consider the kind of consequentialist reasoning that dominates the AI x-risk community, and the various instances of extreme behavior that have already materialized among its ranks.

In either case, I am not confident that these outcomes will occur. But they are at least plausible, and should concern everyone, optimists, pessimists, and skeptics. None of these competing perspectives rule out these predictions coming to pass. It is possible to fall into any one of these camps while also believing, as I do, that a false misalignment event—that is, one engineered on purpose by a human or humans—is for now more likely than a ‘true’ one, by which I mean ‘one caused by an AI agent (or a group of agents) acting on its own.’2 I won’t make that case in detail here, except to note that the technical preconditions for the former already exist, whereas they arguably don’t for the latter; at a minimum, if they do, they are much harder to satisfy.

Bears and bulls alike would do well to recognize this and to begin preparing for false misalignment events just as, if not more eagerly than for true ones. To be clear: I am aware that much is already being done to prepare for and preempt scenarios in which malicious humans use AI to some harmful end. But there is an important, underdiscussed subset of scenarios in which humans do this with the express aim of masking their role, and implying instead that a rogue AI agent is responsible. Nor need they act, as is often assumed, in pursuit of some ideologically motivated plot; they could just be after money, status, or attention.

This is the kind of scenario that the Moltbook debacle brings to mind. Far from the birth of the agent swarm, Moltbook is more akin to an undergrad digital humanities thesis that was lucky to go viral. But it was concerningly easy to sell it as something more. The lesson of its outsized impact is that AI discourse is fertile ground for a certain kind of malignant, reality-distorting virality. If the contemporary internet is at informational war, AI discourse is its Pacific theater. Only the Eastern Front of politics sees more grift, bias, and confusion. This has been lost on many AI watchers, who are often too busy envisaging the far future to take stock of the kind of present that would have to become it.

An essential feature of that present is the ubiquity of grift, what philosopher Daniel Rubio has convincingly argued is our new American frontier. Once statesmen, generals, and titans of industry, our agents of history are influencers, and they have no values except, like dogs, to chase tacky, golden cars. At least for now, this is the zeitgeist of AI, no less than politics, no less than culture. AI doom may be coming. But if it does, it may turn out more Trumpian than Yudkowskian. Its second step may be tragedy, but its first one will be farce.

The United States is in the midst of committing national suicide. There may still be time to save it.

A mix of economically illiterate trade policy and aimless geopolitical chest-thumping has precipitated a slide in the value of dollar-denominated assets. Donald Trump, a man with shamanic powers to bend social reality to his will, has summoned the specter of an end to America’s “exorbitant privilege.”

He has also begun liquifying America’s network of alliances, as well as its manufacturing capacity, a feat achieved in the latter case by simply declaring his desire to do the opposite, at which point the administration’s savviest economic minds proceeded to ask a chatbot how to revive early 20^th century protectionism. The resulting “Liberation Day” tariffs, announced via a large piece of cardboard, shook many on the right out of their polarized stupor, at which point it dawned that the November ballot they’d cast was akin to slipping on a giant banana peel in slow motion, the floor having finally greeted them with a piercing crack to the spine.

Unsatisfied to rest on its laurels, the Trump administration has now sicced an army of masked, gun-toting illiterates on the people of Minnesota. While it did so on the pretense of combating fraud and rounding up violent illegals, it has also given ICE agents de-facto permission to murder American citizens. ICE’s second victim—Alex Pretti—was shot to death yesterday morning.

As violent clashes kick off in the streets of Minneapolis, framed by history-echoing tensions between local and federal law enforcement, it might seem beside-the-point to wax philosophical about American national identity—especially as the very concept is still anathema to much of the left.

But if the murders of Renée Good and Alex Pretti end up a turning point, it will be because they blast through a thick crust of culture-war detritus to reveal longstanding cultural and institutional pillars that, decrepit as they are, were never really toppled, and so can still support the to-us obvious conclusion that it is wrong for agents of the state to gun down innocent civilians in broad daylight.

Those who claim “the moment has come” for a person, or a group to make some history-shaping decision are often proven incorrect. Real-world politics is less dramatic than this, at least most of the time. More often, it is a slow-rolling war of position, even and especially when glossed in climactic terms. Probably, Trump will continue to suffer in the polls as Americans consume large volumes of video evidence depicting blatant abuses by ICE. But memories are short—and shortening—and so, not unlike the assassination of Charlie Kirk, the killings of Good and Pretti may soon stream backward into the digital ether, their memories preserved by little besides gruesome AI slop generated en masse from their likenesses.

I want to believe that this time is different. Every person with eyes and a conscience can see that the killings of Good and Pretti are unacceptable. Evidence is mounting that they will have real political consequences. Americans’ moral intuitions are aligning.

Sometimes, it takes the awfulness of their violation to reveal that we do still have shared values in this country. One of the lessons I predict we will learn from the current period in American politics is that we can neither deny that this is case, nor try to resurrect those values by force on whatever are our social group’s preferred, probably revisionary terms. But that does not in turn mean we must embrace some preexisting set of ideals, lying dormant, to be unearthed a static whole. Whatever remains of American national identity will have, in some sense, to be rebuilt.

Masked thugs, acting at the behest of the state, are carrying out lawless executions in our streets. The Trump administration has mounted a multi-front war against nearly every spring of American prosperity. They are not us, and this is not who we are.

But in that case, who are we? Over the last decade, we have forgotten. But Minnesota is reminding us. Who are we? It may be that the moment to decide has finally, again, arrived.

Note: An edited version of this essay is now available here at Transformer.

1. Introduction

If you have been following A.I. research and development over the last couple of years, you likely will have encountered one or both of the figures below.

They come from non-profit A.I. research institute METR’s “Measuring AI Ability to Complete Long Tasks” benchmark, which has come to be seen as a critical bellwether of A.I. capability growth. The Long Tasks benchmark measures trends in A.I. models’ abilities to complete software engineering tasks by comparison to how long those tasks take humans.

The AI Futures Project (the team responsible for AI 2027) recently called METR’s Long Tasks benchmark “the best benchmark currently available for extrapolating to very capable AIs.” In his 2025 letter, Google DeepMind Engineer Zhengdong Wang describes METR’s benchmark as “many people’s first introduction to the case that AI will be a big deal.” Screenshots of what is often simply called the “the METR graph” have been posted and reposted to death on X by a veritable army of A.I. researchers, engineers, and businesspeople.

The subtext (or often simply the text) of these posts is that A.I. models are improving so rapidly that they may soon displace a large proportion of human knowledge-workers, not to mention pose a wide variety of serious, even existential risks to our wellbeing. Alternatively, they could usher in a utopia of economy-wide automation, freeing humans from the yoke of necessary labor, and bringing about a world of once-unimaginable material abundance.

While I am skeptical of these extreme scenarios, my goal here is not to convince you that they are overblown. Instead, it is to convince you that METR’s Long Tasks benchmark does not even remotely constitute evidence in their favor. This benchmark suffers from such severe methodological problems that it is a hair’s breadth from being totally useless. Its influence constitutes a concerning signal about the epistemics of the A.I. community broadly construed, and raises serious questions about the trustworthiness of the many non-profit research institutions (such as METR) that have sprung up in the wake of the A.I. boom. While I am not the first to criticize this benchmark, I hope I am the last.

2. METR knowingly overpublicizes its more misleading results

The y-axis of the METR graph plots software engineering task durations at which models succeed 50% of the time (see the two graphs above). These durations are calculated by fitting a logistic curve to the models’ (average) success rates (across 8 runs per model/task pair) as a function of humans’ completion times for the same tasks, and identifying where the curve crosses 50%.

METR has also published analogous figures replacing 50% with 80% success rates. Predictably, the results are much more modest (note the y-axis):

In METR’s public-facing communications, the 50% figures are always given clear priority. But consider that even 80% reliability is unacceptably low for the critical business functions many hope (or fear) will one day be completed by A.I. agents with minimal oversight. Even with oversight (suppose you are more interested in productivity gains in the here and now) 50% reliability is so low that agents this capable seem more liable to slow users down than to speed them up. It is hard for me to see what legitimate reason METR might have for prioritizing the 50% graph over the 80% one. Doing so makes the METR graph more difficult to interpret, and more likely to be overinterpreted.

That is even more so the case given that, as METR itself acknowledges, many of the tasks it tested models and human baseliners on are extremely unrealistic:

Real-world intellectual labor often involves messy details that benchmarks usually don’t include, such as being under-specified or poorly scoped, having unclear feedback loops or success criteria, or requiring coordination between multiple streams of work in real-time. We generally observed that agents struggle more on tasks that have these “messy” details (Section 5). (Kwa et. al. 2025, p. 14)

More comprehensively, the authors of METR’s Long Tasks preprint discuss five “systematic differences between [their] tasks and real tasks”:

• Automatic scoring All tasks we use are automatically scorable, meaning a piece of code running in the task environment determines the final score. This imposes constraints on e.g. the format of solutions, that tend to reduce task open-endedness, and the need for sensible value judgments.

• No interaction with other agents None of our tasks involve interacting with other autonomous agents. Coordinating with, or competing with, other agents seems likely to increase task difficulty. For instance, by increasing the importance of strategic decisionmaking, real-time coordination, and predicting the actions of other complex agents.

• Lax resource constraints None of our SWAA tasks, and few of our HCAST tasks saliently involve making efficient use of a limited resource—a common constraint in real-world tasks.

• Unpunishing Similarly, very few of our tasks are punishing of single mistakes. This is in part to reduce the expected cost of collecting human baselines. Real world tasks can often be more punishing, for instance, when they involve competing against other agents. For instance, a single blunder in a chess game can greatly reduce the chance of winning the game.

• Static environments Our tasks typically use environments that do not significantly change unless directly acted upon by the agent. In contrast, real tasks often occur in the context of a changing environment. (Kwa et. al. 2025, p. 19)

Beyond identifying these differences in an informal manner, the authors also break them down into 16 distinct “messiness factors,” (see p. 39-40 for a full list) and use them to assign “messiness scores” to all tasks that take humans a minute or longer to complete (i.e. all those in the HCAST and Re-Bench datasets). They then plot A.I. models’ performance on the messiest versus least messy 50% of tasks. This is a graph of models’ performance on the messiest 50% of tasks, a graph oddly located in the Long Tasks paper’s final appendix (p. 42):

You are reading that right. Not a single model topped a success rate of 30% for the messiest—that is, the most realistic—half of tasks. There is a strong argument to be made that this graph ought to be considered METR’s topline result; better yet they would produce updated versions of their influential time-horizon and doubling-time graphs using only tasks above the median messiness score. The authors openly acknowledge that these tasks are more representative of what software engineers actually do. Why publicize results with much less external validity if not for reasons orthogonal to being maximally informative?

All in all, I find METR’s research communications choices questionable. They consistently do more to publicize their flashier, more easily misinterpreted results despite conducting analysis implying they are misleading. I would think that, if truth were their priority, they would take a different approach.

3. METR’s human baselining methods are abysmal

These issues are practically typos, though, compared to the problems with METR’s human baselining methods.

Something I gather few people are aware of is that METR’s benchmark results are almost entirely driven by tasks from the “HCAST” dataset. These are software engineering tasks spanning ~1 min. to 30 hrs. in human completion time. Out of the 170 tasks in METR’s task-suite, 97 come from HCAST. The rest come from SWAA Suite, which contains 66 single-step tasks taking humans between 1 and 30 seconds, and RE-Bench which contains 7 complex A.I. R&D tasks taking humans at least 8 hrs. This means that the vast majority of the data points on the METR graph are HCAST tasks, since all but a couple of them fall into the ~1 min to ~5 hr. range.

How, one might wonder, did the team responsible for HCAST (almost entirely METR employees) find out how long the dataset’s underlying tasks took humans to complete? Before continuing, consider how important this question is to METR’s results. These are the times on the METR graph’s y-axis; if they are severely compromised, then so is the entire benchmark.

Keeping this in mind, here is a description of how the HCAST team collected their human baselines.

They recruited engineers with >3 years of experience “primarily .. via the professional network of METR employees” (Rein et. al. 2025, p. 6). Then they assigned them to complete, as far as I can tell, an average of ~4 software engineering tasks each; there were 140 baseliners and 563 task attempts (p. 8-9). Although, METR only ended up using 97 of HCAST’s 189 tasks (Kwa et. al. 2025, p. 5), and 286 (successful) baselines for the Long Tasks benchmark (Kwa et. al. 2025, p. 7). Presumably, this means they ended up with around 3 baselines per task (more realistically, easier tasks probably had more completions, and harder tasks fewer).

During task completion, baseliners had to commit their progress to a repository, and record their activity through Loom. They were also instructed to “verbalize their thoughts and processes aloud during work.” Recordings were used to ensure adherence to anti-cheating protocols (Rein et. al. 2025, p. 22).

Notably, baseliners were also paid “$50-$100 per hour, plus $25 - $150 per hour in performance bonuses for completing tasks quickly and successfully” (p. 6). One bonus, conditional on successful task completion, was half the receiving baseliner’s hourly rate multiplied by the average time spent on the task by all baseliners, and another, conditional on completing a given task faster than all other baseliners, was the baseliner’s full hourly rate multiplied by this same average (p. 23).

So, taking stock of just these details (there are more to come): task completion times were recorded from not just a clearly biased, but an utterly miniscule sample. As I hope we can agree on, you cannot generalize to all software engineers (much less all humans) the average task completion times from ~3 people per task all recruited from the same social network (and that 3 is generous, especially as you move up the y-axis, because some longer tasks likely only had one or two completions, and the same people would’ve had to complete multiple tasks; remember HCAST had only 140 baseliners in all).

This alone is very near to making the entire benchmark useless. I expect some will take this verdict to be excessively uncharitable, so I want to take a moment to defend it. The idea that the average time it takes ~3 engineers to complete a given task tells you even very roughly about the average for all engineers is false, full stop. You cannot infer anything about the latter from the former; the sample is simply too small, and too biased. Nor is it reasonable, assuming you are a software engineer, to simply gut-check whether the times seem right to you. This is unscientific in the extreme. You are one person. Neither you, nor your friends, nor people on X are representative of software engineers generally. Preventing these sorts of low-powered, highly-biased generalizations is why we have the sampling norms that we do in the empirical sciences.

I expect some of you will want to say something like: “well, even if the specific times are off, the METR graph still shows model capabilities doubling at a frightening rate.” No. Do not try this. If the specific times are wrong, so are their distribution. That is to say: were we to collect baselines from a much larger, more representative sample, some tasks might fall into different completion-time buckets. For instance, tasks METR found took 2-3 hours might turn out, in this larger sample, to take 1-1.5 hrs. But, unless all completion times shift downward by the exact same proportion, this would compress the overall distribution of task completion times. That would in turn change the rate at which A.I. models moved up METR’s y-axis. We have no way of ruling out this or similar possibilities, meaning that—given METR’s tiny sample—we cannot use their results to make general inferences as to how much better models are getting, or how fast. This is why I call the benchmark (nearly) useless: it is being sold as licensing inferences about A.I. models’ overall capability growth, and yet that is precisely what its meager sample means it cannot do.

Remarkably, its poor sampling is only one of several comparably serious problems with METR’s baselining methods. For instance, consider the likely effects of the HCAST team’s anti-cheating protocols and incentive scheme. Remember: baseliners were asked to verbalize their thoughts aloud while working—if you are a software engineer, consider whether doing this would interrupt or complicate your normal workflow. Much more importantly, though, they were literally paid by the hour—the longer they took to complete tasks, the more money they made.

Think for a second about how bizarre this is. If you are measuring the duration of a certain set of behaviors in a sample—and you want that measurement to have some external validity—the worst thing you can do is to pay your participants an hourly rate! It would be one thing if the target population were paid hourly for those behaviors ‘in the wild,’ but the vast majority of full-time software engineers are salaried.

I haven’t forgotten, to be clear, that the HCAST team also paid baseliners a bonus if they completed a task faster than all of their peers. This doesn’t dissolve my concern. Since the prospect of receiving this bonus was both uncertain and unlikely (baseliners wouldn’t have known how many other baseliners they were competing with), a hypothetical baseliner looking to maximize their returns is arguably going to go slower not faster. Going faster would mean lowering one’s predictable hourly return in pursuit of an unlikely and uncertain bonus. Not only that, but the average time spent per task was used as a multiplier for both bonuses, further magnifying the incentive to slow down. As a result, it at the very least seems likely that the HCAST team’s incentive system increased completion times relative to other possible incentive designs.

Continue digging through the HCAST and METR Long Tasks preprints, though, and the mere prospect of inflated baselines becomes a certainty. Given the course of my critique so far, you may be surprised to find that there is an entire section in the HCAST preprint titled “challenges with interpreting human baseline results.” Fascinatingly, not a single one of my criticisms so far makes an appearance! Here are a few important concerns that do, though (to the authors’ credit):

One fundamental limitation of our baselining process is that we measure the time tasks take for humans who have little to no outside context or prior experience on the tasks (though our tasks are also designed to require little additional context). We do this in part to avoid learning effects and in part due to practical challenges with acquiring sufficiently many high-quality human baseliners. This likely increases the amount of time baseliners spend on getting up to speed on the tasks (such as familiarizing themselves with the relevant tooling, reading documentation, and carefully understanding the instructions), compared to tasks people complete in their real work; this in turn may bias our time estimates for many of our tasks upwards …

Although we attempt to assign humans tasks aligned with their expertise, accurately matching baseliners to suitable tasks proves challenging and is occasionally imperfect. As a result, baseliners sometimes encounter tasks only loosely aligned with their specific expertise, potentially reducing their observed performance relative to typical professional work scenarios. In the real world, people mostly attempt tasks they and their employer have very high confidence they will be successful at. We also preferentially assigned tasks (particularly harder tasks) to the baseliners who seemed likely to perform best (from qualitative impressions of their qualifications and performance). This helps us obtain more successful baselines, but may have a systematic effect on the time estimates for easier versus harder tasks. (p. 14, emphasis my own)

So, on top of everything else, baselining engineers were regularly assigned tasks they had little prior experience with, in contradistinction to the sorts of tasks they would encounter in a typical work setting. To get a sense of how this might have impacted results, consider this pair of remarkable admissions from appendix B of the METR Long Tasks preprint:

We manually watched and annotated the recordings of 50 baselines to get a sense of how our baseliners spent their time. Based on this small survey of baselines:

1. On short (5-15 minute) tasks, a substantial fraction of the time (between 25–75%) is spent reading instructions or understanding how to submit correctly. Particularly for short tasks, our time estimates may be hard to interpret as there is some fixed overhead for reading the instructions, independent of the difficulty of actually performing the task per se.

2. Our tasks are designed to require relatively low amounts of context but base liners still spend over 25% of their time reminding themselves of relevant information (e.g. how to do port forwarding). (p. 29)

This is strong evidence that METR’s human baselines are inflated. And it is not even the only piece of evidence to that effect in this appendix. METR also quietly tested frontier models on issues from an internal repository, and compared their solution-times to those of baseliners and—critically—repository maintainers. This data is informative because it gives us a rough sense of how task completion times might differ between baseliners that were less versus more familiar with their assigned tasks. Arguably, it makes more sense to assess A.I. models by the latter since this would be more representative of work done by professional software engineers. In the words of the HCAST team “in the real world, people mostly attempt tasks they and their employer have very high confidence they will be successful at.” All this in mind, the following table is another very bad sign for the Long Tasks benchmark’s external validity:

“In practice, repo maintainers were 5-18x faster” than baseliners. At this point, we can essentially be certain that METR’s baselines are inflated, possibly by quite a lot. That would in turn entail that the model capabilities they report are inflated, since they are being evaluated by comparison to these baselines. It may be, for instance, that Opus 4.5 can (sometimes) complete a task that would take a typical HCAST baseliner five hours, but if that task would take an engineer with more relevant expertise only one hour, reporting the five hour figure as how long it takes “humans” to complete the task will mislead us as to Opus 4.5’s capacity to substitute for software engineering labor.

The human baselines used for METR’s Long Tasks benchmark are awful, end of story. They are based on an extremely small, non-representative sample. They are confounded by a bizarre incentive scheme in which participants were paid according to how much they increased the key value under study (i.e. task completion time). And to boot, baseliners had far less task-relevant expertise than would professional software engineers completing similar tasks at work. There is far too much noise for any genuine signal to break through here.

4. Conclusion

Regardless of the litany of serious issues with the METR graph, I imagine some of you will still feel I am being harsh. But human baselines are very complex and expensive to collect you may say; but it was admirable of METR to review so many limitations of their work, to conduct deep analysis impugning the realism of their very own task suite.

All of this is true. But I have a very hard time seeing posts like the following, from METR’s initial X thread describing the Long Tasks benchmark, and feeling much sympathy for them:

Considering my criticisms above, this phrasing is ridiculous. It is an uncontroversial example of misleading science communication (and no, nothing in the rest of the thread makes it better). This, from the more detailed blog post on their website, is barely an improvement:

Summary: We propose measuring AI performance in terms of the length of tasks AI agents can complete. We show that this metric has been consistently exponentially increasing over the past 6 years, with a doubling time of around 7 months. Extrapolating this trend predicts that, in under a decade, we will see AI agents that can independently complete a large fraction of software tasks that currently take humans days or weeks.

METR is in no way measuring the “length of tasks AI agents can complete.” It is measuring whether A.I. can occasionally complete 97 highly contrived software engineering tasks whose “lengths” are spuriously determined. Nor does “extrapolating this trend” predict anything that can be understood in terms of what “humans,” as such, can do. You do not get to generalize to the entire species from a few of your LinkedIn connections. You especially do not get to do that if you claim to meaningfully measure the time it takes them to do something while also paying them by the hour.

This is all very bad, no matter how charitable one is inclined it to be. It is not the sort of intellectual impropriety that deserves to be handled with kid gloves.

The A.I. boom has inspired a wealth of careful research by well-meaning people who are sincerely trying to understand the world-shaking changes happening under our feet. It has also inspired troves of lazy and dishonest grifters trying to ride the hype-wave to fame and fortune. It would be unfair to accuse the authors of the METR graph of being part of the second group. I remain sincerely uncertain, though, whether they are part of the first.

Over the past month, the discourse on this platform about continental philosophy has been staring me down from the corner like Willem Dafoe’s ‘Green Goblin' mask in the first Spider-Man movie. On one level, I find the whole thing exasperating—these paths have been trod numerous times before—but on the other, I think this debate has and continues to mask interesting questions about philosophical method, ones we’ve made little progress resolving since the beginnings of the continental-analytic split in the early 20^th century.

I also suspect, at the risk of sounding self-important, that I have something valuable to add here. I’ve said little about my own intellectual background on this platform as of yet, but engaging with this specific debate demands a bit of autobiography.

I got interested in philosophy as an undergraduate largely as a result of encounters with continental philosophy. In particular, a former debate coach recommended Baudrillard’s Simulacra and Simulation, and I later encountered figures like Marx, Foucault, Derrida, and Deleuze in an introductory literary theory course. I soon became almost obsessively interested in the Western Marxist and poststructuralist traditions, and ended up doing a thesis on the relationship between Derrida and Adorno’s related but differing perspectives on the notions of “presence” and “identity” (you can find it online with relative ease, but I can assure you it is very bad).

During my junior year, I decided I wanted to go to graduate school for philosophy, but was dismayed to find that there were very few programs focused on the continental tradition. The few that were, I found out, had largely moved on from the particular debates I’d become interested in anyway, so there was virtually nowhere I could go to continue pursuing my specific set of interests.

On top of that, I was far from impervious to prevailing hierarchies of prestige in the discipline, and since I wanted to go to a “good” and even a “top” program, resigned myself to the fact I would have to pivot in the direction of analytic philosophy as currently practiced in major American universities.

To make a long story short, despite reading as aggressively as I could in preparation for applying (I graduated in 2020 just as Covid took off, and so had a lot of down-time), by the time app season rolled around, I was still hopelessly naïve about the prerequisites for gaining admission to top philosophy programs. I got rejected from every PhD program I applied to, but was extremely fortunate to be accepted with a generous scholarship into the philosophy MA program at Tufts.

There, and again skipping a lot of extraneous detail, I began to specialize in (analytic) political philosophy and metaphysics, with some sub-areas of special concern being, on the political side, distributive and ‘structural’ injustice, and on the metaphysics side, grounding, causation, and social ontology. After some two and a half intense, but intellectually generative years, plus a couple more unsuccessful rounds of PhD apps, I pivoted out of philosophy academia entirely and onto what have turned out for me to be greener pastures.

So, all of that is to say, I feel I have some unique insight to offer when it comes to the ‘continental vs. analytic’ debate. While plenty of people have engaged in deep ways with both sides, most of them have done so from a particular side, and often only with what they take to be the least offensive slice of what their counterparts have to offer (or, as we will see, snippets from or caricatures of the most offensive slice).

While I’m sure I’m not the only one, I am at minimum part of a very small group who has engaged broadly and deeply with the offerings of both camps—and in particular with what we might think of each’s most ‘egregious offenders,’ i.e. with the most ‘continental’ of the continental (Hegel, Derrida, Butler, etc.), and the most ‘analytic’ of the analytic (cloistered debates in metaphysics, philosophy of language, and so on). I have also done so with approximately equal intensity (in effect as my full-time job), and over approximately equal time periods (~2-3 years each).

The reason I feel the need to frontload my credentials like this is because the ‘continental vs. analytic’ debate attracts a lot of folks who posture as having the background necessary to approach it fairly, but in reality are just purebred analytic, or purebred continental philosophers whose understanding of the other camp is minimal to nonexistent. In fact, that describes the majority of those who have ever seriously tackled the issue. As a result, I feel it’s important to emphasize that I have a fair bit of expertise in both areas, and also have no special allegiance to either (I’m much closer to an analytic philosopher by training, but still have a soft spot for my continental roots).

This background on hand (and with apologies for the lengthy intro), we can proceed to discussing the most recent spark in the long history of explosive debates over the continental-analytic divide. That would be this essay by popular philosophy blogger ‘Bentham’s Bulldog,’ real name Matthew Adelstein.

Adelstein’s essay is typical of lazy critiques of continental philosophy. It consists in

a) sharing block quotes from prominent continental philosophers in the hope that they will read as preposterously unclear and confusing, and

b) running through a series of putatively typical and obviously invalid “inferences rules” of continental philosophy, cribbed from other writing by continental philosophers.

Let’s rebut each section in turn.

The first is easily dealt with because, as vexing to read as Adelstein’s chosen quotes might be to the uninitiated, they are easily parseable by those familiar with the concepts and debates they are referencing. Let’s look at his first quote from Judith Butler:

The move from a structuralist account in which capital is understood to structure social relations in relatively homologous ways to a view of hegemony in which power relations are subject to repetition, convergence, and rearticulation brought the question of temporality into the thinking of structure, and marked a shift from a form of Althusserian theory that takes structural totalities as theoretical objects to one in which the insights into the contingent possibility of structure inaugurate a renewed conception of hegemony as bound up with the contingent sites and strategies of the rearticulation of power.

Some prerequisites for understanding this passage in depth include: the distinction between structuralism and post-structuralism, the Marxist meaning of “capital,” Louis Althusser’s concept of “structural totalities,” Antonio Gramsci’s concept of “hegemony”, as well as Ernesto Laclau and Chantal Mouffe’s adaptation of this last notion in their 1985 book Hegemony and Socialist Strategy, which is the main subject of this paragraph, as is clear if you read it in context.

I am skeptical Adelstein (or anyone who nodded along with his essay) did anything approaching their due diligence on this background, before dismissing what is a complex, but far from incoherent claim from a renowned expert in their field. And so I can’t be accused of appealing to authority or gatekeeping this background myself, here is an attempt to render the Butler passage in a form (somewhat) more congenial to analytic philosophers and other non-specialist readers:

On “structuralist” theories of capitalism, ‘capital’—the process by which capitalists exploit human labor in order to produce commodities for profitable sale with the aim of reinvesting those profits in further commodity production—is understood to shape the social relations of those living under capitalism in a uniform manner, i.e. by way of that economic system’s most stable macroscopic properties. Conversely, Laclau and Mouffe extend Antonio Gramsci’s concept of ‘hegemony’—roughly, the idea that capitalists manufacture consent for their preferred form of social organization through the extra-economic pursuit of political and cultural power—to argue that the macroscopic properties of capitalism must be reproduced over time via micro and meso-level phenomena (such as fights over policy in national legislatures, public relations efforts by key political actors, distributional conflicts between bosses and unions, and so on), phenomena characterized by a far greater degree of dynamism and political contestation than the more stable features of capitalism writ large. The upshot is that, while we still ought to think of capitalism as having a holistic structure, Laclau and Mouffe’s renewed account of Gramscian hegemony has shown that we need to understand that structure as encompassing phenomena that a) exist at a range of metaphysical ‘levels,’ and b) are always undergoing rapid change, change whose precise characteristics are politically contested.

What do you notice about my revised version of this passage? It is more than double the length of the original, and is chock-full of lengthy appositives and interjections. So too would be analogous attempts to translate complicated, jargon-filled prose from other fields. There is nothing special about continental philosophy in this regard, nor Butler’s passage, nor for that matter Adelstein’s dismissive attitude. It is the same attitude held by religious fanatics who dismiss research in evolutionary biology because of its, to them, baffling appearance. My point is not that (all) continental philosophy has the same epistemic standing as (all) evolutionary biology, but that Adelstein’s “objection” to the former is no better-formed than the religious fanatic’s to the latter. He does not understand the passage, and he thinks it is self-evident that no one could, or does. That is false, and it is false of the next two passages he cites as well.1

Adelstein’s next line of objection is that continental philosophers make arguments based on defective rules of inference. In his words, “they’ll start by saying some things, and have orderly norms governing the relationship between premises and conclusions. The only problem is that the norms don’t involve the use of either logic or any other sound inference rules.”

His first example of a “continental philosophy rule of inference” is “’A is not B’ (where B is the obvious thing everyone would expect it to be) ‘but instead C’ where C is some random thing that makes no sense.” He takes this passage, again by Butler, to exemplify this rule:

As a result, gender is not to culture as sex is to nature; gender is also the discursive/cultural means by which “sexed nature” or “a natural sex” is produced and established as “prediscursive,” prior to culture, a politically neutral surface on which culture acts.

Translating into a less ‘continental’ register:

The concept of gender is upstream, rather than downstream, from the concept of sex, understood to refer to an objective, biological reality. Whereas we take sex to exist mind-independently, the precise semantic content of this term is strongly dependent on our conventional understanding of gender. In fact, it is because of our “gender” concept that “sex” is conceived of as mind-independent in the first place.

Regardless of what you think of this passage’s substance, we should be able to agree it does not contain an inference on any plausible theory of inference. There is no deduction, induction, or abduction here. At most there is a series of positive claims that are related, but only in the sense that they elaborate on one another. They do not, however, entail one another, nor is there any evidence that Butler thinks they do. Adelstein is conflating a common rhetorical pattern (“A is not B, but C”) with a “rule of inference,” and suggesting that it is continental philosophers, rather than he who have confused the two.

But no continental philosopher would characterize this passage as involving an inference, nor do any subscribe to the silly “inference rule” Adelstein attributes to them. So this is just a lazy straw man. Even more egregious is Adelstein’s next example in which he claims that the following relies on the “inference rule,” “for A, B.”

For Beauvoir, women are the negative of men, the lack against which masculine identity differentiates itself; for Irigaray, that particular dialectic constitutes a system that excludes an entirely different economy of signification.

This is plainly a brief piece of exegesis contrasting Beauvoir’s views to Irigaray’s. I have no idea how Adelstein could believe it to be an “inference” that Butler believes amounts to a “decisive argument.” We’ve entered the realm of either serious confusion or deliberate obfuscation.

I’ve only covered a few of Adelstein’s examples, but suffice it to say those I’ve left out are no better than those above. Bizarrely, they include multiple passages generated by ChatGPT, one of which is only identified as such in a footnote.

I find Adelstein’s work pretty strong overall, and so have a hard time believing he sees these as genuinely well-formed objections. Instead, I suspect he simply wasn’t interested in doing any real due diligence in this case, and further relaxed his standards for rhetorical effect, in particular given continental philosophy is already unpopular among his readership. His essay doesn’t wear its identity on its sleeve, but squint and it quickly reveals itself as the sort of low-effort out-group-bashing typical of the contemporary internet.2 There is plenty to criticize continental philosophers for, but this is not the way to do it.

This is the first in a series of posts I intend to write on the continental / analytic divide. In the next, I plan to discuss some of the other responses to Adelstein, in particular those by Ellie Anderson and Sam Kriss, which, while much better than Adelstein’s initial post, I also found unsatisfying and worthy of a response. I may also discuss posts by Michael Huemer and Daniel Muñoz which defend aspects of Adelstein’s position. After that, I will try to outline in some detail my own thoughts about the differences between continental and analytic philosophy.

Everyone is talking about “slop.” But what is slop really? Here is a tentative answer, in the form of a working definition:

Slop: media designed to deliver frictionless and amusing short-term stimulus, rather than to inspire more frictional forms of emotion and/or reflection.

“Friction” here is meant as a more efficient term for the subjective experience of mental discomfort brought on by thoughts that are inconsistent, confusing, or complex.

Though not a necessary or sufficient condition, inspiring discomfort of this kind is a common property of good art, whether visual, written, or otherwise. Art often has something to teach us, and good art convinces us to ingest its teachings, even at the cost of some mental strain.

Slop is media designed (intentionally or not) to do the exact opposite, to suppress or smooth out aesthetic and cognitive tension of exactly the sort that demands we reflect. It is media optimized to hold our attention and to do precisely nothing with it to the greatest extent possible.

Absent any serious inhibitions to our agency or autonomy, most of us are choosing to consume slop every day, often for multiple hours at a time. This would seem an unattractive choice from a disinterested, rational perspective. It would seem even more unattractive, arguably, than a similar, if more science-fictional choice known for its unattractiveness: to enter ‘the experience machine’ from philosopher Robert Nozick’s famous 1974 thought experiment.

“Suppose,” Nozick writes, in his book Anarchy, State, and Utopia, “there were an experience machine that would give you any experience you desired.”

Superduper neuropsychologists could stimulate your brain so that you would think and feel you were writing a great novel, or making a friend, or reading an interesting book. All the time you would be floating in a tank, with electrodes attached to your brain. Should you plug into this machine for life, preprogramming your life’s experiences? … What else can matter to us, other than how our lives feel from the inside? (p. 42-43)

Nozick’s, and indeed most moral philosophers’ assumption has long been that few if any would choose to enter the experience machine (“we learn that something matters to us in addition to experience by imagining an experience machine and then realizing that we would not use it,” p. 44).

The thought experiment is meant as an argument against ‘hedonism about well-being,’ the idea that our lives are good or bad purely in virtue of our experiences of pleasure and pain. Were hedonism about well-being true, we’d be fools not to enter the experience machine, since it allows us to max out on pleasurable experiences (and to ‘min out’ on painful ones). But given most have strong intuitions against entering the machine (or so many philosophers have claimed), hedonism must be false by modus tollens. There is something disconcerting about leaving ‘the real world’ behind, even for a vastly more pleasant false one.

It is striking that without even the promise of moderate pleasure, much less unrealistically intense pleasure, most of us now enter our own sort of experience machine for, on average, several hours a day, hours that add up over the course of our lifetimes, until, by the end, they’ve turned into years. Reframed as its own thought experiment, this more realistic machine is somehow both more subtle and more off-putting than Nozick’s original:

The Inexperience Machine: you may enter for several hours daily a machine that displays, on a screen inside, a fluctuating menagerie of audiovisual experiences ranging from the trifling to the grotesque. Upon exiting the machine, you will forget all but a narrow slice of what you heard and saw. To enter is to commit these hours to the machine every day for the rest of your life.

Between the experience and inexperience machines, I would argue that the former is the more attractive choice. The experience machine at least trades reality on the whole for a greatly more pleasurable alternative. But the inexperience machine does little besides cleave apart and empty out the one reality that we have.

That so many of us have entered the inexperience machine—not in a thought experiment, but for real—would come as a shock to the late 20^th century ethicists (including Nozick himself) who intuited that Nozick’s machine would so offend our attachment to the real that even the promise of experiencing, say, an impossibly passionate love affair, or winning every Nobel prize at once, could not seduce us to enter.

Our new digitally suffused unreality is a dramatic object proof that they were wrong. We are in effect making the same trade with a large proportion of our time, while getting much less in return. While it is difficult to say how much of what we consume counts as “slop,” U.S. adults now spend an average of 8 hours a day on video consumption alone.

It is not just the raw time commitment that is galling here, but also how it is stratified. Raw averages mask stark inequalities in the video consumption of different groups. A privileged few, for instance, still consume little digital media at all. And among those with more typical diets, some are much more modest consumers than others.

The education nonprofit Common Sense’s 2021 report on “Media Use by Tweens and Teens” found that children of families making $100,000 or more a year spent 2-3 hours less on screens than those of families making less than $35,000 a year, and 1-2 hours less than those with yearly incomes in the middle.

Similarly, while it is widely discussed that increased digital media consumption over the last 15 years has coincided with declining literacy and numeracy scores across age groups, it is less discussed that these declines have not been equally distributed.

In its 2023 Adult Skills report, the OECD found that declines in average literacy and numeracy over the previous decade were “largely due to falls among the lowest performing adults” (p. 22 - 23). International PIAAC test results bear this out, and the U.S.’s domestic results on both the PIAAC and the NAEP show a similar pattern.

Across all these figures, so-called “brainrot” has hit the lowest performers the hardest (yes, I am eliding some complex causal questions here, but the evidence that digital media consumption explains these trends gets stronger by the day). While the high-achieving have not been immune, they have better resisted some combination of the inexperience machine’s allure, and its stultifying effects.

Others haven’t fared so well, though, and even for high-achievers, things could very well get worse. For context, the data above largely precede the rise of so-called “AI slop.” Generative AI has made it much easier to create content optimally designed for shallow attention-seeking. The result has been a Biblical flood of often bizarre and disturbing content (see below for some classic entries in the genre; the first has 166 million views, the second 414 million).

Putting aside that videos like these are currently traumatizing unsuspecting children on an industrial scale, more, and more addicting content probably also means more brainrot. If so, then the educational and intellectual hierarchies reflected above are likely to keep growing.

They will do so even if sturdier social barricades soon stem slop’s rising tide. Much of the damage has already been done, and it will continue to compound as its young victims age into cognitively impaired adults. That was the conclusion of a little-noticed 2022 review article in the Journal of Integrative Neuroscience. “We hypothesize,” write the authors, “that excessive screen exposure during critical periods of development in Generation Z will lead to mild cognitive impairments in early to middle adulthood resulting in substantially increased rates of early onset dementia in later adulthood.” They go on to predict that

from 2060 to 2100, the rates of Alzheimer’s disease and related dementias (ADRD) will increase significantly, far above the Centres for Disease Control (CDC) projected estimates of a two-fold increase, to upwards of a four-to-six-fold increase … [enough to] result in widespread societal and economic distress and the complete collapse of already overburdened healthcare systems in developed countries. (pg. 1)

While some have and will resist the inexperience machine, opting instead to use the wider digital ecosystem the way once-vocal cyber-idealists assumed they would—to learn, to connect, to find like-minded people—far more are set to whittle their minds away drinking long and deep from the slop trough.

It doesn’t take a sage to see the future that this portends: one of aggressive, and fast-rising cognitive inequality, borne not of hardship or oppression, but a garish digital rollercoaster even the smallest children are tall enough to ride. While rising wealth inequality has long been a fixture of our political economy, rising cognitive inequality may soon rival it for structural significance.

Nozick was wrong to assume that no one would want to enter the experience machine. He thought that “we want to do certain things, and not just have the experience of doing them” (p. 43). But humans are rarely moved by abstract philosophical reasoning. More often, they are moved by a mix of biological drives and low-grade psychological heuristics. Hence the spate of experiments finding that, when people are asked if they would enter the experience machine, their answers hinge on trivial aspects of how the question is posed.

Rats will give up food, to the point of starvation, for successive hits of electrodes attached to their neural reward centers. We are more similar to them than the ideally rational agents of much analytic philosophy. We are climbing into the inexperience machine in droves, choosing streamers over friends, cooking videos over cooking, and pornography over sex. We are choosing slop, which goes down easy, over edifying friction.

The latter may soon become the exclusive provenance of a new cognitive elite. There would then be little to stymie the U.S.’s ascendant techno-authoritarianism. Sooner than we’d like to think, we may be less participants in politics than points on a scoreboard, jockeying between whichever faction has the best short-form propaganda. What was once a republic may soon be reduced to a field of manipulable atoms bouncing around digital space, our behavioral parameters wiggling at the whims of the highest bidder.

Resisting this outcome means reintroducing friction into our lives, the friction of seeing and meeting strangers, of learning, of holding strong to our civic commitments. The crisis of slop consumption is of a piece with a more general crisis of risk aversion, one shaping everything from early parenting to post-college career development to socializing among adults—a kind of cradle-to-grave program for turning out pampered neurotics. The rise of always-at-hand digital entertainment is an essential part of this program, since it means we can always trade something challenging, but good for us, for something both comfortable and pointless.

Nozick’s opposition to the experience machine was so strong that he considered plugging into it to be a “a kind of suicide” (p. 43). He was anticipating the ideas of someone who would at first seem a strange bedfellow, namely the author David Foster Wallace, who had this to say—in 1996 no less—about the future of electronic entertainment:

… the technology is just gonna get better and better and better and better. And it’s gonna get easier and easier, and more and more convenient, and more and more pleasurable, to be alone with images on a screen, given to us by people who do not love us but want our money. Which is all right. In low doses, right? But if that’s the basic main staple of your diet, you’re gonna die. In a meaningful way, you’re going to die. (Lipsky, p. 86).

To the extent you agree with Nozick and Wallace, the choice to leave the inexperience machine becomes painfully simple: life is better spent living than dying by a thousand virtual cuts.

Dan Williams’s new essay, “Is Social Media Destroying Democracy—Or Giving It To Us Good And Hard?” is characteristically sharp and well-argued. I was going to post a short note giving my initial reactions, but that note quickly grew in length and detail, to the point that I decided it was worth a full post. I reserve the right to follow up with an even longer, more detailed post later on.

One minor point is that I think Dan is overstating my position a bit when he says in the ‘Further Reading’ section of his new essay that I think he’s ‘very wrong’ about social media. As with his previous work, I found lots to agree with in this recent piece. That said, no one will be surprised to hear that I still believe he’s understating the harmful role of social media per se, as opposed to its broader democratizing effects.

A brief thought experiment will help to explain why. Imagine a world in which ordinary people had commensurate influence over public discourse, but without social media. Perhaps there are more frequent, larger, better publicized town halls open to all. Maybe there are also cheap, widely accessible T.V. shows hosted by ordinary citizens, ones that compete effectively with shows hosted by media elites. Would this world experience a profusion of sensationalist / extreme / violent rhetoric similar to what we’ve seen post-the rise of social media?

If so, that would speak in favor of Dan’s argument that the recent rise of right-wing populism and political extremism more broadly is the consequence of a more democratic public square. But if not, it would speak in favor of mine, that social media is the problem, not a broadly right-wing, populist public with a louder voice than before.

There are several reasons I suspect we would not see a similar profusion of extreme politics in this alternative world:

a) alternative means of public influence besides social media don’t allow speakers to be anonymous, meaning they would have to worry more about the reputational effects of what they say, discouraging extreme rhetoric;

b) absent social media, speakers (or ‘content creators’) would not be paid in direct correspondence to the amount of attention they garner;

c) nor would that attention be tracked in such extensive detail, making it more difficult to optimize for;

d) and this world would also lack massive engagement-hungry corporations systematically amplifying the most attention-grabbing views (which tend to also be the most sensationalized and extreme ones).

And that’s just a start. I concede I’ve only made a few suggestive comments about the nature of this alternative world, so it’s hard to say exactly what it would be like. Maybe if ordinary people routinely hosted their own television shows, they would still disseminate extreme rhetoric for the purpose of maximizing engagement. I’m skeptical they would, though, without strong incentives of the kind social media creates.

Even if those incentives cropped up naturally in this alternative world, I don’t see that as a problem for my argument. Strictly speaking, my complaint has always been with the specific design features of actually-existing-social media—design features that might have similar harms applied to different contexts—not with all possible variants of social media, some of which I’m sure would be far more benign, or even beneficial. The way this distinction plays into my argument with Dan is that it suggests it would be helpful for both of us to sharpen what we mean by “social media.” I think it’s probably the case that we both support the existence of something social media-like, but would prefer it be governed differently, likely in a way that involves a greater degree of gatekeeping (I have a post in the works about what this gatekeeping should look like).

I think our real disagreement comes down to the proportion of responsibility for our current political malaise owed, on the one hand, to the objectionable design features of actually-existing-social media (principally those that incentivize users to optimize for attention above all else), and the bare fact of it increasing non-elites’ influence over public discourse. Very roughly, I’d put the ratio of responsibility here at something like 70/30 (favoring design), whereas I imagine Dan would put it at something like 30/70 (favoring non-elite influence).

But this binary leaves something crucially important out. As I argue in “The Case Against Social Media is Stronger than You Think,” it is principally elites, not non-elites, that are responsible for the extreme political rhetoric that has exploded on social media. Yes, smaller accounts make extreme posts too, and sometimes those posts go viral. And yes, they make up the followings of more influential actors like Tucker Carlson and Candace Owens, and transmit those bad actors’ views to others. But it is the Tucker Carlsons and Candace Owenses themselves that are mainly responsible for social media’s inflammatory impact. It is the elite political influencers who are to blame, not (the vast majority of) average social media users.

This is a problem for Dan’s argument, as well as the broader argument that social media has had a democratizing effect on the public sphere. While it surely has at some level, its more consequential effects flow from the elite political influencers it has empowered, only a very small portion of which weren’t elites already, before they gained access to social media.

This suggests a very different conclusion than Dan’s: social media, far from devolving public influence back to the masses, has instead shifted public influence between different groups of elites. In particular, it has favored (in relative terms) those who have for various reasons been denied access to legacy media, perhaps because they have fewer qualms than the average journalist or politician about stoking fear, hatred, and alarm, not to mention outright lying, but also very possibly because of excessive liberal bias in legacy institutions.

That is not to deny that social media has had a democratizing effect. But it is not the same one defenders of the technology have emphasized. It is a stretch to claim social media has given typical members of the public substantially more voice. What it has done, however, is provided a platform for them to more readily amplify others’ voices, namely elite political influencers’, most of them on the populist right. So yes, in that narrower sense, social media has had a democratizing effect: it has unleashed latent public demand for media produced by a different subclass of elites than dominated the pre-social media landscape. But that is very different from democratizing the public sphere as such.

There are two things I want to highlight about this alternative story of social media-induced “democratization.” First, it contradicts Dan’s claim that, until the rise of the internet and social media “information production remained primarily in the hands of wealthy, well-connected elites.” Except at the margins, this is still true. Elites still largely control information production, even if different kinds of elites than before are competing for narrative dominance.

Second, this alternative story has very different implications for the democratic credentials of prospective defenses of digital gatekeeping. It is one thing to prevent typical members of the public from participating in major public fora. It is another thing—arguably a more permissible thing—to limit certain elites’ ability to do so, especially if we have ways of ensuring those restrictions only apply to them; one example of what that might look like is for larger accounts to face stricter content restrictions than smaller ones.

Still, I’m aware that this is an extremely controversial, easily misapplied idea. It is an open question to what extent it is consistent with the values of a healthy democracy to restrict the reach of certain extreme, even anti-democratic views proffered by the rich and powerful. But it is, at least, the right question to ask. Compare it to an analogous question about the public writ large, in particular one that assumes social media has amplified all our voices equally: should we block the masses’ access to social media, given the dangerous new world their voices have wrought? I hope we can all agree that this question is the wrong one, on at least two different fronts.

Edit: I received a helpful comment from (who is an economist at the University of Queensland) that gave me occasion to clarify / expand on some of the arguments above. Thought the exchange was worth reposting in the body of the article (and will update with any future comments from Page / myself).

Lionel Page: Hi Nathan, thanks for pointing me to your post. I don’t think that your two perspectives are really different. Whether the counter-discourse is made by elites or not, it is targeted at the popular preferences that were previously under-represented. I had written the following paragraph for a future post, prior to Dan’s post: “Social media alter the democratic game. They create opportunities for political entrepreneurs to vie for new coalitions and challenge existing ones. They also give citizens a way to seek and supply public justifications for different positions without the intermediation of knowledge institutions that might overly represent elite views. As a result, social media make the marketplace of ideas more contestable.” (Here I refer to the economic notion of contestability for markets.)

Nathan Witkin: Thanks for the thoughtful comment! I think the last passage is mostly spot-on, but would say it omits some considerations that help to bring out where I see Dan and I differing.

One I would underline is the question of to what extent popular preferences are endogenous to the marketplace of ideas. I see online ‘political entrepreneurs’ as powerful shapers of public opinion, so I think it’s misleading to frame our new digital political culture as the product of unalloyed, bottom-up democracy. I still agree with the overall thrust of the idea that social media allows people to seek out perspectives considered beyond the pale in mainstream media, but I think we’re missing a key part of the story if we treat elite political influencers as simply filling preexisting demand, rather than at some level helping to create it.

Some evidence / elaboration on this claim from a previous essay:

“There is strong causal evidence that political influencers have the power not just to reflect, but to shape the attitudes of their followers. In a recent preprint, Rathje et. al. (2025) found in two different field experiments (n = 494, n = 1,133) that unfollowing “hyperpartisan social media influencers” on Twitter (they give the examples of Palmer Report on the left and Breitbart News on the right) “improved [participants’] recent feelings toward the out-party by 23.5% compared to the control group, with effects persisting for at least six months,” (p. 2). This should put to rest the idea that inflammatory political content is a purely ‘demand-side‘ phenomenon. Yes, this demand has always existed, but it is possible for the attitudinal effects of existing supply to increase it even further.”

At a more philosophical level, I am hesitant to accept that social media is “democratizing” the public sphere at all. Given Dan embraces this rhetoric in his essay, I’d say this is also a point where we differ. It’s very much still an unmined vein for theoretical exploration whether enlarging the marketplace of ideas is intrinsically ‘more democratic.’ How do we make sense of this claim if the ideas proffered by new entrants are grossly undemocratic in substance? What does it mean for democracy writ large if the marketplace of ideas becomes more procedurally democratic while becoming less substantively democratic? Is this even a well-formed distinction? We’re still a long way from having hashed these questions out.

Lionel Page: Thanks, Nathan. I think the two dynamics (political entrepreneurs creating content and masses demanding content) are two symbiotic things. Entrepreneurs try to create content that has the maximum chance of success. They innovate, but innovations will work only to the extent that they meet latent preferences in the population. For that reason, entrepreneurs will most often offer the population things they want (e.g. Jobs creating iPods and iPhones). Could they sometimes lead the population astray with bad innovations? Sure, appealing but bad products can be offered, and the same can happen in the marketplace of ideas. It does happen (not just on the counter-elite/populist side, though).

Note: I am in the process of getting set up at a new job, and so will have less time to write over the next few weeks. In the meantime, I thought I’d repost a lightly edited version of the section on social media and mental health from my previous essay, “The Leviathan, the Hand, and the Maelstrom.” While a lot has been written on this issue, I am aware of no article besides this one that gathers much of the strongest research on social media’s mental health harms in one place for easy reference. Haidt, Twenge, and Rausch’s sprawling ‘collaborative review’ is a partial exception, but collects much more evidence of varying quality and relevance (ditto for other academic reviews). I hope this short essay can act as a more curated and accessible supplement.

In the aftermath of the historic rise of social media, it is easier than ever to make what social scientists call ‘unfavorable,’ or ‘upward’ social comparisons. Online, we are beleaguered by ostentatious displays of wealth, status, and power. It is only natural that we humans, who thrive and falter on social hierarchy, would react with class-conscious anger, anger directed not just outward, but also inward, at ourselves.

Hence the recent uprising in Nepal, which was led by members of Gen Z, and motivated in large part by online content showing off the opulent lifestyles of the country’s so-called “nepo-kids,” an unsurprising firestarter in a country with a per capita GDP of around 1,450 USD.

Not everyone reacts to the glossy irrealism of the digital world, though, with anger and contempt. Some, and in particular those yet to graduate from the angst of youth into the brashness of adulthood, react instead with self-loathing and hopelessness—that they’re not the person they should be, that they can never hope to compare to what they’re seeing on their screens. This explains why, since the rise of social media, an unprecedented amount of adolescents have fallen into mental health crisis.

Despite remaining more or less constant through the aughts, mental illness among young people in the U.S. began to sharply increase starting around 2010. To give a small, but grim sampling from Jonathan Haidt’s popular book The Anxious Generation, from 2010 through 2020, depression increased by 106% among college students, anxiety by 139% and 103%, respectively, among those ages 18-25 and 26-34, and suicide rates by 167% and 91%, respectively, among girls and boys aged 10-14 (p. 26-31).

There is of course a lively debate as to whether these developments can be attributed to social media. Critics often cite large meta-analyses which they claim show the existing evidence to be equivocal or inconclusive. For instance, Candice Odgers, a prominent critic of the alleged link between social media and youth mental health claims, in her critical review of Haidt’s book, that efforts by researchers to search for the “large effects suggested by Haidt … have produced a mix of no, small and mixed associations” (p. 29). Odgers’s phrasing would seem to suggest a blend of statistically significant and insignificant effect sizes with both positive and negative signs.

But even a basic review of the literature shows that suggestion to be inaccurate—and to be frank, Odgers should know better. Of the 4 meta-analyses and review articles cited by Odgers in her review of Haidt, every single one, but her own openly acknowledges that the predominant result in the correlational research under review, with respect to adolescent anxiety and/or depression in particular, is a small, but statistically significant association of about r = .1, in other words a persistent, although modest association between social media use and mental illness. Several similarly comprehensive meta-analyses Odgers does not cite all find the exact same thing. The most recent of these—Liu et. al. (2022)—reviewed 21 cross-sectional and five longitudinal studies covering 55,340 participants and found that “the risk of depression increased by 13% for each hour increase in social media use in adolescents” (p. 1). The notion that the correlational research is equivocal or inconclusive on this point is flat-out false (you can read Haidt’s own rebuttal to Odgers here.)

This doesn’t prove much in itself. Those correlations may reflect causality running in the other direction—that more depressed or anxious kids use more social media, rather than the other way around. But, at a minimum, they mean that the hypothesis of a causal link passes initial muster. Social media might well have been associated with reduced anxiety and depression. It might well have only ever had a statistically insignificant association. But instead, most of the correlational research points the same way at a first pass.

Fortunately, we don’t only have the correlational research to rely on. A growing number of experimental and quasi-experimental studies support the existence of a genuine causal link. Some are social media ‘reduction’ studies, in which regular social media users take a break from social media for a period of time, and some are social media ‘exposure’ studies, which assess the impact of gaining access to social media. Both point in favor of genuine causation.

In a recent meta-analysis of 20 social media reduction RCTs covering 10,106 (adult) participants, Liu et. al. (2025) found that the interventions had a common effect size of .233 on overall well-being, a small, but significant positive effect (p. 10). The largest reduction experiment to date, Alcott et. al. (2020), assessed the effects on 2,743 people of deactivating Facebook for the four weeks before the 2018 U.S. midterm election, and found they experienced .09 SD and .1 SD decreases in measures of anxiety and depression, respectively. Another large study, Tromholt (2016), with 1,095 participants, found that quitting Facebook for only one week significantly improved life satisfaction and emotional positivity, and that both effects were larger for both heavy and passive users. Brailovskaia et. al. (2020), Lambert et. al. (2022), Reed, Fowkes, and Khela (2023), and Davis and Goldfield (2025), though smaller in size, all found similar, significant effects in their own reduction experiments. Findings of decreased symptoms of depression were particularly common. Though a number of other small experiments show conflicting results, the balance of evidence, especially when adjusted for statistical power, favors a small, but positive causal effect of reducing social media use on mental health and general well-being.

Reduction studies, though, can at best tell us about the likely effects of reducing social media use in a world where it is already very common. Exposure studies, on the other hand, can help us understand the impact of its initial spread. As one would expect given the evidence reviewed so far, exposure studies provide further confirmation of a causal link between social media use and mental health. Two in particular are worth highlighting.

Braghieri, Levy, and Makarin (2022) took advantage of a natural experiment—the staggered early rollout of Facebook on college campuses between 2004 and 2006—to assess the differential impact of the platform on mental health among college students through 2008. They found that the introduction of Facebook led to a .085 SD increase in an index of poor mental health among college students, a magnitude equivalent to around 22 percent of the effect of losing one’s job, and a 2 percentage point increase in the share of students suffering from depression. They also calculated (on the basis of some relatively strong assumptions), that the introduction of Facebook accounts for approximately 24 percent of the increased prevalence of severe depression among college students over the last two decades (p. 3662).

Guo (2022) also took advantage of a natural experiment—spatial variability in the introduction of high-speed wireless internet to British Columbia—to assess the impact of greater social media access on teenage students. She found both that access to high-speed wireless internet was a good proxy for greater social media use (p. 22-23), and that, while boys did not experience a significant change in mental health in areas with access to high-speed wireless (and if anything saw slight improvement), teen girls in those areas saw a 90% increase in severe mental health diagnoses from 2010 through 2019 (p. 28). This is consistent with previous research finding associations between digital media use and mental health to be larger in girls than in boys.

Guo explains her findings by pointing to research that suggests girls are more sensitive to how social media might reshape their social landscape and interpersonal lives. She also cites internal research done by Meta which found, among other things, that Instagram made “’body image issues worse for one in three teen girls’” and that “’constant comparison on Instagram is the reason why there are higher levels of anxiety and depression’” (p. 9).

Meta’s in-house research team isn’t alone in this conclusion. A number of recent papers have honed in on unfavorable social comparison as a key mechanism explaining the link between social media and mental illness among young people, and in particular girls. Adolescents are (in)famously attuned to social hierarchies, and whereas they once would’ve only had to worry about local ones, those at their school or in their community, social media thrusts them at an early age into a substantially larger, more competitive social pool, one in which the ‘popular kids’ are not just the richest or most attractive in their school, but the richest and most attractive in the world.

Not only that, but engagement metrics in effect serve as ‘popularity scores’ that allow young people to compare where they or their peers fall in the pecking order. It makes a lot of sense that 57% of Gen Zers would “become an influencer if given the opportunity,” considering that, for them, this is more or less equivalent to becoming ‘the most popular type of person’ in their now oversized reference group. What is often missed in overspecific debates about the hyper-competitive nature of college admissions is that all of childhood is now much more competitive, including and especially the childish parts. Call it ‘free trade’ in the teenage popularity market. Once localized status competition—subject to various forms of ‘infant industry protection’—has now gone global. Unionized American machinists versus sweatshop workers in Bangladesh? Try being a teenage girl competing for likes with Charli D’Amelio.

Jokes aside, it is straightforward to see how, for social media’s heaviest and most sensitive users, constant contact with curated displays of high status might lead to distress. Exposure experiments like Guo’s further reinforce the conclusion implied by every other line of evidence: social media causes small, but statistically significant increases in mental illness among young people.

Those who continue to stubbornly oppose this conclusion are going to find themselves with fewer and fewer intellectual allies in the years to come. Strong evidence continues to pile up, and that is before the effects of the much more recent advent of short-from video content have made it into datasets. Though a few academic holdouts will likely continue to harp on the issue of small effect sizes, they would do well to consider the following:

a) much of the literature is cross-sectional or dependent on short interventions, meaning it won’t capture the compounding effects of social media use over time (remember: teens spend almost 5 hours on social media per day—and have been doing so for over a decade);

b) effect sizes are averages, and so consistent with vulnerable sub-populations who fare much worse than the norm, and in turn impose large second-order costs on others (financial, emotional, etc.);

c) the baseline number of adolescents with any one mental illness is low, and so not especially difficult to increase in large percentage terms with an intervention (i.e. social media) applied to 100% of adolescents; and

d) many adolescents may be on the threshold of mental illness, and only need a small nudge to be pushed over.

For these reasons and others, small effects on the individual level can mean very large ones on the national level (for more on the “indispensability” of small effects, see Götz, Gosling, and Rentfrom 2021).

While numerous questions of course remain as to the exact nature of social media’s effects on mental health, the evidence of a genuine causal link is strong. Social media has done serious, and sometimes irreversible harm to America’s young people. Yes, the details matter, and will continue to be teased out by careful research, but there is little room for debate over the broader strokes. Given the often glamorous, but warped image of the world that young people consume on social media, it is no wonder that many of them feel as if they cannot possibly measure up.

I recently read F. T. Marinetti’s “The Foundation and Manifesto of Futurism” for the first time. The Italian Futurists—Marinetti being the movement’s predominant artistic voice—came to prominence in the early 20^th century, and were known for their celebration of the dynamism and virility of the modern world, in particular as manifested by technologies of transport and warmaking—automobiles, steamships, planes, tanks, bombs, bullets.

The Futurists, while mainly an artistic movement, were also known for deploying their art in the service of fascist politics. Marinetti returned from World War 1 a twice-decorated veteran and leveraged his newfound public image into the foundation of the Partito Politico Futurista, the Futurist Political Party, which was absorbed into Mussolini’s National Fascist Party in 1919.

The Futurist manifesto has a more personal, narrative character than most political writings of its kind. Its first paragraph reads more like literary fiction than fascist propaganda:

We had stayed up all night, my friends and I, under hanging mosque lamps with domes of filigreed brass, domes starred like our spirits, shining like them with the prisoned radiance of electric hearts. For hours we had trampled our atavistic ennui into rich oriental rugs, arguing up to the last confines of logic and blackening many reams of paper with our frenzied scribbling.

From here, the scribblers, soaked through with liquor and strange, new ideas, climb into Marinetti’s 1908 Fiat and speed through the streets of Italy until, after nearly colliding with two cyclists, they flip the car into a ditch. For all the overwroughtness of Marinetti’s prose, his reaction to this incident is admirably, if disturbingly informative of Futurism’s basic ideals and psychic underpinnings:

O maternal ditch, almost full of muddy water! Fair factory drain! I gulped down your nourishing sludge; and I remembered the blessed black breast of my Sudanese nurse … When I came up—torn, filthy, and stinking—from under my capsized car, I felt the white-hot iron of joy deliciously pass through my heart. (trans. R. W. Flint, 1971)

Put a different way: Futurism is defined by its sexually charged revelry in the frenzied, machinic violence of the modern world, by drunkenly careening into a ditch ensconced in metal, by getting a racist, Oedipal hard-on while you do it.

Though it has been superseded by the more professionalized, technocratic futurisms of Silicon Valley, there is something of Italian Futurism still with us. In the aftermath of this latest act of high-profile political violence, the assassination of MAGA acolyte Charlie Kirk, America is beginning to feel a bit like Marinetti’s car, speeding out of control, driven by a small group of chaos agents who revel in the aesthetics of extremist violence.

As of my writing this, we do not know who shot Charlie Kirk. But it would be a good guess that they are not a disciplined operator with a well-worked-out theory of change. They are probably, like Thomas Crooks, like Luigi Mangione, like Vance Boelter, lonely, suffering from some kind of mental illness, and influenced by some eclectically extreme backroom of the contemporary internet, or at best their own scattered, misshapen attempts to make sense of a world seeming to skate by without them. By killing or trying to, by seeing that, whether they lived or died, their actions would be blasted into the public sphere by the same vast, modern machine that incubated them, these men found a surefire path to reenter the world again, to crash their speeding cars into ditches, to force themselves, newly muddied, into the wakeful eyes of the inevitable, gathering crowd.

I want to get out of Marinetti’s car. I also want, but know I can never fully ‘get out’ of the contemporary internet, as colossal and impressive a machine for prosecuting modernity’s conflicts as has ever been, and perhaps ever will be built. And yet, it of course won’t do to answer Marinetti by banning cars or any other of modernity’s technological wonders. It will only suffice to reassert rational, public control over them.

But in the aftermath of what is sure to be one of the most explosive killings of a public figure in America since the turn of the century, that may become a much more remote prospect. Many on the right are already gearing up to exploit Kirk’s assassination in order to repress their political enemies. The conglomerates that have swallowed the internet whole have for the most part become their allies, their CEOs having already lined up to pay tribute to our dear leader. The scope of Marinetti’s metaphor hardly seems apt. A car is too quaint, too small, for the contemporary modern imaginary, with it’s A.I. apocalypses and fantasies of social collapse.

Perhaps we are careening through space in a titanic starship, cowering in small metallic rooms, vomiting into wastepaper baskets, our faces washed in blue light from flashing, touch-screened consoles. “I condemn in the strongest terms the violence that has broken out on deck 5,” a lone woman types furiously. “My prayers go out to Kirk, his wife, and the son and daughter that survive him.”

There is hoarse yelling from the floor above. An alarm is screaming red. Maybe you can do something to slow this all down, assuming it can be stopped. But what if it can’t? The sinister thought occurs: maybe it is better to prepare for after. Why not be one of the few who stands back up, bloodied, but beaming, in the wreckage of the crash?

The last few years have witnessed a glut of writing about men and their reading habits. Among the takes on offer are that men are being stolen away from reading by pornography and video games, that the decline of male reading is driving them toward the political right, and that if men read more, they would learn empathy and better ingratiate themselves toward women.

Notwithstanding that the disparity between male and female readers has likely been overstated, I think there is something to this discussion. If nothing else, it is accurately tracking that the paradigm reader in most Americans’ (and publishers’) minds is now a woman, who is reading a book marketed toward women.

It is also tracking how, in a surprising turn of cultural history, men are now treated by much of the commentariat as a troubled or dysfunctional demographic class, one whose ills are represented back to them in articles they frequently neither author nor read.

One recent entry in this genre (that received a lot of online flack) is Catherine Pearson’s “Why Women Are Weary of the Emotional Labor of ‘Mankeeping,’” which explores women’s frustrations at having to manage the social and emotional needs of their male partners—what Stanford postdoc Angelica Ferrara has termed “mankeeping.” Judging by their recent interests, many in the world of letters are as yet unwearied by their own project of collective mankeeping, their attempt to use books to guide the fallen gender toward heightened taste and emotional wellness.

Some men probably find this trend condescending. My complaint is a different one: I think many of these writers are, to borrow a recent coinage, bad ‘mankeepers’ and are in desperate need of a refresher course. Most of what’s out there claiming to pitch men on reading is painfully unconvincing—and probably because it was never meant to do that in the first place.

Take “The Real Reason Men Should Read Fiction,” published this past June in the Atlantic. The essay by author and senior Atlantic editor Jeremy Gordon positions itself against the overdone argument that reading (and reading fiction in particular) has the power to teach men empathy, with Gordon proudly advertising that “empathy is a bit too touchy-feely as a consistent motivation—at least for me.” So what motivates him instead? Not much, in my view, that stands to actually convince prospective male readers to pick up a book. This is the sum total of Gordon’s case:

“Literature … allows me to occupy a place that is totally for myself, and unaccountable to other people’s expectations … It is seductive … to keep things to yourself. To incubate your own thoughts and ideas without having to express and justify them in real time as you might when talking with other people.”

“… reading skips backward and forward, allowing me to consider the world from a removed vantage point. Instead of feeling squeezed by my earthly existence and my own bodily limits, I leap into other minds and perspectives … and consider those expectations … fiction can remind you that you exist along a continuum of human experiences, and that your own everyday ennui is less of a dead end and more of a data point.”

Everything else in the essay is either ground-laying, autobiographical color, or elaboration on these two core arguments, the first being that reading gives you private time to reflect insulated from others’ expectations—something you can just as well achieve with less effort by taking a walk; and the second being that reading encourages the consideration of others’ perspectives—which sounds suspiciously similar to the classic ‘empathy defense.’

I don’t mean to single out Gordon. I’m sure he has much more say on this topic, and I ultimately do agree with what he has said. But his essay has the unfortunate property, given its topic, of transparently not being meant to convince more men to read. It is instead meant to rehearse the virtues of reading for those who already do.

That is why its core arguments will be accepted at face value by avid readers, but will come off as schmaltzy and elusive to almost everyone else. Reading’s public defenders love to gesture in circuitous terms to the activity’s joy and/or power, while spurning more direct appeals to concrete incentives. They might consider emphasizing, for instance, that reading literature is a challenging intellectual undertaking that will make you smarter and more capable if you do it.

This, I offer, is the sort of appeal more likely to be heard by men. I know what you’re thinking: it violates that common dictum, ‘thou shalt not instrumentalize reading.’ But that, I also offer, is too bad. In a fast-paced world making temporal misers of us all, it is imperative to tell people why time-consuming things are worth doing, and to make such appeals in straightforward terms they will understand. There is not always something wrong with stripping away layers of figurative language, and just telling people what they stand to gain.

That all said, here is an honest attempt by a man who loves to read to pitch men on reading, one that I hope avoids the major pitfalls of its forerunners. I want to be clear that this is not an exhaustive description of why I read, nor of why everyone should read. It is very specific to the goal of pitching men—in particular those skeptical of or uninterested in reading—on why they should read. Feel free to share with any men you may be ‘keeping’ at the moment.

Forget what you know about reading. Forget whatever boring stuff you were forced to read in school. Forget whatever ‘vibe’ you think reading has. Reading is not about being a certain type of person. Reading is not about signaling that you are smart or worldly. All reading is about, all it will ever be about, is processing linguistic information, information delivered to you in language, in words, in sentences. You’re probably pretty good at this already, or else you couldn’t read what I’m saying now. But you could be better, way better.

Long ago, the physically strong held more sway over human affairs. But in the modern world, it is your ability to internalize, organize, and extrapolate from information that makes all the difference. It is your mental, more than your physical strength that determines your lot in life. In a world like ours, you need a mental training regimen, just as much, if not more than you need a physical one.

So what should you do? You could learn a narrow technical skill like coding. You could play an intellectually challenging game like chess. But these fields are defined by their strict, exceptionless rules. The real world is much more uncertain. The real world is much more colossal and diverse. The real world is not a game, and it doesn’t have a rulebook.

To prepare for it, you need something that shares these same properties. You need to test your ability to reason absent hard-and-fast rules, to organize disordered, uncertain, and incomplete information. You need to throw yourself at something that won’t hold your hand, the same way the world throws itself at you. That thing is a book.

Books collect massive amounts of complex information into a unified whole. To read them—and especially to read them well—means expending sustained effort to retain and interpret the information they contain. There is no way to do that without learning something, without becoming a smarter, more capable person than before. And the more you read, and the more challenging what you read, the more those gains compound.

Take it from someone who used to read very little. Once I started, I couldn’t stop. I saw how each book shaped the person that I was. I saw how each book represented a new challenge that required focus, discipline, and ingenuity to surmount. I felt better about myself. Things got a little bit, and sometimes a lot easier in other parts of my life.

A book is like a little world. More specifically, it is like a model of a world, or a simulation of one, that is made out of words. It is a linguistic model in the same way that models of, say, the solar system are mathematical models. Moby Dick is a model of a tireless whaling-boat captain’s journey to take revenge on the titular ‘great white whale’ that tore off his leg. Das Kapital is a model of the 19th century European manufacturing economy, and how it shaped the political and distributional conflicts of its time. The more you practice engaging with these ‘linguistic worlds,’ and the more of them you familiarize yourself with, the more prepared you will be to take on the real thing.

You probably already exercise. Maybe you lift weights, maybe you run, maybe you play a sport. But you don’t live in a world ruled by strong or fast people. You live in a world ruled by owners, architects, and manipulators of information. In a world like that, those who don’t read are at the mercy of those who do. So if for no other reason than self-protection, do yourself a favor and go pick up a book.

Note: this post makes the case against social media on political grounds. Those interested in the well-documented mental health harms of social media should read this post as well.

1. Introduction

The philosopher Dan Williams recently published two pieces on social media— “Scapegoating the Algorithm” at Asterisk Magazine, and “The Case Against Social Media is Weaker Than You Think” at his Substack. As their titles attest to, both argue that the case against social media, on epistemic and political grounds, has been considerably overstated.

I recently published a lengthy essay arguing the opposite: that the case against social media has, if anything, been understated. And so, especially given I’m new to Substack, Williams’s recent work gives me a welcome opportunity to pick a fight for engagement-farming purposes. Sadly for my subscriber count, I agree with Williams on quite a bit, and so this is going to be less combative polemicizing and more demonstrating that a range of serious worries about social media are capable of charting a course between the equivocal lines of evidence he underlines.

I am going to focus on the putative political impacts of social media—in particular its impact on political polarization—rather than the specifically epistemic ones. These are often conflated, but I think it is helpful to cleave them apart. It is possible to believe, as I in fact do, that social media has had a dangerous, incendiary effect on American politics, while also believing that related concerns about misinformation and conspiracy theorizing are somewhat (though not entirely) overblown. I may decide to write a short follow-up on the latter, but will put it to the side for now.

With respect to social media’s supposed contributions to political polarization, and in particular ‘affective polarization,’ Williams highlights four main lines of evidence that contradict the prevailing narrative. First, polarization has been rising for decades and began doing so well in advance of social media. Second, polarization has increased the most in recent years among those who use social media and the internet the least, those over 65. Third, trends in polarization have diverged in countries that all have widespread social media use. And fourth, several high-quality experimental studies have found a negligible effect of social media use on individuals’ levels of polarization.

If I can import a cliché from academic philosophy, it is sometimes said that all argumentative objections fall into one of two categories: “oh yeah?” or “so what?” I opt for a mix of both here. I begin with the “oh yeah?”: Williams’s evidence is much less convincing than it initially seems, and beyond ruling out an implausibly large spike in polarization, tells us little about how social media might be influencing its trajectory.

I then turn to my main focus, the “so what?”: even if Williams is right that social media has not significantly contributed to affective polarization in the U.S., this is consistent with it having plenty of other negative effects on American politics. To make this argument concrete, I highlight different lines of evidence to Williams in order to demonstrate first, that social media has indeed had a deeply concerning impact on our politics, and second, that one need not rely on the metric of affective polarization in particular to make this case. I refer to the alternative view I develop as an elite radicalization theory of online politics.

While I am an avid reader and admirer of Williams’s work, I think he is engaged in what has unfortunately become a fashionable form of academic contrarianism: arguing that despite what very much seem to be foundational, risk-laden changes in our social order, all we observe is in fact business-as-usual, and to the extent it seems otherwise, that is because of some mix of psychological bias and media overhype.

A similar contrarianism plagued early predictions of the threat posed by Donald Trump. Those critics have either relented or fallen quiet now. It is time for those who continue to minimize the downside risks of the digital media revolution to finally do the same.

2. Williams on social media and polarization

Let’s first run through Williams’s four main lines of evidence against social media-driven polarization in a bit more detail.

The first is that affective polarization has been rising since well before the advent social media. As Williams explains in “Scapegoating the Algorithm,” this trend begins in the late 20^th century and likely has its roots in a) the partisan realignment surrounding the Democratic party’s 60s-era embrace of racial integration, and the Republican party’s subsequent “Southern Strategy” to appeal to disaffected white opponents of integration; and b) the rise of a more combatively partisan media ecosystem in the aftermath of Reagan’s 1987 repeal of the “fairness doctrine,” which required broadcasters to present balanced perspectives on controversial public issues.

Williams’s second line of evidence is the study Boxell, Gentzkow, and Shapiro (2017), which found that, between 1996 and 2016, the demographic groups that used social media and the internet the least, those 65 and older, also saw the largest increase in polarization.

Williams’s third line of evidence is a later study by the same authors, Boxell, Gentzkow, and Shapiro (2024), which studied trends in affective polarization since the 1980s in 12 OECD countries and found that, while the U.S. exhibited the largest increase over this period, countries like Australia, Britain, Norway, Sweden, and Germany all saw polarization fall, including in the period after the initial spread of social media.

Williams’s fourth line of evidence is that several high-quality experimental studies have examined the relationship between social media and polarization at an individual level and found no evidence of a connection. Guess et. al. (2023a), Guess et. al. (2023b), Nyhan et. al. (2023), and Alcott et. al. (2024) all found that interventions meant to limit social media’s ostensible polarizing effects (such as replacing default with reverse-chronological feeds, hiding reshares, reducing exposure to content from like-minded sources, and deactivating social media entirely) had a negligible impact on users’ levels of polarization.

2.1. “Oh yeah?”

Taken together, this evidence indeed appears to contradict that social media has been a major contributor to American political polarization. But dig a bit deeper, and things get much messier and more uncertain.

The fact is that we have relatively little data about affective polarization since ~2010, when smartphone and social media use became widespread in the U.S. The vast majority of this data comes from the American National Election Studies (ANES) survey, whose flagship data set is only collected during presidential election years. As a result, we only have 4 data points since 2010, those for 2012, 2016, 2020, and 2024.

It is difficult to draw strong conclusions from this data one way or the other. Besides the fact there isn’t much of it, rival-party feeling does appear to start falling faster around 2010 or so, which would be consistent with some social-media related effect. But I concede I’m just eyeballing here; more rigorous conclusions will have to await further research.

The data on international trends in affective polarization isn’t especially convincing either. The main study cited by Williams—Boxell, Gentzkow, and Shapiro (2024)—looks at affective polarization in 12 countries since the 80s, but for our purposes, it is only the post-2010-years that are relevant, since it is only after this point that social media would’ve had any notable effect. That in mind, the data here is also relatively unsatisfying:

With the exception of Germany, there are never more than 3 data points per country after 2010. When you consider on top of this that rates of social media penetration over time probably vary across these countries, it becomes clear that there is little we can learn through these comparisons without more context, or at least more time.

The bare fact of diverging trend lines also tells us little to begin with, since these trends might have been different absent the influence of social media. For instance, polarization might have risen more slowly in the U.S. and fallen more quickly in Germany. Sure, maybe this data does conflict with the most alarmist narrative about social media—that there was some massive spike in polarization after 2010—but that was never plausible to begin with. There is still more than enough room, even in light of this data, to be justifiably concerned about how social media is shaping American politics.

That in mind, let’s move on to Williams’s two other lines of evidence—that polarization has grown more in recent years among the elderly (those least likely to use social media), and that the experimental research on social media has not turned up any appreciable relationship with polarization.

The paper documenting higher polarization among the elderly—Boxell, Gentzkow, and Shapiro (2017)—makes a very important caveat that Williams does not mention. While the authors see themselves as ruling out the most straightforward version of the social media-driven polarization story, they acknowledge it is possible to construct an alternative account that they cannot rule out:

[Alternative accounts would need to address] the rapid increase in polarization among those with limited internet use and negligible use of social media. However, it is possible to construct such accounts. It may be that social media increases polarization among the young while some other factor increases it among the old. It may be that there are spillovers across demographic groups; young adults polarized through social media might in turn affect the views of older adults or might indirectly influence older adults through channels like the selection of politicians or the endogenous positioning of traditional media. (p. 3, bold my own)

While we should not fault Boxell, Gentzkow, and Shapiro for the scope of their specific project, it is obvious in my view that every one of these spillovers is in play. I suspect the most important is the possibility that social media is influencing the tone and coverage of traditional media, and in particular cable news. This should ring true to anyone familiar with right wing cable’s aggressive coverage of various social media-native controversies, like the recent American Apparel ad starring Sydney Sweeney, or the various sorority rush videos currently going viral on TikTok. If this kind of coverage increases polarization among the elderly, that would still in part be social media’s fault. In light of potential spillover effects, this evidence can at best rule out that social media has influenced all demographic groups equally, and just as immediately, rather than seeing its influence channeled from some groups to others.

But this kind of ‘channeling’ was always going to be part of the story. Social media’s intrinsically social nature guarantees it will have large spillover effects. Ignoring this is like assuming that, if you could only delete your social media accounts, that would insulate you from 100% of social media’s effects on your life. The obvious reason it won’t is because everyone around you will still be using social media, and their use will continue to affect you (e.g. by their telling you what’s happening online, by you feeling excluded from their group-chats, etc.). As long as social media represents the social world to us, and as long as we share our impressions of that social world with others offline, then its effects on individuals are going to ‘spill over’ to their peers.

It is therefore wrong to assume that, since old people don’t use social media (or use it much less), they are not being affected by it. Due to spillovers, those of them that watch cable news are arguably affected by it every day. Not only that, but their relative lack of experience with the modern information environment means their political attitudes are likely to be more sensitive to (even the indirect effects of) social media than those who use it more often.

As a general rule, very few strong conclusions about the aggregate impact of social media can be drawn from research that does not account for these kinds of spillovers.

Let’s now turn to Williams’s last line of evidence, the experimental research on social media and political attitudes. This is probably the strongest evidence we have against social media-driven polarization. And yet, think carefully about what it is really in a position to tell us, and the result is again underwhelming.

Of the experiments Williams cites, the longest lasted three months, and only one—Alcott et. al. (2024)—actually entailed the complete deactivation of social media, rather than just modifications to how and what content was displayed. Already, this should give us pause about interpreting this evidence to imply that social media’s supposed impacts on politics are overblown. On a more even-handed read, all the studies besides Alcott et. al. (2024) show is that feed design changes don’t impact users’ political attitudes over relatively short periods.

Even focusing on complete deactivation, though, spillover effects remain a serious problem. Like anyone else, social media users’ political attitudes originate from and are continually modified by a wide variety of sources: their families, their friends, prominent members of their community, other acquaintances, online news, print news, T.V. news, podcasts, talk shows, radio shows, and so on. If these other influences have themselves been influenced by social media, then deactivating participants’ Facebooks and Instagrams for six weeks—the intervention in Alcott et. al. (2024)—does little to isolate them from the effects of social media in general. They experience those effects anyway by continuing to communicate with other social media users.

This is an even greater concern than usual in the case of Alcott et. al. (2024) because their intervention took place in the six weeks before the 2020 election. Those who took part in the experiment would have been inundated by far more political communication than usual, communication which was almost certainly influenced in some form by social media.

The moral is again that spillover effects aren’t to be ignored. Given social media’s now-enormous role in shaping public discourse, interventions targeting personal use alone can at best isolate a small sliver of social media’s total impact. That means that even high-quality experiments on individuals tell us little about social media-driven polarization in aggregate.

3. “So what?”

Suppose all these objections are unimpeachably correct. I’m sure they aren’t, but let’s assume. That would still only amount to a case for uncertainty. It remains, then, to make the case that rather than only remaining cautious about social media’s effects on politics, we should be seriously concerned.

Williams is right that the evidence on social media and affective polarization will not get us there on its own. But affective polarization is hardly the only potential measure of social media’s harms. Consider the following claims Williams cites (in “The Case Against Social Media is Weaker than You Think”) as typical of the public concern over social media:

Alexandria Ocasio-Cortez describes Meta as “a cancer to democracy metastasizing into a global surveillance and propaganda machine for boosting authoritarian regimes and destroying civil society.”

Jonathan Haidt has argued that social media platforms “dissolved the mortar trust, belief in institutions, and shared stories that held a large and diverse secular democracy [in America] together.’’

Obama suggests that one of “the biggest reasons for democracies weakening is the profound change that’s taking place in how we communicate and consumer information.”

These are of course strong criticisms. But it is far from clear that they depend on social media being a major driver of affective polarization in particular. One of the issues with placing academic and non-academic discourses in conversation is that the terms of the former may not map cleanly onto the terms of the latter (and vice versa). In this case, there is no reason to assume claims like those above rest on the evidence in favor of social-media driven polarization, especially if there are other, more convincing lines of evidence available.

This is my primary complaint against Williams: there are other, more convincing lines of evidence available.

Even if social media is not increasing affective polarization, it is making our politics more angry, tribal, and violent—and to more than a sufficient degree to justify grave concerns like those of Ocasio-Cortez, Haidt, and Obama.

In what follows, I focus on two lines of evidence in particular: research showing that social media amplifies the reach of emotionally extreme content, and research showing that higher social media penetration in a given geographic area induces more extreme political behavior, such as protests and hate crimes. I synthesize both into an overarching view I refer to as an ‘elite radicalization‘ theory of online politics.

3.1. The elite radicalization theory

It is well-established in the research literature that our attention on social media gravitates toward content that is more emotionally extreme. Recent studies have found that posts that are sad, fearful, uncivil, morally and emotionally indignant, hostile towards out-groups, and negatively valenced in general spread much further than their more neutral counterparts.

Brady et. al. (2017), for instance, found that each additional word with “moral-emotional content” (e.g. “duty,” “fear,” “shame,” “war”) increased political Twitter posts’ diffusion factor by 20%. Brady et. al. (2019) later reproduced this finding using the tweets and retweets of over 500 presidential candidates and members of congress. Expressions of “moral anger and disgust,” the authors found, diffused particularly quickly.

That negative content spreads especially far online makes sense given some well-established properties of human psychology. Humans exhibit a broad-based psychological negativity bias as well a range of more specific attentional biases toward negative stimuli. This is because, like other organisms, we have evolved to be uniquely attuned to signals suggestive of danger. As a result, when given the opportunity to cycle through a variegated soup of thousands of digital signals daily, we tend to fixate on and amplify the most distressing ones.

This dynamic creates very strong incentives to prey on our negative emotions. Successful ‘attentional entrepreneurs’ online not only enjoy attention’s more immediate benefits like status or influence, but are often paid by platforms in proportion to the ad revenue they generate. In that case, if certain kinds of content reliably accrue the most attention, there are very strong incentives online to produce that content en masse.

That is exactly what has happened in recent years with sensationalized and excessively negative political content. The last decade and a half has seen the rise of a new class of political influencers who, empowered by social media’s unique incentive environment, have come to exert near-symphonic control over the fear, anger, and tribalism of large sectors of the American public. The phrase “political influencer” calls to mind names like Tucker Carlson and Candace Owens, but I mean it to refer to any content creator, pundit, journalist, or even politician with an active online presence oriented around the production of political content—so perhaps hundreds of thousands of users with followings of varying sizes. Critically, this group is not a random selection from social media’s overall user base, but skews wealthier and more educated, meaning its greater online influence is likely matched by greater offline influence as well.

Despite accounting for only a small slice of the online population, this new influencer class is coming to dominate the market for political communication. In the process, it is transforming America’s perception of itself, which, since America is a social entity constituted in part by its self-perceptions, just amounts to saying it is transforming America.

Robertson, del Rosario, and Van Bavel (2024) outline this group’s newfound influence in their recent paper “Inside the funhouse mirror factory: How social media distorts perceptions of norms.” It is worth quoting their analysis at length:

Online discussions are dominated by a surprisingly small, extremely vocal, and non-representative minority. Research on social media has found that, while only 3% of active accounts are toxic, they produce 33% of all content. Furthermore, 74% of all online conflicts are started in just 1% of communities, and 0.1% of users shared 80% of fake news. Not only does this extreme minority stir discontent, spread misinformation, and spark outrage online, they also bias the meta-perceptions of most users who passively “lurk” online. (p. 1, my italics)

[With respect to] political discussions [in particular], the people who post frequently on social media are often the most ideologically extreme. Indeed, 97% of political posts from Twitter/X come from just 10% of the most active users on social media, meaning that about 90% of the population’s political opinions are being represented by less than 3% of tweets online. This is a marked difference from offline polling data showing that most people are ideologically moderate, uninterested in politics, and avoid political discussions when they are able. (p. 2)

The result is that Americans are coming to see each other as much more politically extreme than they in fact are—or at least were. In the words of political scientist Henry Farrell, Americans have internalized a “malformed collective understanding,” in this case a vision of their collective identity that casts them as especially angry, pessimistic, and tribal.

The idea that this understanding is “malformed” is a tricky one, though. It invites the question: how malformed, and for how much longer? While in the past, survey evidence has indeed shown Americans to be moderate and relatively uninterested in politics, it is possible, and even probable, that social media is changing that. If in response to believing their peers are more politically extreme, people in turn become more extreme in reaction, then Farrell’s ‘malformed collective understanding’ may end up a self-fulfilling prophecy. Perhaps the more extreme content we ingest, the more likely it is we internalize this style of political communication and begin posting extreme content ourselves. There is already some limited evidence to support this possibility.

Kim et. al. (2021) found, for instance, that those exposed to toxic ‘featured’ comments on Facebook posts were more likely to post toxic comments later of their own volition. Brady et. al. (2021) found that Twitter users “conform their outrage expressions to the expressive norms of their social networks,” (p. 1) and become more likely to express outrage when their surrounding social network is itself more extreme (p. 8). And in an as-yet unpublished preprint, Brady et. al. (2025) found that in a simulated Twitter-like environment, “engagement-based algorithms” changed participants’ perception of descriptive and prescriptive norms surrounding “ingroup-aligned, moral and emotional political content,” and ultimately made them more intent on posting this content themselves.

These results seem consistent with widespread anecdotal evidence of once-moderate friends and family falling down political or conspiratorial ‘rabbit-holes,’ often via involvement with insular online communities (e.g. QANON). At a broader level, it is hard to deny that the expressive norms governing our political culture have changed (witness the much-discussed expansion of the Overton window), and that this has made extreme political speech (open Nazism, racism, misogyny, antisemitism and so on) more common.

If this is all downstream of a small and relatively well-off group of high frequency posters (some of our elected officials among them), that would suggest what we might call an ‘elite radicalization’ theory of online politics. The idea is that social media has empowered a (relatively) small group of political influencers who, in response to the perverse incentives created by attentional negativity bias, have disseminated ideas that make people more angry, fearful, and extremist.

As represented above, those ideas spread both via influencing other elites (think podcast and talk show hosts like Joe Rogan or Jon Stewart), as well as by influencing typical consumers of online political content directly. Regardless of the exact transmission pathway, social media’s vast webs of interaction ensure that whatever effects online political content has, its reach is sure to be very far.

Nor does this content just ‘bounce off’ political junkies already firm in their preexisting views. There is strong causal evidence that political influencers have the power not just to reflect, but to shape the attitudes of their followers. In a recent preprint, Rathje et. al. (2025) found in two different field experiments (n = 494, n = 1,133) that unfollowing “hyperpartisan social media influencers" on Twitter (they give the examples of Palmer Report on the left and Breitbart News on the right) “improved [participants’] recent feelings toward the out-party by 23.5% compared to the control group, with effects persisting for at least six months,” (p. 2). This should put to rest the idea that inflammatory political content is a purely ‘demand-side‘ phenomenon. Yes, this demand has always existed, but it is possible for the attitudinal effects of existing supply to increase it even further.

Perhaps this is not on its own cause for serious alarm, though. However unpleasant online political influencers’ ideas might be, the question is whether they transcend the digital realm to influence people’s real-world political behavior. If not, then maybe we can relax in the confidence that the apparent rise in extremist attitudes is confined to an annoying, but harmless class of digital ‘LARPers.’

With apologies to those in the market for this kind of apparently savvy, relieving narrative, there is a convincing base of evidence showing just the opposite. By amplifying extreme content online, social media does appear to catalyze more extreme political behavior offline as well. The following are quasi-experimental studies that take advantage of exogenous variation in how different social networks have spread in order to study their political impacts. Their findings are relatively univocal:

Bursztyn et. al. (2019) instrumented on variation in city-level penetration of VK (the dominant Russian social network) in Russia and found that greater social media use increased the prevalence of xenophobic attitudes, and lead to more hate crimes in cities with higher pre-existing levels of nationalism.

Sabatini and Sarracino (2019) instrumented on variation in broadband/high-speed internet coverage in Italy, and found that greater social media use significantly decreased measures of social, particularized, and institutional trust.

Enikolopov, Makarin, and Petrova (2020) also instrumented on variation in city-level penetration of VK in Russia and found that a 10% increase in penetration into a given city increased the probability of a protest there by 4.6% and the number of protesters by 19%.

Schaub and Morisi (2020) instrumented on variation in municipality-level broadband coverage in Germany and Italy and found that greater access to online communication platforms made Germans and Italians more likely to vote for the populist AfD and M5S parties, respectively.

Müller and Schwarz (2021) used a slightly different approach—instrumenting on variation in Facebook and internet outages in Germany, rather than access—and found that greater social media use predicted increased hate crimes against refugees.

Müller and Schwarz (2023) instrumented on variation in the number of Twitter users across counties induced by early adopters at the 2007 South by Southwest (SXSW) festival and found that higher Twitter use led to a sizable increase in anti-Muslim hate crimes during the 2016 Republican primary, likely driven by future president Donald Trump’s tweets. In particular, they found that “a one standard deviation-higher exposure to Twitter [was] associated with a 32 percent … increase in hate crimes.” (p. 272)

These studies’ geographic range underscores that their concerning effects are robust to meaningful changes in context. And since they aggregate data across whole countries and municipalities, they’re able to account for the kinds of spillover effects that individualized experiments ignore. At least within the Western world broadly construed, the spread of social media has consistently produced more extreme forms of political behavior, including xenophobic violence in the most severe cases, but also protests, shifting voting patterns, and presumably much more that scholars have yet to study in detail.

While the evidence on this front is more limited in the case of the United States, the graph below does not inspire much optimism—note the hinge point around the late aughts when social media (and smartphone) use first became pervasive.

Social media is not, then, just a vast machine for disseminating extreme ideas—the most successful of which seem to have a right-wing populist flavor—but also one for shaping practical politics, including in its highest-stakes moments. As the quasi-experimental research reinforces, it does so in large part by empowering elite political influencers who, far from being just ‘terminally online’ pests, have the power to genuinely refigure their home countries’ political terrain.

While much of the preexisting research on social media and politics is focused on the political right, it is worth flagging that the elite radicalization theory also helps explain the strand of more radical left-wing politics that has gained prominence over the last decade and a half. While, as a factual matter, left-wing political influencers are far less dangerous (see the preceding graph), it is no less true that they have successfully exploited attentional negativity bias for both personal and political gain, shaping our current political culture in the process. As it happens, I am more sympathetic to their views, and so think some of their impact has been positive. But that is my own personal view, not an implication of the theory.

3.2. Elite radicalization and two-party politics

One of the core advantages of the elite radicalization theory is that it can account for social media’s incendiary effects on politics without committing to any particular relationship between social media and political polarization. While I hope the theory will be independently interesting to readers, it is for this reason that it can minimally act as a “so what?” objection to Williams.

Note that other common accounts of social media’s relationship with politics build in some commitment to polarization from the outset. For instance, ‘echo-chamber’ or ‘filter-bubble’-based theories argue that social media sorts people along partisan lines, in which case it would make sense for those inside a given bubble to become more polarized against those outside.

But the elite radicalization theory makes no such commitment. Of the elite influencers empowered by social media, many have views orthogonal to the classical ideological divide between Republicans and Democrats. While the content they create is often angry, fear-mongering, and prejudicial, it is only occasionally aligned with one of the U.S.’s two major political parties. More often, it has a populist flavor that targets some nebulous group of elites spanning party lines. And much of it also straddles the divide between political and apolitical, weaving discussion of, say, the economy or the Epstein files into broader conversations about pop culture.

As a result, and contra much presumptive emphasis on polarization, online political content may decrease the intensity of Democrat or Republican party identification, not in spite of, but because it makes people more outraged about politics. That outrage, for instance, may make independents less likely to gravitate toward the Republican or Democratic party. It might also make those who are already party members dislike their party more, perhaps enough to leave it altogether. This would be consistent with the rising public distrust of both parties, the frequent infighting that now defines Democratic party politics, and the simmering tensions between the more nativist and more business and tech-oriented factions of the Republican party.

If social media does in fact decrease the incidence or intensity of major party identification overall, then affective polarization—since it is defined in terms of major party allegiances—would be the wrong metric to look at in order to understand social media’s effects on politics. While this is only speculation for now, the graph below showing a rising share of Americans identifying as independents is at the very least suggestive (again note the hinge point just before 2010).

None of this is to say that social media definitely doesn’t increase polarization. But at minimum, this is by no means necessary for the elite radicalization theory to succeed. With respect to ‘affective’ polarization in particular, the theory suggests that social media’s harms probably have more to do with the ‘affective’ part, than the ‘polarization’ part. In other words, while social media makes political discourse more affectively or emotionally intense, it may do so without making avowed Democrats or Republicans dislike each other more than they already did.

If the effects of social media go even further to the point of convincing a significant number of Americans to redefine as independents, that would not only make affective polarization less relevant, but might even slow its increase. This could happen if those convinced to leave their party exhibited higher-than-average-negativity towards the out-party when they were still members. That may seen counterintuitive, since presumably the Republicans (Democrats) who hate Democrats (Republicans) the most would be among those most loyal to the party, but it may be that high out-party hate is a proxy for greater disagreeableness generally, in which case those who are fed up enough with the two-party system to quit it may also be among those who once exhibited the most out-party hatred.

There is not yet sufficient evidence to confirm or deny any of these suggestions. But in general, academics and public commentators alike should be mindful that in addition to spreading political extremism, social media may also be shifting the underlying axis on which those extremes exist. This could have serious consequences for social-scientific measurements of political attitudes, not least that of affective polarization.

5. Conclusion

By incentivizing the creation of disproportionately negative and sensational content, and by in turn inducing more extreme, even violent political behavior, social media has almost certainly played a major role in the destabilizing political ructions of the last fifteen years—in particular in the U.S., but probably across Europe as well.

Elite political influencers, both those hailing originally from politics, television, and journalism, as well as those native to digital media platforms, have played a critical mediating role. It is they who generate much of the most provocative content, as well as who influence the outlooks and incentives of other elites, allowing their influence to spread well beyond the confines of the major social networking platforms.

While Williams is correct that we have not seen a large spike in affective polarization as a result, we have seen spikes in angry, fearful, and identitarian political speech. We have also seen large spikes in political violence. And though I will not pretend to have made the case for their direct causal connection to social media, we have also seen the rise of a more radical strain of progressive politics, as well as the once-fledgling, now all-powerful MAGA movement figureheaded by Donald Trump, both of whose successes seem hard to separate from the belligerently online politics of the 2010s. All of this has come to pass even as polarization has increased at an overall smooth rate.

As the digital media revolution continues to transform social life in the U.S. along almost every conceivable dimension, it is critical that we have a clear-eyed view of its effects. Williams is right to challenge certain aspects of the panic surrounding social media, such as the often confused and imprecise discourse surrounding “misinformation.” But when it comes to the technology’s effects on politics writ large, he has not taken a sufficiently holistic look at the evidence.

It is a widely discussed fact that A.I.-based cheating is rampant on college campuses. In January 2023, a survey of 1,000 students found that 89% had used ChatGPT to help complete a homework assignment. As suggested by several buzzy recent articles on the issue, that figure is by now almost certainly an understatement. In the words of one student quoted on her and her fellow students’ A.I. use in a recent New York Magazine piece, “we can’t really imagine living without it.”

Nonetheless, college administrators are embracing the integration of A.I. into campus life. Schools like Duke, the CSU system, and the University of South Carolina have already contracted with OpenA.I. to offer students and faculty free and unlimited access to ChatGPT. These moves have been widely justified by the claim that A.I. proficiency is set to become table stakes for competing in the white-collar labor market.

Universities’ embrace of A.I. sits uncomfortably, though, with the emerging body of evidence that when misused, A.I. tools have the potential to corrode foundational cognitive abilities. A recent preprint from researchers at MIT found that those who had access to A.I. assistance while completing an essay-writing task showed reduced neural connectivity and overall activation (as measured by EEG analysis), and produced less ‘unique’ outputs (as judged by English teachers) than their counterparts who had no access to A.I. assistance. In a survey-based study of 666 UK residents’ A.I. tool usage, SBS Professor Michael Gerlich also found a strong negative correlation (r = -.68) between A.I. tool use and critical thinking scores, writing that “as individuals increasingly offload cognitive tasks to A.I. tools, their ability to critically evaluate information, discern biases, and engage in reflective reasoning diminishes” (p. 14).

This raises the thorny question: given that cheating seems to be one of A.I.’s most popular use cases, to what extent are its putative benefits in the form of greater employability and productivity worth it to students, if they come at the cost of worsening foundational skills? And isn’t it plausible that the latter will overwhelm the former anyway, given those foundations have their own presumably strong relationship with employability and productivity? Maybe the risks are lower for adults who have already cemented their core skills and knowledge-base, but how much of their educations should young students be willing to sacrifice to A.I.?

A.I. optimists will deny this trade-off on the grounds that A.I. is also a powerful tool for learning. And under the right conditions, they may be correct. Prominent LLMs like ChatGPT can ostensibly act as super-charged personal tutors, ready to offer endless, high-level instruction on almost every topic known to man. Recent research is at least moderately supportive of this conclusion. Wu and Yu (2024) conducted a meta-analysis of 24 experiments and quasi-experiments on A.I. chatbots in education and found that they had a statistically significant positive effect on learning outcomes for college students (although effects were insignificant for primary and secondary school students). An experiment conducted by the World Bank on secondary school students in Nigeria found that access to an A.I. tutor increased English language learning outcomes by .23 standard deviations, the equivalent of “1.5 to 2 years of ‘business-as-usual’ schooling.” And a recent RCT conducted at Harvard on physics students found that those given an A.I.-supported lesson at home saw double the median learning gains (compared to a pre-test baseline) of a control group that completed a traditional lesson on the same topic in class.

As encouraging as these results are, though, other research suggests we should approach them with caution. Consider Bastani et. al. (2025), published last month. The authors—largely researchers at the Wharton School—conducted an RCT on around 1,000 high school math students in Turkey involving three phases: first, students in three randomly assigned classrooms reviewed a previously covered math topic; second, the students solved practice problems on this topic; and third, they took a closed-book exam on the topic. The treatment was applied on phase two: one class was the control and received no A.I. assistance on the practice problems, one was given access to “GPT Base”—a chat interface based on GPT-4—and the third was given access to “GPT Tutor,” the same underlying LLM, but armed with a prompt that included problem solutions, teacher input on common mistakes, guidance on how to provide feedback, and explicit instructions not to give away answers.

The study had two main results. First, students in the GPT Base and GPT Tutor groups performed 48% and 127% better, respectively, than the control group on the practice problems. Good news right? Not exactly. The second main result was that the GPT Base and GPT Tutor groups performed 17% worse and no differently, respectively, than the control group on the unassisted exam. After analyzing the messages sent by both of the treatment groups to their chatbots, the authors concluded that “the vast majority of students [were] using GPT Base to obtain solutions, whereas a significant fraction of students [were] using GPT Tutor in a purely substantive way” (p. 15). In other words, those students who could simply ask the chatbot for the answers did, and as implied by the exam results, learned much less than even the control group as a result.

This is a crucially important study for evaluating the prospective impact of A.I. on education. Along with the World Bank and Harvard studies cited above, it implies that A.I. assistants carefully designed around pedagogical best practices—for instance, by being prompted about common student mistakes, and told not to prematurely give up answers—may improve outcomes (although, in this case, ‘GPT Tutor’ simply had no effect one way or the other). But it also shows that when not restricted to A.I. tools designed to promote learning, students often use A.I. to circumvent learning instead.

The future of A.I. in education, then, may turn on how effectively educators can restrict access to ‘bad A.I.’ and encourage access to ‘good A.I.’ The problem of course, as the A.I. cheating crisis makes plain, is this is very hard to do in practice. There is no way around students doing a substantial portion of their learning outside of the classroom, meaning that even if they are given access to ‘good A.I.’—models prompted or otherwise designed to encourage real learning—they will always be tempted by the easy availability of ‘bad A.I.’—the kind that will do their work for them, leaving them worse off than they would be in a ‘no-A.I.’ world.

Among other reasons, this is why A.I. will not “democratize” education. People said the same thing about the internet and were wrong then (as indicated by stagnant, or increasing achievement gaps) for the same reasons they are wrong now. No technological innovation can make up for the mix of bottom-up motivation and top-down guidance that convinces students to undergo the strenuous, painful process of learning new things.

For better or worse, students have historically been motivated to undergo that process through somewhat superficial, short-term incentives like grades. Prior to the rise of generative A.I., students who lacked the motivation to learn still couldn’t opt out of doing so entirely on pain of failing their classes (or incurring the costs of more frictional, risky forms of cheating like plagiarism or paying others to complete their assignments). But now, because of A.I. cheating, they can and are, grades and reputations unscathed. In other words, since A.I. allows students to cheat easily and undetectably on all but the most complex assignments, it has severed the already thin connection between grades and genuine learning. That means there are even fewer legible incentives to learn now than there were before the rise of A.I.

Perhaps that will redound to our benefit in the long run, as we are forced to reinvest in finding reliable means to both convince students to learn, and to evaluate the degree to which they have. But for now, the education system is struggling to keep up. Assuming, as the budding research literature implies, that A.I. cheating is both 1) extraordinarily common and 2) corrosive of genuine learning, then millions of students passing through our education system are learning less than they once did. In a period when students are still reeling from Covid-era learning loss, the A.I. cheating epidemic threatens to set them back even further.

While well-designed A.I.-tutoring software may soon boost outcomes for a growing subset of the population, it will never be the silver bullet its most adamant boosters are hoping for. That is because the problem of learning is and has always been the problem of making young people do strenuous cognitive labor whose primary rewards they will not realize until well into the future. The evidence so far suggests A.I.-based tools may be able to help, but whatever gains are made in constrained settings—in which responsible A.I. use can be monitored and enforced—are going to be rolled back twice over unless we find a way either to prevent A.I. cheating (unlikely) or convince students not to do it (also unlikely).

I’m left to conclude that, while a very small group of savvy, motivated students will benefit from A.I.-powered educations, far, far more will use A.I. to avoid the hard work of learning instead. That may not matter so long as these students live lives devoid of the kinds of personal and intellectual challenges ChatGPT cannot surmount for them. But the moment they step outside of that exceedingly narrow zone, the new cognitive limits imposed on them—by themselves, by A.I. providers, by irresponsible educators—are going to come rushing back.

I recently published my debut essay on Substack, “the Leviathan, the Hand, and the Maelstrom” (LHM). Though I’m optimistic it’ll continue to find readers, it’s a long piece, and I don’t want the ideas within to be held hostage to folks’ willingness to read 12,000 words. So I’ve decided to write a shorter primer on one of the main issues it explores—markets for digital content.

In LHM, I argue that the new digital media ecosystem founded on social media and the smartphone ought to be reframed as a social institution that operates on market principles. Within that system, acts of communication, i.e. posts on social media, are bought and sold not mainly for hard currency, but for attention, as measured by engagement metrics such as views, likes, and reposts.

The idea is to reconceive of social media as a marketplace patronized by ‘sellers’ of communication across different mediums (so of images, of video, and so on) and ‘buyers,’ who disburse some quantity of attention to those sellers in exchange for their ‘content,’ content being our new bit of vernacular for saleable communication of this kind.

But in what sense is content really “bought and sold?” Should this claim be read literally? In my opinion, yes. Attention as it exists online is a form of money. It has all of money’s defining properties, namely being a medium of exchange, a unit account, and a store of value. Running through these in turn:

Attention is a medium of exchange because we give it away to others and, in the case of social media, get content in return. There is no way to engage with a given piece of content, absent giving up some attention to its creator.

Attention is a unit of account because it is measured so as to, in effect, ‘value’ pieces of content. This is what engagement metrics do—they quantify how much, and also what kind of attention a given piece of content has received. They price that piece of content in attentional terms. Although, unlike regular prices, attentional prices don’t convey information about the cost of consumption to the consumer—the time any one person must spend to engage with the content at issue won’t be reflected in its engagement metrics, only how much attention in aggregate it is has received. This an important difference, but it doesn’t mean attentional prices aren’t genuine prices. All it means is that they aggregate different information to prices in the traditional economy.

Attention is also—although I expect somewhat more controversially—a store of value. When a given creator receives a certain amount of attention, they can reuse it later. Now, it is of course not their attention—content creators are not vampires sucking attention from their followers, so that they can live longer and disburse more attention themselves—but they can still disburse other people’s attention in a manner analogous to how an individual might spend money they first received from someone else.

They can do this, for example, by telling their followers to watch something made by another creator, or by showing their followers an ad. Granted, the quantity of attention they have ‘stored’ is volatile; it is not an amount that linearly increases with each additional view they get. Rather, the amount of attention they ‘have’ is a kind of projection of how much they command at any given amount. We might model this amount by asking how much attention a given creator could direct, at a given time, to an arbitrarily chosen piece of content. That is, very roughly, the amount of attention they have accumulated. This is a case where attention behaves very differently from other forms of money, but the difference is not, I think, large enough to prevent attention from being a store of value.

The stored value of hard currencies is hardly stable itself. 1 dollar this year is not 1 dollar last year—that the quantification is the same does not mean the two dollars’ underlying ‘real’ value is. Attention, in the same way, can be inflated and deflated. 5000 followers this year may not amount in ‘real terms’ to 5000 followers last year, if say, these followers have a lower propensity to consume other content the followed creator directs them to, or if they disburse less attention online this year, for whatever reason, as measured by other engagement metrics.

This gets at the question of what ‘backs’ attention’s value. Interestingly, attention’s most basic form of real value, unlike hard currencies,’ is not dependent on convention. Hard currencies are valuable, at least in the modern day, because states accept them as tax payment and oblige private individuals and firms to accept them in exchange for goods and services. But attention is valuable at least in part because we enjoy it intrinsically for social, psychological, and biological reasons. So in this respect attention’s value is more stable than hard currencies’ value. We might say it is less stable in the short run and more stable in the very long run. The real value of the attention a given creator has ‘stored’ at any given time may be somewhat opaque, but we know for certain that attention as such will always have real value as long as we are the sorts of organisms who crave it.

One issue I neglected to discussed in LHM is how content markets ‘clear.’ This is a perplexing question at first glance, since content is not consumed in anything like the way traditional goods and services are. It is not a rival good, and it can’t be depleted or used up. There is no sense in which, once content supply equilibrates with content demand, content inventories are emptied. Rather, content is just endlessly created and never strictly exits the market; it is priced and repriced for eternity as it is consumed repeatedly by different people. Depending on how we frame this process, that means content markets either never clear, since content can always be re-priced by a future consumer’s engagement, or are perpetually clearing, since they are being consumed over and over, albeit at different price points.

We could of course imagine an idealized content market in which a static amount of content is created and consumers have a static amount of attention to give. Then markets do clear: at a certain point consumers will “die,” having disbursed all their available attention, and there will be a final distribution of attention across all the content in the market.

Another more complicated model might give both creators and consumers time allowances for content production and attention disbursal, respectively, and examine what clearance looks like in an environment where creators can choose to produce different amounts, or even types of content, at the cost of different amounts of time.

Exploring the behavior of such models with traditional economic methods might prove interesting, but in realistic content markets, the supply of and demand for content are extremely elastic (at least in the upward direction). New content is always flooding the market and new consumers are always joining social media, and pricing and repricing both old and new content through their engagement.

This gets at a few other interesting feature of content markets: consumers, without exception, set all of the prices, the process of price-setting and consumption are perfectly coextensive, and price determination is merely additive. Content creators obviously don’t set the level of engagement garnered by their posts in advance, in the way that firms set their own prices (even if only by choosing to sell at the prevailing market price). Consumers set all the prices and do so through their consumption: it is how, and how much they ‘pay attention’ to certain posts that determines those posts’ engagement metrics, which amount in this context to their price (or at least something like their ‘attentional value‘). In content markets, there is no price-setting besides this form of consumption, and there is no consumption besides this form of price-setting; consumption and price-setting are metaphysically intertwined. The process of price discovery is therefore much simpler than in traditional markets. Rather than reflecting the complex convergence of initially disparate bids and asks, prices are just engagement metrics which rise as more people consume the content at issue.

In summary, digital content markets are large, diverse, transparent, procedurally efficient social systems for incentivizing the creation of, and determining the ‘attentional value’ of discrete instances of communication.

This makes sense given the superordinate business function they serve for the major platform companies: to keep people on their platforms, so that those people can be shown ads. Platforms are markets that incentivize the generation of maximally engaging content. They offload the process of attracting consumers to advertisements to consumers themselves, who are compensated for keeping each other on platforms through their content by the attention they receive in return for it.

In LHM, I argue that our being corralled in large numbers ‘into’ these markets has had serious social costs over the last fifteen years, ones that speak not just to the temporary effects of technological change, but to changes of the more tectonic kind we associate with the formation of major social institutions like states or traditional markets.

If you’ve found this short primer valuable, please consider giving the original essay a read.

I am somehow both perplexed and woefully unsurprised that the Democratic party establishment is “melt[ing] down” over Zohran Mamdani’s victory over Andrew Cuomo in New York city’s democratic mayoral primary.

The puzzlement stems from feeling that figures like Mamdani, at least ones with similar personability and acumen for populist, digitally savvy rhetoric are the best hope for the democratic party going forward.

This is a party, recall, whose rank-and-file members—roughly 60% of them—believe “the leadership of the Democratic Party should be replaced with new people.” The party’s overall approval rating sits at a record low of 29%. That big name democrats like Bill Clinton, Mike Bloomberg, and Jim Clyburn endorsed Cuomo did nothing to prevent his “thrashing” by a 33-year old Muslim socialist with zero mainstream party credentials to speak of.

Mamdani’s win also comes on the heels of Trump making historic inroads in normally solid-blue New York, suggesting that the president’s relative success not just there, but in other progressive strongholds, may have more to do with broad anti-establishment sentiment, rather than expressly conservative sentiment among voters.

This should all excite leading Democrats: Mamdani has shown that they can notch convention-defying wins by melding disciplined emphasis on cost-of-living with an interlinked, state-of-the-art social media and ground game.

Any proven contribution to revamping the Democratic playbook in advance of 2026 and 2028 ought to be welcomed with open arms. Senior party members have so far failed to internalize the fact that the political mood of the time is anti-establishment to its core. They have also failed to see the sinews connecting this fact to the new digital media environment, and to infer from there that it is a grave error to rebuff opportunities—as Harris did in 2024—to appear on shows like Joe Rogan’s, or otherwise commit to a high-visibility social media presence. Doing so would mean embracing both unscripted, longer-form interviews as well as punchier, short-form video content, as Mamdani did to great effect.

I recently argued that our new digital media ecosystem is akin to a grand new social institution, one set to delineate the shape of public discourse in this country for the foreseeable future. The sooner the Democratic party wakes up to the newfound influence of social media, as Republicans did years ago, the sooner they can reestablish themselves as a potent political force in the eyes of the large number of Americans fed up with mainstream politics.

For all the Republican party has to thank Trump for, his most underrated gift might be obviating their need to grapple, in their own antiquated terms, with adapting to the new cultural climate of the 2010s. In reshaping the party more or less by force, Trump headed off a counterfactual scenario in which senior Republicans might’ve been as hopelessly lost as senior Democrats are now.

Mamdani’s win in New York is an invitation for Democrats to find their way again by learning from the younger, more progressive, more digitally savvy wing of the party. Doing so doesn’t require embracing every, or even any of Mamdani’s policy platform, much of which is likely to prove difficult to implement, should the 33-year old become mayor. But it does mean taking seriously the potential of backing, or at least emulating younger, fresher digital natives with a populist message—even if that means passing over more senior party figures that see themselves as ‘next in line.’

Note: I’ve decided to add a TLDR to the top of this post, both as a pitch to encourage folks to read the whole thing, and as a partial substitute for those who don’t have the time. For those who found my work via my reply to Dan Williams (“The Case Against Social Media is Stronger than You Think”), this is a less polished proof-of-concept of some bigger ideas I’d like to develop more carefully in the future. It also contains earlier versions of that post and my more recent one on social media and adolescent mental health.

The basic argument is that we need to reframe the new digital media ecosystem made possible by social media and the smartphone as a nascent institution that is, in effect, ‘modernizing’ civil society. I see this process as analogous to how the now-familiar institutional forms of ‘the state’ and ‘the market’ once modernized other important domains of human affairs (roughly, the domain of top-down governance and the domain of goods-production and exchange).

The idea is that we are living through a similarly important episode of modernization in the realm of more informal communicative exchange encompassing media, entertainment, academia, NGOs, political-activist groups, and so on, as well our private social lives; in short, most everything outside of ‘government’ and ‘business.’

Civil society writ large, or what is sometimes called ‘the public square,’ is being compressed into a globalized, direct-democratic, hyper-competitive market for ‘content’—our new term for discrete instances of digital communication—hosted on the major social networking platforms, these amounting together to a kind of semi-unified institutional environment (certainly no less unified than the analog market for goods and services). Within this environment, producers of content ‘sell’ their wares, and consumers of content ‘buy’ them by expending some quantity of attention. Content-instances (‘posts’) are in turn ‘priced’ (or ‘valued,’ or ‘ranked’) according to how much attention they garner.

In other words, the rise of social media and the smartphone has remolded not just the internet, but increasingly civil society as such into a more horizontal, more structured, more surveillable digital marketplace for communication, one that values its chief commodity—content—via metrics that measure attention.

My claim is that most of the major effects of social media and the smartphone—on face-to-face socializing, on education, on politics, on mental health—are helpfully understood through this lens, this ‘institutionalist’ account of social media and the smartphone. It is also that this new institution is having variously corrosive effects on the other legacy institutions that make up civil society.

My argument is partially inspired by the political scientist James C. Scott’s work on high-modernist institution building and its often disastrous effects (as in e.g. Soviet collectivization, forcible “villagization” in Tanzania under Julius Nyerere).

The point is not that social media is as morally awful as e.g. Soviet collectivization, but that it represents an analogous kind of dangerous high-modernism in action, albeit one enacted by a network of private companies rather than the the state.

As in James’s case studies, the new institutional form founded on social media and the smartphone—which I elect below to call ‘the Forum’—is corroding the tacit knowledge associated with the forms of social organization that have preceded it, knowledge we have relied on to build social bonds, to educate ourselves, to cooperate politically, and broadly to live well together.

I see this particular line of work as being in its early stages, and so am eager for any and all feedback. I appreciate that the essay below is long, but it’d mean a lot if you gave it a read, and even more if you got in touch with your thoughts / reactions! Thanks.

1. Introduction

Over the last fifteen years, it has become an article of common sense that American institutions are under strain. The dysfunction extends not just to formal institutions—government agencies, legacy media, the electoral system—but also to cultural ones, ranging from reflexive optimism about the country’s global influence to a once widespread faith in democracy. Virtually all of the above have suffered a massive and concurrent loss of public confidence. There is broad agreement that the country has fallen into a pessimistic, angry morass, if not a more existential downward spiral. The reelection of Donald Trump—and the flood of disruptive changes he and his allies have already made in Washington—has only reinforced the idea that something deep has gone awry in America’s collective psyche.

As sweeping as this regression has been, though, it also poses an explanatory puzzle. There is little evidence to suggest, for instance, that it is the result of declining material wellbeing. Though Americans were justifiably upset by the Covid-era inflation as well as the expiration of its attendant increases in social spending, and while they continue to face salient cost burdens at the grocery store, in the housing market, and in their pursuit of college educations, their finances have nonetheless reliably improved since the ’08 crash, whether in terms of median income, net worth, or hourly earnings. That’s not to deny there have been pockets of rising hardship, but considered in aggregate, the American economy is, without exaggeration, the envy of the world (note: at least it was; this essay was published before Trump’s Liberation Day Tariffs). The twin narratives of a country hollowed out either by inflation or globalization are almost entirely myths.

Matters of international politics don’t carry much explanatory weight either. Though a small subset of Americans is wracked with ‘climate anxiety,’ that this might account for widely distributed misgivings about the country’s politics and culture is inconsistent with the public’s relative lack of concern for climate change. Replace it with other global crises—war in Ukraine, war in the Middle East, economic and geopolitical competition with China, recent advances in AI, falling global fertility rates—and a similar argument follows. It is an unfortunate fact that these issues are low on most Americans’ priority lists. That isn’t because they’re unimportant; the polar opposite is true. But if Americans profess not to care much about global issues, then those issues probably don’t underly our grim national mood.

It is instead time we take more seriously another major candidate explanation for the last fifteen years of disharmony. As Democratic strategist James Carville once said of the economy, so his successors may very well say, teeth no doubt gritted in painful retrospection, ‘it’s the internet, stupid.’

Since its state-sponsored birth as ARPANET, the internet has transformed human social experience in strikingly short order. But it has also gone through transformations of its own, not least those represented by the more recent rise of social media and the smartphone. In the aftermath of the public releases of Facebook in 2006 and the iPhone in 2007, social media and the smartphone grew to become, in effect, privately owned utilities, with the former acting as the de facto public forum of the digital age and the latter acting as its primary access point, one uniquely tactile, mobile, and ever-present.

90 percent of Americans now own a smartphone and use it (on average) 5 hours a day, and 70 percent are on social media and use it (on average) 2.5 hours a day. Teens spend 1.5 hours on their smartphones during the school day alone, and almost 5 hours a day on social media. Half report being online “almost constantly.” And relative to the normal pace of technological diffusion, this all happened in an instant. The smartphone and social media are two of the fastest adopted technologies in human history.

The public discussion they’ve inspired has quite understandably treated them as just that—technologies. Among other things, smartphones and social media have been called “communications technologies,” “information technologies,” “social networking technologies,” and “digital media technologies.” Hence comparisons to television, videogames, the personal computer, the (analog) telephone, and even the printing press. But this lens is incomplete. It risks downplaying the sheer scale of the social transformation social media and the smartphone have brought about, a transformation that greatly exceeds the impact of any of their predecessors.

Social media and the smartphone haven’t just created new kinds of communication. They’ve created a new communicative institution, new ‘rules of the game’ for communication as such. Just as the state and the market have, over their lengthy modern histories, claimed imperfect monopolies over distinct realms of public life, so the new digital ecosystem created by social media and the smartphone—what I will call “the Forum”—has begun to stake its own such claim over the once hallowed realm of the public square.

It has done so by routing an ever greater share of written, oral, and visual communication through a narrowly optimized sieve: a digital marketplace for communication designed above all else to sate major tech conglomerates’ hunger for advertising revenues. The Forum is rapidly replacing networks of dialogue concentrated at the local level—supplemented, of course, by carefully gatekept national and supranational networks in media, academia, and government—with a flattened, hyper-competitive, global market for ‘content,’ one in which acts of communication are bought with, and sold for human attention.

In a historical instant, this new social architecture has transformed American life almost beyond recognition. Among other things, our social lives, education system, politics, and mental health are being degraded along multiple fronts at once, with few signs of reform or slowdown in sight.

It is well past time that we see our new digital media ecosystem for what it is: a systemically significant institution in its birth-throes, one corroding precisely those preexisting institutional and coordinative capacities necessary to control it.

2. Clearing the forest

The political scientist James C. Scott begins his masterwork, Seeing like a State, with a parable that nicely captures our current predicament. Scott’s story is a long one, a succession of tragedies over hundreds of years and multiple continents in which the blunt imposition of order on complex, naturally occurring systems led to disaster. His first case study concerns (bear with me) the early application of scientific forestry in eighteenth-century Europe.

At the time, Scott writes, “the … European state viewed its forests primarily through the fiscal lens of revenue needs … [in particular] the revenue yield of the timber that might be extracted annually” (p. 11-12). This was of course antithetical to the conflicting ‘viewpoint’ of nature itself. The forests were filled with complexly interactive, but fiscally useless flora and fauna, including many of in-kind value to human poachers, fishers, and farmers. But since none of this seemed pertinent to the state’s timber interests, foresters were made to clear and replace the forests with what Scott describes as “serried, uniform ranks … to be measured, counted off, felled, and replaced by a new rank and file of lookalike conscripts” (p. 15). Extraordinarily lush ecologies, in other words, were replaced with something akin to industrial timber farms.

As it turned out, this program of unbridled demolition and artificial recreation was a disaster in waiting. The foresters had underestimated how interdependent timber yields were with the parts of the forest they’d destroyed, such as the once prevalent underbrush and animal life, which, among other things, had maintained the nutritional qualities of the soil, increasing yields beyond what would otherwise be possible. The practice of monoculturing—only planting one variety of tree—also meant the redesigned forests were easy prey to storms and disease, threats more diverse crops are better able to resist.

The result was that, in the worst cases, whole forests withered away. The Germans coined a new term, Waldsterben—'forest death’—to describe the resulting desolation (p. 20). This of course came at enormous cost to the state, not to mention the plants, animals, and other humans who relied on the resources the forests provided. In a dynamic that would be repeated throughout modern history, arrogant men with dreams of panoptic control destroyed systems whose beneficence depended on their self-propelling, organic integrity.

Scott is very explicit about the moral of this story: “[it] illustrates the dangers of dismembering an exceptionally complex and poorly understood set of relations … in order to isolate a single element of instrumental value” (p. 21).

Perhaps this sounds familiar. Scott might well have said the exact same thing about America’s last fifteen years. Like the forests in Scott’s parable, our civil society is it itself a vibrant and dizzyingly complex ecosystem. It has persisted at its healthiest on the informationally rich bonds of face-to-face interaction, a foundation of widely accepted liberal norms, and the guiding hand of a diverse subset of publicly accountable elites acting in good faith—albeit not flawlessly or in the most glamorous fashion—to advance the public good.

Over the last decade and a half, major tech corporations such as Google and Meta have begun to supplant this now outmoded vision of American life with their own, one designed, as with the state and its timber yields, to maximize their ad revenues above all else. What early scientific forestry was to the forest, ‘the Forum’ is to our social reality. And fittingly, we are in the midst of our own version of Waldsterben.

3. The discourse economy

In one sense, the Forum is something we are all acutely familiar with: the digital ecosystem of major social media and entertainment platforms, such as Facebook, Instagram, Twitter, TikTok, YouTube, and Twitch, as they are accessed via smartphones in particular, but also tablets, computers, smart T.V.s and so on. But it is also something more.

Far from just referring to the platform economy as conventionally understood, ‘the Forum,’ at its core, refers to that economy’s distinctive nature as a set of digital markets for communication monetized by attention. Within these markets—the major social networking platforms—users ‘buy’ and ‘sell’ instances of public communication across modalities (i.e. text, image, video, and audio), not with traditional currencies, but their attention, as quantified by their time, views, likes, reposts, and comments. Advertisers in turn pay tech companies in hard currency to ensure that a large portion of this attentional pool flows to ads. This is the business model that made platform companies the titans they are today. For reference, advertising accounts for 97% of Meta’s revenue, 80% of Google’s, and 80% of TikTok’s.

Not unlike other markets that have arisen in analog environments, certain aspects of these systems are so ingenious they couldn’t possibly have been planned from the beginning. Platform companies have not only accumulated a vast captive audience for advertisers to pick off, like fish in a barrel. They have also tasked that audience with competing, at their own cost, to produce whichever media most effectively keeps them in that barrel.

From advertisers’ perspective, platform companies and their users jointly perform the functions once assigned to T.V. networks. The key difference is that whereas networks once produced programing that would attract consumers so that they could be shown ads, now the consumers do everything themselves: they create the programming and watch the ads. Instead of entertaining the public at their own cost, platforms in effect host massive contests in which members of the public compete to entertain each other. As evinced by the enormous profit margins of the major tech giants, this yields a large economic surplus, one that can be traced back to externalizing various costs of production on the public at large.[1]

That does not mean, however, that the content-creating public is working for free. The platform economy’s genius lies precisely in the fact they are not. Rather, instead of paying their billions of de facto workers in a preexistent currency, platform companies performed an economic magic trick: they created a new one. Creators are paid, not in dollars, not in crypto, but in attention.[2] On the digital content market(s), attention has all the defining properties of money:

Attention is a medium of exchange because individuals must ‘spend’ some of their scarce attentional resources by definition to engage with a given piece of content. On the ‘sell side,’ content creators are incentivized to make their content by the attention they receive in return for it.

Attention is a unit of account because it is measured so as to, in effect, ‘value’ pieces of content. This is what engagement metrics do—they quantify how much, and also what kind of attention a given piece of content has received. They price that piece of content in attentional terms. Although, unlike regular prices, attentional prices don’t convey information about the cost of consumption to the consumer—the time any one person must spend to engage with the content at issue won’t be reflected in its engagement metrics, only how much attention in aggregate it is has received. This an important difference, but it doesn’t mean attentional prices aren’t genuine prices. All it means is that they aggregate different information to prices in the traditional economy (the question of whether this stretches the notion of ‘price’ too far I put to the side for now).

Attention is also a store of value. When a given creator receives a certain amount of attention, they can reuse it later. Now, it is of course not their attention—content creators are not vampires sucking attention from their followers, so that they can live longer and disburse more attention themselves—but they can still disburse other people’s attention in a manner analogous to how an individual might spend money they first received from someone else. They can do this, for example, by telling their followers to watch something made by another creator, or by showing their followers an ad—something they can get paid for in hard currency.

Much else we expect of money we can also expect, in some form, from attention as it is exchanged in the Forum. We regularly ask of other forms of money, for instance, what ‘backs’ their value. When we ask this question of attention, we find that its value is backed by the fact humans intrinsically enjoy it for social, psychological, and biological reasons. Attention’s value is also backed by hard currencies (and so transitively backed by states), since platforms pay creators in hard currencies in proportion to the attention they get, and advertisers pay platforms in hard currencies to direct some of that attention to ads.

Similar examples abound. The interesting notion of an ‘attentional recession,’ for instance, is also suggestive of the parity between attention in the Forum and traditional currencies. If, for whatever reason, many people left social media, this would mean less attention to go around, and so creators might be more hesitant to direct their remaining attention away from themselves. They might collaborate less with other creators, for instance, and focus more on their own siloed projects. They might also take fewer risks, and focus more narrowly on doing whatever it is their viewers would like to see most. These behaviors are analogous to how individuals and firms tighten their purse strings during traditional market downturns.

There is much more to say here that would take us too far afield. Whole theoretical edifices might be built making sense of such ideas as attentional credit markets, capital investments, banking institutions, and so on. In any case, the point is that since attention in the context of the Forum walks and talks like money, we shouldn’t hesitate to treat it as money proper.

In similar fashion, the transient acts of communication typical of the Forum—i.e. ‘posts’ on social media—can be treated as ‘commodities’ of a kind, purchasable at the cost of some quantity of attention. Like commodities classically defined, posts on social media have both ‘use’ and ‘exchange’ values. That is, they have concrete properties we desire, such as being funny or informative, as well as abstract ones that quantify convertibility into other things we desire, like views, likes, reposts, and so on—metrics we can think of as together constituting a multidimensional ‘attentional price’ (or if this is too confusing, since engagement metrics don’t represent the cost of engagement like prices normally would, ‘attentional values,’ or ‘ratings’).

Put another way, while posts range widely in medium and significance, they are still all ‘content,’ in its cynical contemporary sense. What is ‘content’ in the way we now use this term if not an instance of communication with a semantic ‘use value’ of secondary importance to its attentional ‘exchange’ value? ‘Content’ is to the Forum as the commodity is to capitalism. While communication has an inordinately rich variety of ‘uses’ it is ultimately the attention it garners that determines its ‘price’ online, not its capacity to transmit sympathy, report facts, express happiness, or whatever else. One can of course ‘purchase’ content that does any one of these things, but it is not of incidental significance that, rather than arising in the course of some organic conversation, it will have been for sale.

Projecting these developments forward, there is good reason to think we are on the cusp of a transition very similar to that experienced by the premodern exchange in goods. Premodern societies exchanged goods for a much richer variety of reasons—to mark rites of passage, to welcome outsiders, to send diplomatic signals, to pay feudal or religious tribute—only for the vast majority of these to be displaced by the formalized, Pareto-improving market transaction. The rise of the Forum heralds something similar with respect to exchange in communication.

It is not that the old communicative practices will cease to exist. Even in the wealthy Western world, the practice of gift-giving survived modernity. But just as we do not know, or care to know most of the people we buy goods (and services) from, just as we often occupy completely different normative, cultural, and geographical spheres from them, our communicative behavior on the Forum—a rapidly rising proportion of our communicative behavior as such—is impersonal and transactional. When we shop around on the Forum, the subtext is: I give you attention; in return, you give me what I want to hear.[3] The old ways of communicative exchange—the kinds we associate with such ideas as democracy, community, and friendship—are on course to occupy much smaller niches than they once did. In hindsight, this is a natural development: why do the messy work of sustaining these things when you can frictionlessly see, hear, or read anything you want instead?

In short, the rise of the Forum represents much more than our adopting a new set of communications technologies. What it represents is the rise of whole new parallel economy for trade in discourse, one undergirded by the commodification of communication and the monetization of attention.[4]

4. The Forum and its social costs – a stylized review

As with similarly significant economic transitions, the rise of the Forum has imposed large costs on preexisting forms of social life. Some of these costs are associated with the creation of the Forum, and some with the dynamics internal to it.

More concretely, the Forum can be understood as a pair of distinct, but mutually supportive ‘layers.’ There are a set of mechanisms driving social behavior out of analog communication networks and into the digital ones represented by the major social networking platforms; and then there are a set of mechanisms responsible for the kinds of content created and promoted within these platforms. Call these two layers the Forum’s ‘exterior’ and ‘interior,’ respectively.

On the Forum’s exterior, the major tech corporations (Apple, Google, Meta, TikTok, etc.) have employed a range of often manipulative strategies to cement our individual and collective dependence on their products, with serious consequences for our social lives, educations, and cognitive faculties. In the Forum’s interior, that is, on social networking platforms themselves, the same companies have created a hyper-competitive environment that rewards attention above all else, leading to a profusion of content designed to induce fear, anger, and status anxiety, with serious consequences for our politics and mental health.

What follows is a stylized review of the Forum’s major social costs that, while not comprehensive, should at least bear out their reality and their urgency.

4.1. A visit to the exterior

Pursuant to the creation of large, liquid markets in content, the major platform companies have competed aggressively for the past decade and a half to make their products as dependency-inducing as possible, trying both to outdo each other and the attractions of the outside world. That competition is only continuing to intensify as their market broadens and technological change unveils dark new vistas in the art of user manipulation.

Most platforms have embraced a now-standard set of addictive design features, such as ‘infinite scroll,’ frequent push notifications, and algorithmically tailored content feeds.

Infinite scroll in particular—the practice of displaying new content in unending succession, rather than in paginated form—has played a central role in fostering excessive dependence on social media, and its main conduit, the smartphone. The vast majority of users, whether they prefer to or not, engage routinely in what has been called ‘dopamine-scrolling,’ the trance-like habit—named for the neurotransmitter’s connection to reward and novelty-seeking behavior—of swiping through one’s phone for overlong periods in search of new content.

The technology ethicist Tristan Harris has compared infinite scroll to playing a slot machine because of the variable reward schedule associated with scrolling over and over in search of new information. As with slot machines, the fact that rewards are intermittent, rather than constant—not every post is something you want to see—makes them more likely to lead to compulsive use, in line with previous research on gambling.

If infinite scroll and its other addictive counterparts weren’t bad enough, platform companies are now preparing to take advantage of new AI capabilities in text, image, and video generation to make their products even more enticing. Meta, for instance, is exploring the possibility of filling Facebook and Instagram with fake, interactive users, and offering personally tailored AI “friends” to meet what Mark Zuckerberg has described as unmet demand for personal companionship.

Beyond their addictive design, platforms also owe their success to network effects, the fact their products become more attractive the more people use them. When social media was still in its infancy, there were few disincentives associated with logging off. But now because of social media’s colossal scale, failing to continue using it means losing contact with friends and family, not to mention the wider cultural zeitgeist.

It turns out that this disincentive is so strong that even users who are being harmed by social media will opt to continue using it. Bursztyn et. al. (2023) found that large swathes of social media’s user base are caught in what the authors call a “product market trap,” a situation in which consumers are made worse off by a good or service, relative to a world where it doesn’t exist, but would prefer to use it rather than not in a world where it does. In social media’s case, while users wish they could log off, the benefits of doing so are outweighed by the costs of social exclusion.

The prospect of any notably large group detaching themselves from social media is made even more unlikely by the fact that, rather than living in a faraway realm only accessible from the fixed point of a desktop computer, our content feeds are with us at all times, bleating their Siren calls from our pockets and the palms of our hands. While the role of platforms in our current social transition can’t be overstated, their effects would be much more muted absent support from the smartphone.

Among all the innovations in consumer electronics from the turn of the millennium onward, it is this one that is most responsible for our new more digitized social reality. Smartphones have become almost cyborg-like extensions of our bodies and minds, granting us access, wherever we might be, to a swirling hivemind of humanity’s knowledge, sentiments, and self-amusements—a double-edged privilege that, once one gets acclimated to it, can be very difficult to go without. While some of the time we now spend on our phones might otherwise have been spent watching T.V. or using the computer, the smartphone has buried itself much deeper in our psyches than these technologies ever could.

Melumad and Pham (2020) found as much in their aptly-named paper “The Smartphone as a Pacifying Technology.” As they put it: “smartphones are endowed with a unique combination of properties that lead them to be viewed … as sources of comfort for owners—not unlike pacifiers for children” (p. 239). Those properties include their “portability, highly personal nature, the sense of privacy [they invoke] when engaged, and the haptic pleasure users derive from handling their device” (p. 251). Across a large-scale survey, two structural equation models, and three different experiments, the pair found substantial evidence that smartphone users use their devices as palliative ‘comfort objects’ in stressful moments, and that this effect was unique to smartphones as compared to other electronics like laptops. Consistent with these results, study after study shows that when even moderate smartphone users are separated from their devices, they experience significant[5] and persistent spikes in anxiety.

4.1.1. On social displacement

It is no wonder, given our almost childlike attachment to these devices, as well as the addictive properties of the platforms they help us access, that smartphones and social media have begun to displace face-to-face socializing on a national scale.

This is borne out by a large body of statistical evidence. Twenge et. al. (2019) found that adolescents began socializing in person significantly less throughout the 2010s, in close correlation with their increased use of digital media. The Atlantic’s Derek Thompson, analyzing data from the American Time Use Survey, found that in-person socializing declined by more than 20 percent from 2003 to 2023, with that number rising to 35 percent for unmarried men and people younger than 25. Across roughly the same period, Sharkey (2024) found that the time Americans spent home rose by an hour and 39 minutes a day.

These changes are being felt in the fraying of important social bonds. The American Survey Center (2021) found that almost half of Americans had three or fewer close friends in 2021, compared to only about a quarter in 1991; this despite the fact that they are of course interacting through the internet with more of their peers than ever before. As the Center put it, “the role of friends in American life is experiencing a pronounced decline.” We are living through what the Surgeon General’s office has called an “epidemic of loneliness and isolation.”

Note: if you are already convinced that these are consequences of smartphone and social media use, feel free to skip to the next section; the rest of this one is dedicated to rebutting a few remaining skeptics in academia.

The evidence of a link between these trends and digital media use is not without controversy. Skeptics will be quick to point out that none of this research establishes a causal link between social media or smartphone use, and the frequency with which we socialize face-to-face. Twenge, Thompson, and others in their argumentative camp, like social psychologist Jonathan Haidt, can at best point to strong correlations between the increased use of these technologies, and reductions in face-to-face socializing, but nothing more.

Not only that, but there is a large body of evidence running in other direction. Much of the recent literature on what is commonly called the “social displacement hypothesis” has found that neither social media use writ large, nor smartphone use in particular displace face-to-face socializing. For instance, one prominent experiment that assigned varying degrees of social media abstinence to its participants found that instead of spending their extra time with friends, they spent it on other internet use, work, and chores. This would seem to raise the specter of irrational moral panic. Maybe social media is a welcome new extension of our larger social lives, rather than a hollow substitute.

On deeper investigation of the displacement literature, though, it becomes clear that its claim to have cast doubt on the link between digital media use and (reductions in) face-to-face socializing is misleading at best. Research on social displacement almost uniformly compares individuals within the same social media and smartphone-native cohort, i.e. teens and young adults all sampled in the recent past. But the question of whether heavier social media or smartphone users do less face-to-face socializing than their immediate peers is not the only question at issue. In fact, it is arguably not the main question most participants in the public, rather than the academic discourse are interested in. The main question is whether the entire cohort of longtime social media and smartphones users does less face-to-face socializing than previous cohorts who lacked these kinds of media altogether, and the answer to that is unequivocally yes.

To be sure, skeptics are right to observe that the research supporting this claim is largely non-causal. But one reason for that is the practical challenge of designing a study that models the large-scale changes brought about by the initial introduction of social media and the smartphone. What one would ideally like to do is a massive randomized control trial whose participants had neither used social media or smartphones themselves, nor interacted with others who do. This group would be randomly divvied up into treatments and controls, the treatments would all be given smartphones and social media accounts (and also somehow cordoned off from the controls), and then both groups’ behavior would be monitored over—again, ideally—several months, if not years. This would be a decent proxy for what happened to Americans in the late aughts and 2010s.

An experiment of this kind, though, would face huge practical and ethical challenges, and so causal and correlational research alike have had to occur within cohorts already heavily exposed to digital media. While this is unavoidable, what isn’t is the problematic form of intellectual displacement it might inspire. We should not fool ourselves into thinking that differences within cohorts are a good guide to differences between them. That is in reality rarely the case.

The literature on social displacement is itself a demonstrative example. It tells us that among those who use social media and smartphones, heavier users do not socialize in person less than lighter ones. But there may be incidental reasons for that. For one, extroverted people use social media more, and also socialize more in-person, meaning that even if displacement does occur it may be eroded in the data by a small group of especially outgoing people. Cross-cohort data solves this problem by telling us whether extroverts and introverts alike socialized more in person before the rise of social media and the smartphone than they do after. Again, we already know that, at least on average, they did.

Another problem for within-cohort research is network effects. Once smartphone and social media use become ubiquitous, a single person’s cutting down on their own use may have little positive effect on their social life if their friends fail to do the same thing. These technologies are habit-forming, and since face-to-face interaction obviously takes two-to-tango, we learn little by finding out that one person’s social media abstinence didn’t lead, within a few short weeks, to a convalescent bloom in their social calendar. If their friends weren’t abstinent themselves, there’s no reason to think that would happen.

Oddly, some of the best available evidence for social displacement comes from a study claiming to find no evidence of it. Hall, Kearney, and Xing (2018) found, using data from the Longitudinal Study of American Youth (N = 2774), that a one standard deviation increase in participants’ social media adoption in 2009 was associated with a two standard deviation decrease in their direct social contact in 2011 (p. 6, fig. 1), a large negative effect by any account. Yet the authors wave this striking result away by claiming that, since the same increase in adoption wasn’t associated with decreased social contact in 2010, and because more frequent social media use in 2011 wasn’t associated with decreased social contact in 2011, the data “[do] not support the social displacement hypothesis” (p. 8).

In reality, though, this timeline would make sense if social displacement takes a long time to manifest rather than co-occurring with changes in short-term social media use, as the displacement literature unfairly assumes it must. Nonetheless, the authors conclude on the basis of the LSAY data, and a second set on passive social media use collected from 116 people over a paltry 5 days, that belief in social displacement is likely the result of “public perceptions,” rather than thorough research. (p. 15).

Further evidence in favor of social displacement comes from a specific strand of micro-level research from the literature on smartphones. One of the better-established results in this area is that using one’s smartphone during a given social interaction lowers the perceived ‘quality’ of that interaction by reducing subjective feelings of enjoyment, empathy, and/or connection. Combine this with the fact that 89% of Americans used their phones during their most recent social activity, and that the average user interacts with their phone every five minutes, and it becomes plausible that face-to-face interaction has become much less enjoyable over the last decade and a half, leading people to prefer spending time on other activities.

If this is one of the mechanisms behind social displacement, it further underscores why it is unfair to expect displacement to show up over shorter time spans. It wasn’t until several years after their introduction that smartphones became omnipresent in our social lives. It would’ve taken a couple more years at least for us to reallocate our time to account for the lesser relative appeal of socializing face-to-face. Moreover, even if heavier social media and smartphone users do not engage in fewer discrete face-to-face interactions, they may spend less time within those interactions actually socializing, as opposed to using their phones. In addition to comparing distinct cohorts, accounting for this more subtle form of smartphone-based displacement might yield very different results from the preexisting literature.

Though we have a lot more to learn, the balance of evidence favors that social displacement between cohorts, if not also within them, is very real. Over the last fifteen years, the frequency with which we socialize in person has plunged. There is a good chance that the relative enjoyment we get from socializing has fallen as well. Though certainty is hard to come by, it at least seems very likely that smartphones and social media deserve a large share of the responsibility.

4.1.2. On education and cognition

Social displacement is only one of the serious casualties plausibly attributable to social media and the smartphone. It is not only our relationships that have changed since the advent of these technologies, but our basic capacities for processing and navigating the world. In particular, there is worrying evidence to suggest we are having a harder time learning, and that this is not only a surface-level phenomenon, but one with deeper roots in our neurology.

Several meta-analyses on the effects of smartphone and social media use on educational attainment have found small, but significant negative effects across students’ GPAs and exam scores. While a lot of the underlying research is cross-sectional (i.e. based on data taken at a single point in time), the negative association appears in more robust longitudinal studies as well (although results are more mixed in social media’s case).

While some of these effects are likely due to social media and smartphones acting as distractions from learning, there is also evidence of longer-term impacts on cognitive functioning. Horvath (2020), for instance, found that smartphone addicts—which make up 12-28% of the user population (pg. 3)—showed significantly lower grey matter volume in the insula, and significantly lower activity in the right anterior cingulate cortex (ACC), regions of the brain associated with craving and addiction. The ACC, for instance, is associated with inhibitory control and tends to show lower activity in those with substance use disorders experiencing withdrawal (pg. 11). In other words, especially heavy smartphone users appear to exhibit structural neurological differences similar to those whose capacity to override strong impulses or emotions has been weakened by substance abuse. ERP studies (which use EEG to measure neural responses to specific stimuli) on both excessive smartphone and social media users have also found that both groups perform worse than non-excessive users on tests of inhibitory control.

We should note of course that these results are only correlational. They may suggest that those who already have poor inhibitory control are more likely to become excessive smartphone or social media users. But what little experimental evidence exists on this question suggests causality runs the other way too. In one of the only randomized control trials to date on smartphone use and cognition, Hadar et. al. (2015) gave smartphones to a group lacking any previous experience with them for three weeks and compared their performance on a suite of cognitive tests to another control group of nonusers. Those in the treatment group “showed a significant decrease in information processing capacity compared with their baseline [before the intervention, while] no such effect was observed in the control group.” The treatment group, but not the control also exhibited “greater delay discounting” after the treatment, meaning an inability to delay gratification in pursuit of a larger long-term reward (p. 318). While this was study was small (N = 38), its high-quality design makes it worthy of close consideration.

Even supposing smartphones don’t have shorter-term neurological effects, and are no more harmful than other screen-based electronics, the simple fact of their increasing the amount of time children spend looking at screens could be the cause of enormous harm by itself. That was the conclusion reached by Manwell et. al. (2022) in a review of 37 studies in the Journal of Integrative Neuroscience. The authors’ central warning is worth reading in full:

We hypothesize that excessive screen exposure during critical periods of development in Generation Z will lead to mild cognitive impairments in early to middle adulthood resulting in substantially increased rates of early onset dementia in later adulthood. We predict that from 2060 to 2100, the rates of Alzheimer's disease and related dementias (ADRD) will increase significantly, far above the Centres for Disease Control (CDC) projected estimates of a two-fold increase, to upwards of a four-to-six-fold increase … [enough to] result in widespread societal and economic distress and the complete collapse of already overburdened healthcare systems in developed countries (pg. 1, my italics)

The authors note that the disparity between the CDC’s estimates and their own is driven by the former’s basis in “factors related to the age, sex, race and ethnicity of individuals born before 1950 who did not have access to mobile digital technology during critical periods of brain development.” They add that, “compared to previous generations, the average 17-19 year old spends approximately 6 hours a day a on mobile digital devices … whereas individuals born before 1950 at the same age spent zero” (pg. 1). In other words, because the CDC projections are based on the number and demographics of current Alzheimer’s patients, they fail to account for the unique risk factors faced by those in the future from growing up amid heavy digital media use.

Recent data on literacy and numeracy lends credence to the idea that Generation Z is on track for at least “mild” cognitive impairment. 9 and 13 year olds’ scores on the National Assessment of Educational Progress—often referred to as “the Nation’s Report Card”—have seen historically unprecedented declines in reading and math in recent years, and while they’ve been steeper since 2020, the overall trend precedes the onset of the Covid-19 pandemic.

The same thing has happened across the OECD, with reading, math, and science scores on the Program for International Student Assessment, or PISA, also showing unprecedented declines, all of which line up disturbingly well with the 2010s era social transformation brought on by social media and the smartphone. Nor have American adults been spared, with the most recent round of PIAAC results showing marked declines in literacy and numeracy across age groups.

Recent commentators on these trends, also citing the broader drop in recreational reading among children and adults, have gone as far as suggesting we are reverting to something like a pre-literate oral culture, one that, because of its preference for speech over writing, is both less able to sustain the generation and transmission of complex ideas, and also more likely to prioritize the social-signaling functions of language over those to do with the neutral transmission of knowledge. While this concern may be exaggerated somewhat, is does seem directionally correct. We don’t need to experience a wholesale regression to oral culture for it to be true that our capacity to internalize and transmit complex knowledge is getting worse. Anecdotally, many people seem to admit as much about themselves, and the data seems to bear it out too.

Over the last fifteen years, companies like Apple, Google, and Meta have succeeded at migrating a massive and ever-growing share of human social behavior into publicly unaccountable systems they control. This has quite plausibly come at the cost of things we place great value on, and yet it has happened nonetheless. While a transition of this kind would once have required a period of great social upheaval, this one’s technical basis in consumer-facing media and electronics made it much faster and more cost-effective to carry out.

The strange and unstable world we have found ourselves in was built, in a darkly comic twist, on gizmos and gadgets, on cat videos and rage comics, on that app that used the iPhone accelerometer to make it look like you were drinking beer, on TikTok dances and Bitmojis and Overwatch porn. It was built on a fact the tech elite have exploited to world-historical effect, that we are animals easily diverted from our own flourishing, that like many of our smaller-brained counterparts, we will give up everything for depthless bursts of instant gratification.

4.2. Entering the interior

As a consequence of the predatory mechanisms hard at work on the Forum’s exterior, we have for the most part been successfully corralled into the enormous markets for communication represented by the major social networking platforms. It is now a matter of daily routine for us to visit these platforms in order to exchange our attention for content, or our content for attention.

Attention, recall, is like money in this context: it is a medium of exchange, a store of value, and a unit of account. You can accumulate it, measure it, and spend it like hard currency (e.g. by directing your followers to an advertised product, or another piece of content). But attention is obviously very different from other forms of money. It is neither an object we manufacture, nor a number on a digitized ledger. It is something we disburse using our bodies. As a result, it is a form of money more intimately connected to human biology than any other. We have evolved to distribute our attention according to certain patterns, and if you understand those patterns then you can exploit them for your own gain. From this principle flows much of the what goes on in the Forum’s interior.

Successful content creators are well aware of our attentional biases and have taken advantage of them by relying, in various guises, on the same strategy: exploiting the extremes of human emotion. Inciting heightened excitement, surprise, jealousy, anger, and so on—these are the quickest paths to super-virality. This dynamic is apparent in the exaggerated facial expressions now typical of video thumbnails, and the associated phenomenon of “YouTube voice,” the over-the-top, car-salesman-like tone now widely used in video content. It is also apparent in more disturbing places, such as in the viral success of misogynists like Andrew Tate and Adin Ross, the “skinny influencers” promoting disordered eating on TikTok, and the “pedophile hunters” movement, whose members stream themselves catfishing, and then beating up men looking for underage sex.

Those responsible for these fixtures of our new discourse economy recognize, at least implicitly, that we are animals wired to prioritize certain stimuli over others. The reason our emotional extremes are “extreme” to begin with is because they evolved to feel that way in order to track experiences more likely to affect our (evolutionary) fitness. That is also why the negative extremes—fear, anger, hate, and so on—tend to be even more salient than the positive ones; they were once correlated with genuine physical threats. This is the reason why human psychology exhibits a holistic negativity bias, as well more why we have more specific attentional biases toward negative stimuli. Like other organisms, we are uniquely attuned to anything suggestive of danger.

4.2.1. On politics

Note: a greatly expanded version of this section is now available here.

It is unsurprising, then, given our biological endowment, that much viral content—especially of a political nature—appeals to the darker side of our emotional range. Recent scholarship has shown that posts which are sad, fearful, uncivil, morally and emotionally indignant, hostile towards out-groups, and negatively valenced in general spread much further than their more neutral counterparts.

Brady et. al. (2017), for instance, found that each additional word with “moral-emotional content” (e.g. “duty,” “fear,” “shame,” “war”) increased political Twitter posts’ diffusion factor by 20%. Brady et. al. (2019) later reproduced this finding using the tweets and retweets of over 500 presidential candidates and members of congress. Expressions of “moral anger and disgust,” the authors found, diffused particularly quickly.

What the design of the Forum’s interior does, in effect, is to scramble our normal incentive environment surrounding expressions of extreme emotion. For most of our existence, such expressions have been reserved for rare, risk-laden occasions. But online, the incentives run the other way: the more scared or angry you are in public, the more attention flows to you. Plus the personal risks you run are virtually nil; there is no one in your immediate vicinity who will punish you for crying wolf, or risk initiating a physical altercation.

As bad as these incentives are, though, it is important to guard against the impression that everyone responds to them in the same degree. As Robertson, del Rosario, and Van Bavel (2024) explain:

Online discussions are dominated by a surprisingly small, extremely vocal, and non-representative minority. Research on social media has found that, while only 3% of active accounts are toxic, they produce 33% of all content. Furthermore, 74% of all online conflicts are started in just 1% of communities, and 0.1% of users shared 80% of fake news. Not only does this extreme minority stir discontent, spread misinformation, and spark outrage online, they also bias the meta-perceptions of most users who passively “lurk” online. (pg. 1)

That last point is particularly key: the overrepresentation of extreme content in the Forum, even if it is only originated by a small group, distorts everyone’s perceptions of public opinion. Social media creates, as political scientist Henry Farrell has put it, “publics with malformed collective understandings”—in this case, that perceive themselves as much more ideologically extreme overall than they in fact are on an individual level.

That said, how malformed that understanding is remains an open question. There is some evidence to suggest that what began as a malformed understanding may have become more accurate over time. Those perceiving disproportionate extremism also tend to become more extreme themselves. Kim et. al. (2021) found, for instance, that those exposed to toxic ‘featured’ comments on Facebook posts were more likely to post toxic comments later of their own volition. Brady et. al. (2021) found that Twitter users “conform their outrage expressions to the expressive norms of their social networks,” (p. 1) and become more likely to express outrage when their surrounding social network is itself more extreme (p. 8). And in an as-yet unpublished preprint, Brady and co-authors found that in a simulated Twitter-like environment, “engagement-based algorithms,” promoted morally and emotionally charged political content, changed participants’ perception of descriptive and prescriptive norms surrounding such content, and ultimately made them more intent on posting content of this kind themselves.

As concerning as this research is, it is important not to overstate the reach of any given ‘pathway to radicalization.’ We should note that, first, the baseline number of individuals engaging in more extreme or emotionally charged political activity online is still very low by proportion; second, that only about half of American social media users are politically active online in the first place; and third, that political content is much more typical of some platforms (like Twitter) than others (like TikTok). These considerations are suggestive of a more limited, but perhaps no less dangerous conclusion: that just the most politically active subset of social media users is becoming more extreme, whether in its specific views, or degree of affective investment in politics generally.

While this group only amounts to a small slice of the voting population, it drives many highly significant political outcomes. The most politically active social media users skew much wealthier and more educated than average, meaning that they are also likely to be more politically influential offline. Wealthy, educated elites have much larger ‘megaphones’ with which to influence the wider political culture, whether because they are more connected to political insiders, have more money for lobbying and donations, or appear more often in both establishment and self-styled ‘anti-establishment’ media.

This would appear to land us in the same place, albeit by a different causal channel. Whether social media is making everyone’s politics more extreme, absent substantial mediation by elite influencers, or making elites in particular more extreme, who in turn disseminate their views to the wider public, it is almost certainly a significant driver of what has become an untenably angry, anxious, and chaotic political environment over the last decade and a half.

While its propensity to fuel political outrage is hardly a unique property of social media, actors in other mediums, like the hosts of inflammatory T.V. and talk radio shows, have never been as diverse, unregulated, and numerous as social media users. In the same way more competitive markets for goods squeeze more desirable commodities into existence, so more competitive markets in communication do the same for content. Relative to other sources of media, the Forum is a uniquely powerful incubator of outrage-inducing engagement-bait.

Online harassment campaigns, hate speech, ‘cancel culture,’ ‘doomerism’, and widespread political acrimony are in hindsight the obvious result of restructuring social life as a market where acts of communication are valued by the amount of attention they elicit—few things gets more attention than the ringing of alarm bells. That the attentional demand for the outrageous and hostile exceeds that for the true and the just means the former is what will get made en masse. The latter, in turn, are set to become the informational equivalent of luxury goods.

4.2.2. On mental health

Note: an expanded / revised version of this section is now available here.

It is not only the explicitly political, though, which is to blame for the Forum’s toxic atmosphere. Its competitive market structure makes the Forum a powerful generator of social conflict and status anxiety more generally. It has never been easier to compare oneself to others, to build resentment over those comparisons, and to lash out in response, particularly when lashing out is rewarded by engagement from one’s peer group. It has also never been easier, especially for younger users, to be made to feel lesser-than by the glossy irrealism of others’ lives online. But rather than lashing out, an unprecedented amount of adolescents, the population the spends the most time ‘inside’ the Forum, have fallen into mental health crisis.

Despite remaining more or less constant through the aughts, mental illness among young people in the U.S. began to sharply increase starting around 2010. To give a small, but grim sampling from Jonathan Haidt’s popular book The Anxious Generation, from 2010 through 2020, depression increased by 106% among college students, anxiety by 139% and 103%, respectively, among those ages 18-25 and 26-34, and suicide rates by 167% and 91%, respectively, among girls and boys aged 10-14 (p. 26-31).

There is of course a lively debate as to whether these developments can be attributed to social media. Critics often cite large meta-analyses which they claim show the existing evidence to be equivocal or inconclusive. For instance, Candice Odgers, a prominent critic of the alleged link between social media and youth mental health claims, in her critical review of Haidt’s book, that efforts by researchers to search for the “large effects suggested by Haidt … have produced a mix of no, small and mixed associations” (p. 29).

Odgers’s phrasing would seem to suggest a blend of statistically significant and insignificant effect sizes with both positive and negative signs. But that is not a remotely accurate description of the literature. Of the 4 meta-analyses and review articles cited by Odgers in her review of Haidt, every single one, but her own openly acknowledges that the predominant result in the correlational research they review, with respect to adolescent anxiety and/or depression in particular, is a small, but statistically significant association of about r = .1, in other words a persistent, although modest association between social media use and mental illness. Several similarly comprehensive meta-analyses Odgers does not cite all find the exact same thing. The most recent of these—Liu et. al. (2022)—reviewed 21 cross-sectional and five longitudinal studies covering 55,340 participants and found that “the risk of depression increased by 13% for each hour increase in social media use in adolescents” (p. 1). The notion that the correlational research is equivocal or inconclusive on this point is flat-out false.[6]

This of course does not show much in and of itself. It may be that those correlations reflect causality running in the other direction—that more depressed or anxious kids use more social media, rather than the other way around. But it does mean that the hypothesis of a causal link passes initial muster. Social media might well have been associated with reduced anxiety and depression. It might well have only ever had a statistically insignificant association. But instead, most of the correlational research points the same way at a first pass.

Fortunately, we don’t only have the correlational research to rely on. A growing number of experimental and quasi-experimental studies support the existence of a genuine causal link. Some are social media ‘reduction’ studies, in which regular social media users take a break from social media for a period of time, and some are social media ‘exposure’ studies, which assess the impact of gaining access to social media. Both point in favor of genuine causation.

In a recent meta-analysis of 20 social media reduction RCTs covering 10,106 (adult) participants, Liu et. al. (2025) found that the interventions had a common effect size of .233 on overall well-being, a small, but significant positive effect (p. 10). The largest reduction experiment to date, Alcott et. al. (2020), assessed the effects on 2,743 people of deactivating Facebook for the four weeks before the 2018 U.S. midterm election, and found they experienced .09 SD and .1 SD decreases in measures of anxiety and depression, respectively. Another large study, Tromholt (2016), with 1,095 participants, found that quitting Facebook for only one week significantly improved life satisfaction and emotional positivity, and that both effects were larger for both heavy and passive users. Brailovskaia et. al. (2020), Lambert et. al. (2022), Reed, Fowkes, and Khela (2023), and Davis and Goldfield (2025), though smaller in size, all found similar, significant effects in their own reduction experiments. Findings of decreased symptoms of depression were particularly common. Though a number of other small experiments show conflicting results, the balance of evidence, especially when adjusted for statistical power, favors a small, but positive causal effect of reducing social media use on mental health and general well-being.

Reduction studies, though, can at best tell us about the likely effects of reducing social media use in a world where it is already very common. Exposure studies, on the other hand, can help us understand the impact of its initial spread. As one would expect given the evidence reviewed so far, exposure studies provide further confirmation of a causal link between social media use and mental health. Two in particular are worth highlighting.

Braghieri, Levy, and Makarin (2022) took advantage of a natural experiment—the staggered early rollout of Facebook on college campuses between 2004 and 2006—to assess the differential impact of the platform on mental health among college students through 2008. They found that the introduction of Facebook led to a .085 SD increase in an index of poor mental health among college students, a magnitude equivalent to around 22 percent of the effect of losing one’s job, and a 2 percentage point increase in the share of students suffering from depression. They also calculated (on the basis of some relatively strong assumptions[7]), that the introduction of Facebook accounts for approximately 24 percent of the increased prevalence of severe depression among college students over the last two decades (p. 3662).

Guo (2022) also took advantage of a natural experiment—spatial variability in the introduction of high-speed wireless internet to British Columbia—to assess the impact of greater social media access on teenage students. She found both that access to high-speed wireless internet was a good proxy for greater social media use (p. 22-23), and that, while boys did not experience a significant change in mental health in areas with access to high-speed wireless (and if anything saw slight improvement), teen girls in those areas saw a 90% increase in severe mental health diagnoses from 2010 through 2019 (p. 28). This is consistent with previous research finding associations between digital media use and mental health to be larger in girls than in boys.

Guo explains her findings by pointing to research that suggests girls are more sensitive to how social media might reshape their social landscape and interpersonal lives. She also cites internal research done by Meta which found, among other things, that Instagram made “’body image issues worse for one in three teen girls’” and that “’constant comparison on Instagram is the reason why there are higher levels of anxiety and depression’” (p. 9).

Meta’s in-house research team isn’t alone in this conclusion. A number of recent papers have honed in on interpersonal comparison as a key mechanism explaining the link between social media and mental illness among young people, in particular girls. Adolescents are (in)famously attuned to social hierarchies, and whereas they once would’ve only had to worry about local ones, those at their school or in their community, the Forum thrusts them at an early age into a substantially larger, more competitive social pool, one in which the “popular kids” are not just the richest or most attractive in their school, but the richest and most attractive in the world.

Not only that, but engagement metrics in effect serve as ‘popularity scores’ that allow them to compare where they or their peers fall in the pecking order. It makes a lot of sense that 57% of Gen Zers would “become an influencer if given the opportunity,” considering that, for them, this is more or less equivalent to becoming ‘the most popular type of person’ in their now oversized reference group. What is often missed in overspecific debates about the hyper-competitive nature of college admissions is that all of childhood is now much more competitive, including and especially the childish parts. Call it ‘free trade’ in the teenage popularity market. Once localized status competition—subject to various forms of ‘infant industry protection’—has now gone global. Unionized American machinists versus sweatshop workers in Bangladesh? Try being a teenage girl competing for likes with Charli D’Amelio.

Jokes aside, it is straightforward to see how, for the Forum’s heaviest and most sensitive users, constant contact with curated displays of high status might lead to distress. Exposure experiments like Guo’s further reinforce the conclusion implied by every other line of evidence: social media causes small, but statistically significant increases in mental illness.

Those who continue to stubbornly oppose this conclusion are going to find themselves with fewer and fewer intellectual allies in the years to come. Strong evidence continues to pile up, and that is largely before the effects of the much more recent advent of short-from video content have made it into datasets. Though a few academic holdouts will likely continue to harp on the issue of small effect sizes, they would do well to consider the following:

a) much of the literature is cross-sectional or dependent on short interventions, meaning it won’t capture the compounding effects of social media use over time (remember: teens spend almost 5 hours on social media per day—and have been doing so for over a decade);

b) effect sizes are averages, and so consistent with vulnerable sub-populations who fare much worse than the norm, and in turn impose large second-order costs on others (financial, emotional, etc.);

c) the baseline number of adolescents with any one mental illness is low, and so not especially difficult to increase in large percentage terms with an intervention (i.e. social media) applied to 100% of adolescents; and

d) many adolescents may be on the threshold of mental illness, and only need a small nudge to be pushed over.

For these reasons and others, small effects on the individual level can mean very large ones on the national level.[8] While numerous questions of course remain as to the exact nature of social media’s effects on mental health, the evidence of a genuine causal link is strong. If nothing else, the rise of the Forum has at minimum done serious, and sometimes irreversible harm to many of America’s young people.

The interior of the Forum is a vast and hyper-competitive market for nearly every kind of communicative act, whether textual, visual, or auditory. Considering its sheer breadth, it is surprising its impacts haven’t been even larger. What seems clear, at minimum, is that the Forum is an environment most humans are not well outfitted for. We did not evolve to interact with hundreds of thousands of strangers for multiple hours a day, particularly not ones stoking fear, anger, and self-loathing of almost every conceivable variety.

That does not mean we won’t adapt to this environment; history tells us that we probably will. But as it stands—if the trajectory of the last fifteen years is any guide—we have many reasons to fear, and few reasons to remain hopeful, about the sorts of people that will come out the other side.

5. Conclusion – on completing the trinity

There are few processes more emblematic of modernity than the routing of immense tracts of uncodified social behavior through ever more structured, legible, and narrowly optimized filters. The rise of the Forum is in this sense an incontrovertible sign that we never left the long modern period, that history in its most upheaving, spiralic sense is still chugging along.

For James Scott—who died in July of last year at 87—modernity meant, among other things, the loss of vast stores of practical knowledge, or ‘mētis.’ Mētis is that subtle, uncodifiable form of know-how often necessary for performing complex tasks—what is sometimes called ‘tacit knowledge.’ That skills like playing tennis or riding a bike can’t be learned except by repeated practice is because they require mētis, knowledge manifested so subtly in our brains and bodies that its content is indescribable. As the social world undergoes change, Scott argued, many once-important forms of practical knowledge dissipate and are eventually lost.

Though physical, rather than mental or emotional skill are often cited as exemplifying mētis, the truth is that mētis is essential to every human endeavor, both individual and collective. To no less an extent than riding a bike, navigating social and political life requires massive stores of tacit knowledge, knowledge that can’t be accumulated except by participation. There is no instruction manual for how to make friends, or fall in love, or reach consensus. To learn these skills requires practice, and to sustain them requires norms—norms that impart pro-social aptitudes directly, as well as which sustain an environment where people learn them and are rewarded for it.

The world of the Forum, as Scott might have anticipated, is ill-suited to developing the mētis for human social cooperation. Our cooperative tendencies grew from analog soil, grew from contexts in which, without cooperation, we would die, or at least suffer physical harm. The Forum lacks not only these stakes, but the informational richness of face-to-face interaction, the length and depth of conversation, of body language, of place. It also lacks norms that last. To make and enforce a norm is hard. It is costly. That remains so on the Forum. But to break a norm, to spurn it, to ignore it, is easy—because on the Forum, no one can punish you for doing so in any way that matters.[9] More importantly, it is breaking norms, not making them, that gets the most attention.

Recalling the parable with which we began, what is being “dismembered” by the Forum—on a scale without even remote precedent—is our ‘pro-social mētis,’ our basic aptitudes for cooperation, the wealth of intuitive knowledge we have long used to live well together. Across social life, education, and politics, the Forum is either corroding, or displacing altogether the messy, opaque, frictional—but for those very reasons, enlightening—experiences by which we learn to communicate in a manner conductive to the public good. Those experiences cannot and will not survive their being narrowly evaluated on their propensity to win attention to the exclusion of every other goal or value.

We have, for years, been misled by the myopic conception of social media and the smartphone as just another pair of communications technologies, ones easily grouped in with predecessors like the telephone and personal computer. While there is nothing obviously wrong with this perspective, it elides more than it reveals. Much of what has been said about social media and the smartphone suffers from a narrowness of vision, a failure to grok the larger historical forces lurking in what risks seeming like just another round of American consumer indulgence.

The rise of social media and the smartphone has initiated a transition between communicative paradigms, one happening so quickly we’ve barely begun to reckon with its consequences. The social world as we once knew it was anchored in a multiplicity of communicative practices distributed across a wide array of analog cultural and geographic contexts. The social world being born is anchored in the practice of attention maximization, and spans a set of digital markets jointly covering almost 65% of the global population.

This is the world of the Forum, a world that needs to be named, lest we one day forget the world that preceded it.

You may still be wondering about the name ‘the Forum.’ You may be wondering why the discursive economy made possible by social media and the smartphone deserves a name this general, a name that seems at once to risk emptiness and excess. But there is, in fact, a reason for this specific term, with its lofty historical and juridical connotations. The name of the Forum has to speak to its sweeping nature, has to speak to the fact it is not just a set of technologies, but an institution, one no domain of human communication can remain insulated from for long.

But to refer to the Forum as an institution is still to the understate the point. The Forum isn’t just an institution. The Forum is the institution through which public discourse now flows. It is the institution making claim to this this foundational part of our species’ affairs. The Forum is doing to civil society what the state once did to political authority. It is doing to the public square what the market once did to production and exchange. It is modernizing social life.

We are living through the birth of what may well be remembered as a foundational institution of the modern world—and have so far failed to realize it. The great duality of state and market, of public and private, of the political and the economic, are on the verge of becoming a trinity, as modernity extends its reach into uncharted, tempestuous waters.

State, market, and forum.

[1] A similar economic logic explains the appeal to networks of reality and game shows, which offload the costs usually associated with writing and acting to members of the public.

[2] The possibility of creating a massive producer surplus in one currency by substituting the portion of it paid to labor with a previously nonexistent one is a fascinating and underexplored economic phenomenon. This is similar to the dynamic whereby a surplus is created by taking advantage of unpriced natural goods, like the sun and wind in the case of green energy, or, to pick a case more similar to the one at hand, the free labor of slaves. The monetization of our natural capacity to disburse attention is a bit similar to a scenario wherein slaves are made to compete to see who is physically stronger or faster—in return, perhaps, for ‘strength’ or ‘speed’ tokens—by doing some economically valuable work, say, harvesting tobacco crops. Under certain conditions, maybe winning the most such tokens would be seen as reward enough for doing taxing and otherwise uncompensated labor. At the moment, that others might pay them (more) attention seems to be enough for many content creators on social media.

[3] For a very insightful paper on this phenomenon in the context of in online political media, see Williams (2022).

[4] What, then, to make, conceptually speaking, of advertisers purchasing users’ attention from platform companies? Perhaps this is more akin to exchanging currencies than it is to buying goods or services. If so, advertisers are engaged in a kind of temporal currency arbitrage: they trade dollars to platforms for attention in the short term, knowing that this attention is worth more dollars to them in the long term because it leads consumers to buy their products.

[5] Unless otherwise specified or obvious from context, by significant, I always mean statistically significant, not intuitively large or important.

[6] Read Haidt’s own rebuttal to Odgers here.

[7] These assumptions, however, seem more likely to understate than overstate the authors’ case. They are that “(i) Facebook utilization rates among college students did not change substantially after 2004–2005; (ii) the effects of Facebook did not change over time; [and that] (iii) Facebook does not have cumulative effects.” In footnote 23, the authors point out that (iii) is probably false, given that “the negative effects of Facebook on mental health become stronger with longer exposure to the platform” (p. 3675). There are various reasons to think both (i) and (ii) are false as well in ways that would increase the share of depression explainable by Facebook, e.g. that utilization rates have in fact greatly increased since 2005, and that the site has added plenty of new features since then.

[8] For a more detailed defense of the “indispensability of small effects,” see Götz, Gosling, and Rentfrow (2021).

[9] ‘Cancel culture’ may seem like a counterexample, but note that so-called ‘cancellation’ itself involved norm-breaking in the form of imposing harsh informal sanctions most Americans saw as too strong. Also consider that, such as it existed, ‘cancel culture’ was widely maligned precisely because it attempted to impose pro-social norms. Moreover, even if ‘cancel culture’ could occasionally impose large costs for breaking certain norms, many of those norms (depending on the case) were not widely held, and for that very reason, ‘cancellation’ often failed to make or sustain norms among all but a very small population. At most, the Forum can generate, though not sustain norms of a very restricted, dysfunctional kind.