Everyone who has worked in a lab has had the experience of buying a reagent that looks perfect in the manufacturer’s catalog only to have it fail repeatedly in your hands. The frustration of producing bad Western blots is a right of passage in research. Every basic science lab has at least one box full of “bad antibodies” that never performed as promised. Most researchers just assume that they are doing something wrong and move on to another antibody.

But what if the reason for these failures isn’t the researcher but the catalog itself? In May 2026, research integrity sleuth Reese Richardson found that hundreds of antibody validation images by Thermo-Fisher were modified.

A new analysis published on August 26 found that at least 15 more antibody vendors - including Abcam, Thermo Fisher, MilliporeSigma, Novus Biologicals, GeneTex, and Santa Cruz Biotechnology - have been selling antibodies using doctored or fabricated validation images. The count now stands at roughly 18,000–19,000 manipulated images across ~17,500 products, up from the 450 images from Thermo Fisher’s catalog that first surfaced in May 2026. For those who haven’t toiled doing biology research, monoclonal antibodies are an essential reagent, used in experiments like Western Blots to measure the level of expression of a target protein. Researchers rely on these antibodies to do what they claim, and the discovery that tens of thousands of the validations images are doctored is a devastating setback. One researcher called it “without a doubt the worst case of scientific fraud to hit the antibody world.”

The scope of this fraud is staggering, but what about the costs?

Yes it sucks to spend $200 and waste a week trying to do an experiment with a bad antibody. That undoubtedly adds to the cost of doing basic research. But the total costs of this fraud are MUCH larger.

Most people recognize the enormous time cost of this fraud - every basic scientist has wasted days trying to make bad antibodies work. But what about the opportunity costs? How many graduate students or post docs wasted crucial years trying to replicate an image a manufacturer faked? What important discoveries were delayed by bad reagents? What life saving discoveries were never made because researchers were chasing irreproducible blots?

Let’s step back and ask a different questions: Why do antibody manufacturers fake an image in the first place? There are really only two motives: to make a specific band appear, or to make a non-specific band disappear. Richardson’s catalog shows evidence of both. Both are harmful but in different ways. Thousands of images share the same recurring “background patterns” identical noise with the bands simply repositioned to land at whatever molecular weight the listing needed.

And alongside that, hundreds of images show apparent painting-out or stitching-together of background signal, erasing the non-specific bands that a researcher could waste months, or years, chasing. The harms from nonspecific bands are easy to underestimate. A band a few kilodaltons too heavy or too light can easily fool someone into thinking they’ve found their protein when they haven’t. This likely underlies may irreproducible antibody-based findings: protein misidentification, false co-localization, spurious knockdown “validation.” How many papers and studies were ruined because a researcher chased a nonspecific band that a faked image told them was clean? Nobody has yet quantified this, but the general problem is not hypothetical: Yale pathologist David Rimm has said that in early checks, three of five antibodies his lab tested failed to give reproducible results at all. And in one documented case, an ELISA test that turned out to detect the wrong protein entirely cost the researchers who relied on it two years of work and roughly half a million dollars.

It’s harder to measure the clinical cost of false positives. How many drug candidates made it to clinical testing because of bad antibodies?

The major bottleneck is phase 2, where only about 31% of drugs succeed, and the dominant reason for failure there is lack of clinical efficacy: the drug simply doesn’t do what the preclinical target-validation work said it would. That target-validation work runs on antibodies. How many phase 2 efficacy failures trace back to a target that only looked good in a Western blot because the validation data was fake?

To state the obvious: Clinical trials are enormously expensive. In 2015 Economists estimated that roughly $28 billion a year is spent in the US alone on preclinical research that turns out to be irreproducible, and identified faulty biological reagents - antibodies chief among them - as the single largest contributor.

What is the human cost of these fraudulent western blots? How many patients were harmed by ineffective therapies developed using bad antibodies? How many would have benefited had they been enrolled in the minority of trials which succeed?

We’re all paying for this fraud whether you’ve ever pipetted or not, and the companies that are deliberately selling bad antibodies need to be held accountable.

The companies argue that a manipulated image doesn’t prove the underlying antibody is defective. But the product is not only the antibody; it includes a the clinical validation. What’s being sold isn’t only a vial of antibody; it’s the vial plus a claim: “here is proof this binds what we say it binds.” In a legal sense, every researcher who purchased an antibody based on a company-supplied Western blot has acted in reliance of data that turns out to have been fabricated. This means more than just reimbursing researchers - or the funding agencies - for the many boxes of bad antibodies. There ought to be a class action lawsuit to recover the total costs of this fabrication. The costs are more than just the dollars spent purchasing bad reagents, but the time wasted trying to validate them, and the patient lives lost in trials of ineffective therapies.

On August 19, 2026, Merck and Moderna announced via press release that the Phase 3 trial (INTerpath-001) of their personalized mRNA therapy intismeran autogene, combined with pembrolizumab, met its primary endpoint of recurrence-free survival and a secondary endpoint of distant metastasis-free survival in patients with stage IIB–IV melanoma. This is similar to the work they published a few months ago: the Phase 2b (KEYNOTE-942) results, which showed a 44–49% relative reduction in risk of recurrence or death versus Keytruda alone. The full results are not yet peer reviewed/published, but this is a exciting result that represents a big advance in cancer treatment.

Within hours, the framing on social media became “AI just cured cancer!” - implicitly or explicitly linking this result to large language models, GPUs, and gigawatt datacenters. This is far from the truth. While INTerpath-001 is likely a big step forward but it’s neither “a cure” nor is it “AI” in the sense most of the tech bros hyping it mean. We ought to remain skeptical, especially until we see the actual peer reviewed results (not just a press release). We also need to clarify what AI did that enabled this breakthrough and what it didn’t do!

What really happened?

First, the timeline. The neoantigen-selection pipeline underlying this vaccine was locked in before the original transformer paper was even published and was already generating clinical data years before ChatGPT existed (November 2022). Whatever “AI” did the work here, it isn’t the “AI” anyone was tweeting about implied.

Let’s look at the timeline more closely:

• June 2016 - Merck and Moderna sign their original collaboration and license agreement for mRNA-based personalized cancer vaccines.

• June 2017 - The paper describing transformer models - Attention is All You Need - is published

• November 15, 2017 - First patient dosed in the Phase 1 trial (KEYNOTE-603, NCT03313778).

• November 30, 2022 - ChatGPT Launches

• December 2022 / 2023 — Phase 2b (KEYNOTE-942) positive data announced; published in The Lancet, January 2024.

• June 2026 — Updated Phase 2 data at ASCO.

• August 19, 2026 — Phase 3 (INTerpath-001) readout.

So what did enable this breakthrough?

So if it wasn’t Large Language Models that enabled this breakthrough, what did?There are several key breakthroughs in computer science that enable personalized mRNA vaccines.

1. Deep sequencing allows you to find cancer specific mutations.

2. Neo-antigen/MHC-binding prediction algorithms identify epitopes that can be recognized by the host immune system.

3. Proprietary algorithms (like Moderna’s mRNA Design studio) allow you to design RNA sequences that will be stable and immuno-stimulatory.

All of this technology is exciting, but it’s not new. When I took a graduate course on vaccine development at Brown in Spring 2005, all of these technologies existed in various forms. They have all been refined since, but they clearly predate the 100 billion parameter contemporary LLMs that tech bros have implied are responsible.

What’s the secret sauce in these algorithms?

Neoantigen/MHC-binding prediction tools have historically ranged from simple position-weight matrices, to random forests (e.g., ForestMHC), to shallow/LSTM neural networks (e.g., MHCnuggets, DeepHLApan) trained on curated, structured features (peptide sequence, HLA allele, binding affinity, expression level) from relatively small datasets of thousands to low-millions of examples.

Moderna’s own patents don’t disclose which of these their algorithm actually is, but their sequencing partner’s does. Merck and Moderna have both used Personalis’s ImmunoID NeXT platform since mRNA-4157’s clinical development began, a relationship extended repeatedly, most recently with a $50M Merck investment in Personalis in December 2024. Personalis has published the on their MHC-peptide presentation ranking algorithm: SHERPA (Systematic HLA Epitope Ranking Pan Algorithm). The methods section states: “All models were trained using a gradient boosted decision tree algorithm implemented using an open-source package XGBoost.” SHERPA is trained on 2.15 million peptides across 167 HLA alleles, using laboriously hand-engineered features like peptide sequence, HLA binding-pocket sequence, transcript expression, flanking amino acids. We don’t know if Moderna’s approach is identical to SHERPA, but they clearly view this approach as valuable since they continue to collaborate. The key point is that this is a narrow, feature-engineered classifier built on a decade-old approach, not a contemporary foundation model.

I’m reminded of the old joke that machine learning is code written in python and AI is hype written in powerpoint. It’s not wrong to call these models machine learning - or AI if you must - but it’s incorrect to conflate these models with contemporary LLMs.

I’ve done a lot of work with machine learning algorithms, including XGBoost. They are incredibly useful for bioinformatics, but they don’t remotely resemble a transformer-based foundation model trained on internet-scale unstructured data. These are both technically AI in the way a bicycle and a Boeing 787 are both technically vehicles, and like bicycles and aircraft, they are suited to very different things. Just as the wide-body jet would be ill-suited for a trip to the movies, the 100 billion parameter LLM probably isn’t ideal for designing personalized cancer vaccines.

If not AI, who does deserve the credit?

This is not to diminish the accomplishment of developing a personalized cancer vaccine, a technology that will hopefully be lifesaving for millions of people. But we should remember where the credit belongs:

• Three decades of work on tumor immunology and neoantigen biology

• Development of an immunopeptidomics data that matured through the 2010s; this is arguably the real breakthrough, more so than any specific algorithm

• Technology for mRNA delivery and rapid personalized manufacturing

• Development of immune checkpoint inhibitors that enable this approach in the first place!

• Nine years of unglamorous clinical trial execution that began before transformers were even described!

The problem with the “AI cured cancer” framing is that it doesn’t just overstate what a gradient-boosted tree can do, it points the credit at the wrong decade and the wrong discipline. The people who deserve the headline are immunologists who spent thirty years defining cancer neo-antigens, the engineers who figured out how manufacture mRNA at scale, and the physicians who ran nine years of adjuvant melanoma studies. Misinterpreting this breakthrough as a win for foundation-models leads to two big misperceptions: people expect Claude and ChatGPT to spit out cancer cures on demand, and the field that actually did this work doesn’t get the credit - or the funding - it deserves.

“Kids today get 72 jabs” is an oft-repeated rallying cry among those virulently opposed to vaccines. Like most viral numbers, “72 jabs” has become a truism accepted unquestioningly by people who’ve never bothered to count. And much like the “72 virgins” supposedly promised to suicide bombers - a claim that turns out to rest on a mistranslation, not scripture - the myth of “the 72 jabs” is built on nothing but a fundamental misunderstanding of what words mean and how to count.

So let’s count.

How many vaccines do kids actually get?

The CDC recommends children be vaccinated against 15 diseases:

1. Hepatitis B (HepB)

2. Rotavirus

3. Diphtheria

4. Tetanus

5. Pertussis (whooping cough)

6. Haemophilus influenzae type b (Hib)

7. Polio (inactivated polio virus; IPV)

8. Pneumococcal disease (PCV13/15/20)

9. Measles

10. Mumps

11. Rubella

12. Varicella (chickenpox)

13. Hepatitis A

14. Human papillomavirus (HPV)

15. Meningococcal disease (MenACWY)

Plus three more recommended under shared clinical decision-making:

16. Influenza

17. COVID-19

18. Meningococcal B

Eighteen. You’d need all your fingers and most of your toes to count that high, assuming standard-issue anatomy. Eighteen is a big number. But 72 is a much bigger number. So “72 jabs” can’t possibly refer to the number of diseases kids are protected against by vaccines.

Maybe a “jab” is the injection itself?

Maybe the “jab” isn’t the vaccine but the big scary needle used to deliver it?

Thanks to combination vaccines, kids routinely get protection against several diseases in a single shot. MMR (measles-mumps-rubella) and DTaP (diphtheria-tetanus-acellular pertussis) each fold three vaccines into one injection, cutting down on both cost and pain. Newer combinations manage to combine four (MMRV), five (Pentacel: DTaP + IPV + Hib), or even six (DTaP + IPV + Hib + HepB) vaccines into a single injection.

And some vaccines, like rotavirus and the live oral polio vaccine used in some countries, aren’t injections at all; they’re an oral liquid.

If we add it all up, by injection, across the full childhood schedule we get:

* Depending on the time of year, newborns may also receive a monoclonal antibody that protects against respiratory syncytial virus (RSV). This is an antibody, not a vaccine. We’ll count it anyway.

** At birth, newborns receive an injection of vitamin K to protect against vitamin K deficiency bleeding (VKDB), a devastating condition historically known as hemorrhagic disease of the newborn (HDN). This is a vitamin, not a vaccine. We’ll count it anyway.

Tally it up and kids receive roughly 28 to 32 injections between birth and age 18. Thirty-two is still a big number, but it isn’t remotely close to 72.

Dishonest accounting: doses as separate “jabs”

If your doctor prescribed an antibiotic twice a day for a week, would you say you’d taken one antibiotic, or fourteen? If you take just a baby aspirin daily, do you say you’re on 365 different medications a year?

It’s some disingenuous, Enron-level intellectual accounting to count multiple doses of the same vaccine as different “jabs.” Ask yourself why do anti-vaxxers reach for this kind of accounting only when it comes to vaccines.

But let’s indulge their innumerate fiction and count every individual dose, of every vaccine, separately:

Universal vaccines

1. HepB: 3 doses

2. Rotavirus: 2 doses

3. DTaP: 5 doses

4. Hib: 3 doses

5. PCV: 4 doses

6. IPV: 4 doses

7. MMR: 2 doses

8. Varicella: 2 doses

9. HepA: 2 doses

10. Tdap: 1 dose

11. HPV: 2 doses (typical)

12. MenACWY: 2 doses

Optional

13. MenB: 2 doses

14. Influenza: at most 17 doses

15. COVID-19: at most 17 doses*

* This assumes the COVID vaccine remains on the schedule for decades following the pandemic which is total speculation.

Even counting every dose separately, the total comes to 32. If we include the MenB and Influenza the total comes to roughly 51 doses across 18 years, still nowhere near 72. Even if we assume the COVID vaccine continues to be given annually forever - a total speculation - we still only get 69 doses! Closer but even with the most dishonest possible accounting we don’t get there!

Maybe a “jab” means the antigens inside the vaccine?

What if “jab” is trying, crudely, to describe the antigen load in vaccines? If so, the number has moved in exactly the opposite direction anti-vaxxers assume.

Older inactivated whole-cell and live-attenuated vaccines containing many antigens have been replaced by newer, purified vaccines with far fewer antigens. A kid in the 1980s received thousands of antigens; the whole-cell pertussis vaccine alone had over 3,000. Because modern vaccines are highly purified, a child today receives fewer than 150 antigens in total. Thus, my kids encounter about 95% fewer vaccine antigens today than I did as a child in 1986, despite being protected against many more diseases. Of course, this is a drop in the bucket compared to the billions of antigens that kids are exposed to in their food and environment.

Bottom line

However you define “jab,” the number 72 doesn’t show up:

• Vaccines: ~15–18

• Injections: ~28–32

• Individual doses: ~50–55

• Antigens: <150 - a tiny fraction of what kids received a generation ago

There’s nothing mystical about the number 72. Anti-vaxxers didn’t arrive at it by counting anything. They arrived at it the same way most viral misinformation spreads: by repeating a big, scary number loudly and often, hoping that nobody stops to check their math.

So next time you hear someone bemoaning the “72 jabs”, ask them to show their work.

Most of the public conversation about direct-to-consumer whole-body MRI is about the risk of false positives, leading to unnecessary followup, biopsies, cost, and anxiety. This is an important issue for doctors and patients. But the existential threat to the whole body MRI industry isn’t false positives, it’s false negatives.

The standard objection to whole-body MRI (WB-MRI) screening is that because these scans are extremely sensitive, they lead to false positives. A 2020 review of 6,214 whole body MRI scans found that 95% of scans showed at least one abnormality identified and 30% required further followup. This leads to specialist consultations, repeat scans, and in some cases invasive procedures like biopsies. But the existential risk to the WB-MRI industry lies in the false negatives. Unlike the negative externalities of false positives - anxiety, costs, and complications - which are borne by the medical system and by patients, false negatives are a retained liability for the WB-MRI companies.

Every missed finding on a whole body MRI is potentially millions of dollars of future liability. As the companies scan more people their revenue grows linearly, but their risk grows even faster with both every patient and every year of followup. Customers can “churn” deciding not to pay thousands of dollars a year to the company, but the liability for missed diagnosis persists:

You might wonder, why can’t whole body MRI companies limit their liability by erring more on the side of false positives instead of false negative?

They try to, but they can’t because of the fundamental physical and cognitive trade-offs required to offer whole body MRI scans. In order to keep scan times tolerable (under 60 minutes) they have to settle for thicker slices, fewer pulse sequences, and incomplete coverage. The companies must accept lower resolution and less informative sequences than an organ specific scan in order to keep the scan short.

The other issue is the lack of clinical context. Everyone who orders imaging knows the more information you give the radiologist the better the read will be. Ordering a MRI brain for “weakness” is going to be much less useful than specifying “sudden onset right upper extremity weakness.” The lack of context makes it hard for even the most thorough and careful radiologist to find everything on a scan.

Targeted scans also offer some legal protections. An MRI ordered for a specific complaint carries a narrower implicit claim: “we looked at what we were asked to look at.” A whole-body screening product sold as a comprehensive health check carries a much broader one: “we looked at everything, and you’re clear.”

This bring us back to the lawsuit (Clifford v. Prenuvo) and the defenses the company has tried, unsuccessfully, to use.

The radiologist who interpreted the scan was independent contractor of separate company. This is a common architecture in the direct to consumer imaging space: the scanning company and the reading radiologist are often separate legal entities. It’s plausibly there in part to manage liability exposure. But similar structures don’t protect hospitals from liability. It’s doubtful it will protect Prenuvo.

Prenuvo has also tried to compel arbitration, arguing that the matter should be handled outside a New York courtroom and has separately sought to limit damages. Both motions have been contested, and the damages motion has already failed when court ruled that New York law governs, since Clifford received care in New York, rejecting Prenuvo’s attempt to invoke California’s damages cap, leaving the company facing uncapped malpractice damages.

Normally radiology malpractice risk is priced, roughly, on a per-study basis: a given miss rate, a given claim rate, a given average payout, spread across a large volume of studies with a fairly short and predictable window between the read and the point where an error would surface. Whole-body screening MRI is different: the liability isn’t priced at the time of the scan. It matures years, sometimes decades, later, whenever the customer’s disease actually shows up. A person who has a “clean” scan at 35 has, in effect, purchased a permanent evidentiary artifact for future lawsuits. If they’re diagnosed with pancreatic cancer at 48, or have a stroke at 40, their lawyer doesn’t need to prove the disease was fully formed and obviously visible thirteen years earlier. They only need a radiologist willing to testify, in hindsight, that something on that scan - a subtle asymmetry, an early marrow signal, a vessel that looked “a little narrow” - should have prompted a follow-up recommendation. This presents a massive problem at scale for these companies. Revenue will scale roughly with customer acquisition and repeat-scan behavior, then flattens as churn sets in. Liability exposure increases steadily both with more customers and as each person who was ever scanned ages. Thus a business model that looked promising as a startup gets less and less attractive the larger and more established the company becomes.

Direct-to-consumer genetic testing went through a similar arc, strong early growth, thin regulatory scrutiny, and then a slow accumulation of legal and regulatory attention that undermined profitability. After its founding in 2006 and bouyed by attention from health influences, 23andMe grew rapidly in the 2010s, genotyping over 10 million people, and going public for billions in 2021. It ultimately declared bankruptcy in 2025.

This is not a prediction that the entire whole-body MRI industry will collapse. Some companies undoubtedly will, but others will adapt. Expect to see narrower marketing claims, tighter contracts, and higher prices. What looks today like a high-margin, high-growth health tech startup will, in ten years, look like a specialty insurer with an MRI machine bolted on: razor thin margins, a huge standing legal team, and frequent settlements.

It All Began With a Null Result

The first study of the “July Phenomenon” was undertaken by Buchwald et al in 1989, and looked at costs of care at a major teaching hospital early versus later in the academic year. It found no significant increase. Despite this negative result, the mechanism was too seductive to need evidence. Over the next fifteen years the “July phenomenon” seeped into the vocabulary of every teaching hospital in the country - passed along in orientation talks, nervous jokes on the wards, and the occasional cautionary op-ed - largely on the strength of how obviously true it felt. The name morphed into the more succinct “July Effect” but the sentiment remained unchanged: if you are a patient in July, be afraid.

Enter the Economists…

The folklore finally got some supporting evidence in 2005, when economists Huckman and Barro released a National Bureau of Economic Research (NBER) working paper titled Cohort Turnover and Productivity: The July Phenomenon in Teaching Hospitals. Using five years of admissions from a large multi-state hospital sample, they reported that teaching hospitals saw length of stay jump right after the July turnover and stay elevated for months, and that hospitals of medium teaching intensity showed a significant bump in mortality over the same window, on the order of a 4% rise in risk-adjusted mortality in July–August, translating to roughly 8 to 14 “accelerated” deaths per year at an average major teaching hospital.

If we examine the data more closely, several flaws are immediately obvious. Mortality and length of stay are lowest in July - exactly the opposite of what the “July Effect” would predict. The table that analyzes hospitals by degree of teaching is an exercise in p-hacking: only the middle third of major teaching hospitals saw a statistically significant increase. (This was just one of 20 groups analyzed for mortality, without correcting for multiple comparisons...) On the other hand, the hospitals with the highest teaching intensity, those most dependent on trainees, where the effect should have been worst, avoided the mortality bump altogether. The authors’ own read was that “higher supervision levels mitigated the disruption,” suggesting the effect was something a well-designed system could absorb.

Blaming the Doctors for the Weather?

In 2010 the “July effect” got even more evidence, leading to a surge of mentions in the popular press. Phillips and Barker examined every computerized U.S. death record over 28 years, focusing on fatal medication errors specifically, as an unambiguous marker of a serious mistake. They found that inside medical institutions, fatal medication errors spiked in July and in no other month, and this spike appeared only in counties containing teaching hospitals, where July medication-error mortality ran about 10% above expected. The more teaching hospitals a county had, the bigger the July spike.

Superficially, this is a very worrisome finding. But once again, when we examine the methodology more closely major flaws are apparent. Interestingly, they found no change in fatal medication errors after 2003 when the first work hour restrictions were enacted, something we might expect to see since that’s exactly why work hours were implemented! More importantly, because the computerized death certificates did not include the precise location of death, they could not measure deaths in teaching hospitals, only deaths in counties that contained a teaching hospital. Because the nations 1400 teaching hospitals cluster in urban areas, this introduces a striking bias in the data: urban vs rural. We also need to closely scrutinize their definition: “fatal medication errors” was defined by ICD9 codes E850-E858. This does not attribute who was responsible - and in most cases reflected patient overdoses that occurred outside of the hospital and subsequently died after admission. It is well known that exposure to heat is associated with an increase in the risk of drug overdoses, and that months with greater temperature have more overdose deaths. Does this association between urban deaths in July reflect the doctors or just the weather?

Torturing the Data Until It Confessed

A few years later the “July Effect” got another boost, and another round of headlines. Jena and colleagues (2013) looked at high-risk acute myocardial infarction patients from the national inpatient sample from 2002-2008 and ran a difference-in-differences comparison of May versus July across teaching-intensive and non-teaching hospitals. Among the highest-risk heart-attack patients, mortality in teaching-intensive hospitals was 18.8% in May but 22.7% in July, while non-teaching hospitals showed no such May-to-July gap.

Once again, the devil is in the methodological details.

Why did they compare July with May and not with June? Because the effect isn’t significant if you do! Why did they only look at “high risk” patients? Not only does the effect go away, but July is a month with the lowest mortality for low risk patients; substantially lower than at a non-teaching hospital! If there really is a “July effect” wouldn’t we expect August to be almost as dangerous? In fact August is one of the safest months with one of the three lowest adjusted mortality rates. Once again, the authors make no attempt to correct for the 24 comparisons performed in just this figure:

A decade later, in 2023, a different group revisited this question with 18 years of national data on heart-attack patients - a far longer and more contemporary window than the 2002–2008 slice Jena used. Using an much larger population (n=1.3 million admissions for acute myocardial infarction, compared to the 75,000 used by Jena et al) they found that July AMI admissions actually had lower mortality than those in May. The conclusion was blunt: there was no July effect for in-hospital mortality in this contemporary AMI population.

The same story has played out many times in the scientific literature. A small methodologically dubious initial study finds a weakly significant “July effect.” Despite the limitations the authors and their institution hype it up, and the media credulously publishes terrifying click-bait headlines. Meanwhile, larger and more rigorous studies fail to replicate the finding and get published in lower tier journals with much less fanfare. For example, several studies looking at condition specific outcomes, found null results where the effect should have been the loudest: no July effect in appendicitis, in trauma, in acute ischemic stroke, in acute cardiovascular management, or in mechanically ventilated ARDS patients.

Six Million Patients Later…

In 2023, Zogg and colleagues, put the myth to the test with a systematic review and random-effects meta-analysis of the entire literature, including every July-effect study published before December 2019. This totaled 113 studies of which 81% of them found no July effect. When they pooled the numbers, the “July Effect” disappeared:

• Mortality: odds ratio 1.01 (95% CI 0.98–1.05).

• Major morbidity: 1.01 (0.99–1.04).

• Readmission across 5.98 million patient encounters and 7 conditions: no effect.

They found no signal by country, specialty, or condition. They found no change over time. The authors’ recommendation was essentially to stop running these studies and redirect the energy toward the year-round quality and efficiency problems that probably generate the perception of a July effect in the first place.

Why a wrong idea lived so long

The July effect had everything a medical myth needs. A mechanism that seemed intuitive and a permanent supply of confirmation bias: every bad outcome in July must be due to the “July effect” whereas the same outcome in June could be attributed simply to bad luck. It was buoyed by researchers who cut corners methodologically to sustain the myth, and health journalists who amplified it for clicks.

There are real costs to July. Academic centers do see a rise in length of stay, procedure times. But this may reflect the opposite cause: new interns are slower because they are more careful. Just like the 2005 NBER theorized: the system was built to absorb the shock, and the aggregate data says it largely does. So the next time someone tells you not to get sick in July, you can tell them the truth. The first study on the July effect found nothing. After three decades, millions of patients, and over a hundred studies we can say that definitively the “July Effect” is a myth.

The Pitch

On June 17, Midjourney announced a pivot from making AI slop videos to medical imaging: a new division called Midjourney Medical, and a flagship product called The Midjourney Scanner. In a much-hyped video, they show a person descending into a water-filled chamber surrounded by ultrasound transducers. Magically, the device produces a whole-body scan in just 60 seconds.

Midjourney’s stated ambition is to deploy 50,000 scanners over six years and deliver a billion full-body scans a month, starting with a flagship “Midjourney Spa” planned for San Francisco’s Union Square at the end of 2027, with scanners installed alongside hot tubs, saunas, and cold plunges.

Despite the fact that the device has only scanned about a dozen people and is not FDA cleared, the internet was quick to label it “disruptive,” “life-saving,” and a “game changer.” We’ve heard this startup pitch before: a quick, simple, cheap test will radically improve our health and prolong our lives. What’s the evidence that the Midjourney Scanner will finally deliver on it?

I know something about this topic. I’m a critical care physician who uses ultrasound every day in the ICU to care for people with critical illness. I’ve consulted for companies that design ultrasound machines for years, I’ve done a lot of work with AI and signal processing, and I previously built a technology company using high-frequency sound for clinical monitoring. All of that makes me very skeptical of Midjourney’s claims.

The Physics Ultrasound Can’t Get Around

Ultrasound doesn’t penetrate bone, nor does it pass through air well. This makes imaging the brain and lung parenchyma essentially impossible. It can’t see into bone, which limits its utility for musculoskeletal imaging. Bowel gas - air in the GI tract - frequently makes it impossible to fully visualize abdominal organs like the stomach, intestines, and pancreas.

These limitations come down to acoustic impedance. Ultrasound works by sending sound waves into tissue and listening for the echoes that bounce back at every interface where acoustic impedance changes. Soft-tissue interfaces (liver-to-kidney, muscle-to-fat) have modest impedance mismatches, so enough sound reflects back to build a useful image. Bone and air sit at opposite extremes of that spectrum. Bone’s impedance is so much higher than soft tissue that almost all the sound energy reflects right off its surface, producing a bright white line and then nothing but shadow behind it — which is why ultrasound can tell you where a rib is but not what lies beyond it. Air does the reverse: its impedance is so much lower than tissue that the beam reflects almost entirely at the tissue-air boundary, scattering chaotically rather than producing a clean image. That’s why aerated lung looks like noise on ultrasound, and why gas-filled bowel loops can obscure the pancreas, stomach, and other organs sitting behind them.

Another issue: fat attenuates the ultrasound signal. Obesity can substantially degrade image quality when visualizing deeper structures. Obesity isn’t an edge case: roughly 40% of U.S. adults have obesity, and the figure is higher still among the population most likely to seek out wellness screening. A scanner whose image quality degrades meaningfully in a large share of its target market is a real limitation, not an corner case to wave away.

Any “whole body” claim has to contend with the fact that a meaningful fraction of the body - the entire skeleton, the brain, the lungs, and much of the abdomen - sits in ultrasound’s blind spot for reasons that are physical, not technological. No amount of chip miniaturization or AI reconstruction changes the underlying physics of how sound interacts with bone, air, and fat. Ultrasound isn’t really amenable to a “whole body scan” if it can’t visualize so much of the body.

The Tank Doesn’t Replace a Skilled Operator

Ultrasound exams are often dynamic. If you’ve had one, you may have been asked to roll or move so the sonographer could find a better window into the heart, liver, or gallbladder. By cleverly rotating or compressing, a skilled operator can find acoustic windows that a static setup never would.

The Midjourney Scanner immerses the patient in a tank of water to try to achieve something similar — and to be fair, this is a real trick: we sometimes use a bucket of water ourselves to visualize joints in the hand. But it introduces problems of its own. The tank needs to be cleaned between patients (more on this below). Bubbles clinging to skin can degrade resolution. And not everyone can tolerate being immersed in a tank. None of these is a show-stopper on its own, but together they will limit both applicability and throughput.

Resolution Isn’t the Same as Diagnosis

Ultrasound’s theoretical resolution is very high, but that isn’t the same thing as being able to identify structures. Ultrasound depends on differences in acoustic impedance to generate an image, and cancer and other pathology aren’t always obvious that way. This is fantastic for the heart and for many organs — thyroid, kidney, liver — ultrasound is great. For others — pancreas, colon, stomach — it’s much less reliable. That makes it hard to see ultrasound tomography replacing existing modalities (MRI and CT), especially for cancer screening.

To illustrate the point, compare the Midjourney tomography images with an abdominal MRI. The MRI is sharper and far more clearly delineates the abdominal organs.

Ultrasound also isn’t really one modality, it’s many: B-mode, M-mode, 3D modes, several flavors of Doppler. Together these give you information about blood flow and dynamic processes like cardiac motion that a single static image can’t. It’s not yet clear how many of these techniques the Midjourney Scanner can do; so far they’ve only demonstrated low-quality static images, though Doppler and other modes are presumably possible. Adding them would improve the clinical value of the scan, but it would also add scan time, since switching between pulse sequences isn’t free. Physics dictates a tradeoff between scan quality and time, which means today’s low-quality one-minute scan is likely to become tomorrow’s slightly-higher-quality scan that takes several minutes.

The Economics Don’t Add Up

Ultrasound tomography might be lower resolution, but if it delivers scans faster, could that drive costs down?

Theoretically, a one-minute scan would let far more people use a scanner each day. But much of the time spent on clinical imaging is moving patients in and out and cleaning the scanner between them. Midjourney would presumably need some way to clean the water between users — you wouldn’t want to get into a liquid scanner after someone with C. diff or another communicable disease. If that means draining and refilling the bath each time, a 1-minute scan followed by a 15-minute “clean cycle” would cut throughput dramatically, and degrade the unit economics along with it.

The numbers Midjourney itself has published only sharpen this problem. A billion scans a month across 50,000 scanners works out to roughly 670 scans per scanner per day — one every two minutes, around the clock, with zero downtime for cleaning, maintenance, or the inevitable queue of people drying off. Even setting aside infection control, that’s an extraordinarily optimistic throughput assumption for a device that requires lowering a human body into and out of a water bath.

And the unit Holz showed off costs real money to build: licensing alone runs $10 million a year just for the imaging chips, before accounting for the tank, the lowering mechanism, the water-handling and filtration system, the real estate, and staff. (That’s also substantially more than the $1-3 million for an MRI scanner). The fact that the company is framing this as a “spa” rather than a clinic or a medical device is itself a tell about how they’re managing the regulatory and reimbursement problem, which brings us to the next issue.

One way to bring costs down is to automate image interpretation with AI. AI is genuinely good at certain narrow medical-imaging tasks. But there’s no massive training set for this new modality, and I wouldn’t expect AI to be reliable at reading these scans until the company has accumulated millions of them, paired with gold-standard scans and biopsies to validate them. Until then, they’ll need human radiologists reading every scan. This is a strange position for an AI company to be in. Midjourney’s entire reputation rests on a model trained on billions of existing images. The scanner produces a genuinely novel imaging modality, but there’s no comparable corpus of full-body ultrasound tomography paired with ground-truth diagnoses for a model to learn from. Building that dataset, at the resolution and diversity needed for an AI to reliably flag a pancreatic mass or an early aneurysm, would take years and a research infrastructure that looks nothing like training an image generator. Until then, “AI-powered diagnosis” is mostly aspirational, and the cost structure looks a lot more like traditional radiologist squinting at blurry images than like a software-margin automate tech business.

Where’s the Evidence?

Everything in medicine is supposed to be based on evidence. Proving that low-dose CT lung cancer screening saves lives took over a decade and, at more than $250 million, and was the most expensive trials the NIH has ever funded. Where are the comparable studies for this new modality? So far, just hype, and without evidence, insurance won’t pay.

Holz has been candid that the technology is early: by his own account, only about a dozen people had been scanned by launch, and the device has no FDA clearance of any kind. That’s not a minor technicality. It’s the difference between “device shown not to cause harm and to do roughly what it claims” and “device shown to work in a product demo.”

There are two ways any diagnostic test can fail: a false negative (false reassurance) or a false positive (needless anxiety and workup). Both are harmful, and low-quality images make both more likely. Someone with a “clean” wellness scan may feel falsely reassured and skip evidence-based cancer screening they actually need, missing a curable early-stage cancer. Someone with an incidental finding may face follow-up CT scans, ionizing radiation, and biopsy complications chasing something that would never have caused symptoms.

False positives and false negatives aren’t only a patient-safety issue, they’re a massive liability for the company. Missed diagnosis is the leading cause of medical malpractice claims, and a “whole body” scan implicitly promises to catch whatever’s there. With blurry, low-quality images, the risk of retrospective diagnosis is real: in hindsight, something will always have been visible. Even a 1% false-negative rate, applied across a billion scans a month, isn’t a hypothetical liability problem.

Ultrasound Isn’t Automatically “Safe” Just Because It’s Non-Ionizing

There’s a common assumption that because ultrasound doesn’t use ionizing radiation, it’s inherently safe. That’s not quite right. Sound waves can heat tissue, and the edges of the beam can produce mechanical stress on cells. We don’t know how much acoustic power the Midjourney Scanner delivers, but with 40 Butterfly transducers running simultaneously, the cumulative thermal and mechanical exposure could be significant — especially for more vulnerable tissue, including a developing fetus.

FDA-cleared ultrasound devices follow the ALARA principle (as low as reasonably achievable), and fetal ultrasound in particular is specifically designed to minimize thermal and mechanical effects. A whole-body scanner necessarily uses much greater total power, and it’s not clear it can be engineered with those same safeguards. Midjourney could simply warn against scanning during pregnancy — but in a casual spa setting, how many early pregnancies, especially in the first trimester (the most vulnerable window) will go undisclosed or unnoticed before someone steps into the tank?

“I’ve Bought This Horse Before”

The scan may be quick, but interpretation will be slow as long as it depends on human radiologists. Ultrasound tomography can’t image large parts of the body, so it’s hard to see it replacing whole-body MRI. There will always be wellness-minded early adopters willing to pay out of pocket for a low-quality partial-body scan — but wider adoption will require more than hype.

Bloomberg’s Mark Gurman, who covered the launch, called it Midjourney’s “big hardware debut,” built around the explicit goal of “upending the MRI machine market.” That’s a big claim for a device that, by its own founder’s account, can’t yet point to clinical validation, regulatory clearance, or a viable cleaning-and-throughput model. None of this means the underlying engineering isn’t impressive, or that semiconductor ultrasound-on-chip technology isn’t a real and valuable advance; Butterfly’s chip technology has legitimate clinical uses. But “impressive engineering demo” and “device ready to replace MRI as a population-health screening tool” are very different claims.

A decade ago, a Silicon Valley company promised to improve health by making routine testing easier, more convenient, and cheaper. The company said that widely available testing using its revolutionary technology, placed in local health centers, would save lives. Its founder promised that easier medical testing would usher in “a world in which no one ever has to say goodbye too soon.”

That company was, of course, Theranos. Its founder is now in federal prison.

There’s no reason to think this latest Silicon Valley health-tech venture is fraudulent. But the hype is nearly identical, and the failure to live up to it feels almost inevitable.

VO₂max is best understood not as a single trait, but as the product of a chain, as described by the Fick equation:

VO₂max = Cardiac Output × Arteriovenous O₂ Difference

Or, expanding the Fick equation more fully: pulmonary ventilation × extraction efficiency × hemoglobin concentration × cardiac output (stroke volume × heart rate) × capillary density × mitochondrial oxidative capacity. Each of these steps can be a bottleneck. In the short term, training, and in the long term evolution, can widen any these bottlenecks.

Humans: Exceptional, But Not Maximal

Humans are remarkable endurance animals. Elite distance runners can sustain 75–85% of VO₂max for the duration of a marathon, a feat of metabolic efficiency that almost no other mammal can match over comparable time scales. Our capacity for thermoregulation via eccrine sweating, combined with upright bipedal locomotion that dissociates respiration from stride frequency, likely gave early Homo a decisive advantage in persistence hunting: the ability to run prey to exhaustion across open savanna in midday heat. That hypothesis (which I’ve explored in a previous post) may also bear on why human brains are so large relative to body size.

In terms of raw VO₂max numbers, elite performance by sport looks roughly like this:

Elite rowers can have a a VO₂max of 75 ml/kg/min and the highest absolute VO2max of up to 7 liters/min; however we typically normalize for body weight and express VO₂max in ml/kg/min, which tends to slightly penalize heavier athletes like rowers. (As we’ll see this also has big implications across species…) Elite swimmers’ VO₂max can be 80 ml/kg/min, and distance runners’ up to 85 ml/kg/min. Cyclists and cross country skiers are the highest recorded with values as high as 90 and 96 ml/kg/min, respectively. Cross-country skiers hold the human record for good reason: their sport demands simultaneous maximal engagement of upper and lower body musculature, maximizing the venous return and cardiac output that constrain peak oxygen delivery.

Human athletic achievement is impressive. Elite marathoners sustain ~75–85% VO₂ max for hours. Humans can outrun horses in extreme heat and over distance. But evolution has optimized humans for endurance efficiency rather than for maximum oxygen flux. How do we compare to other animals and what can we learn about physiology from their adaptations?

Sled Dogs: The Mammalian Ultra-Marathoners

Every March, teams of sled dogs depart Willow, Alaska on the Iditarod: a 938 mile race to Nome, run in legs totaling roughly 16 hours of effort per day, over 8–11 days (the current record is 7 days, 14 hours). The dogs run in temperatures that can drop below −40°C, consuming 10,000–12,000 kcal per day. They do not develop the rhabdomyolysis that would destroy a human athlete attempting equivalent work.

Published VO₂max values for sled dogs reach 198 ml/kg/min. Unpublished measurements on Iditarod-trained sled dogs report up to 200–240 ml/kg/min — roughly 2.5 to 3 times the best human values.

How?

The sled dog’s adaptations hit nearly every node in the cascade simultaneously. Cardiac output is enormous, driven by a massive stroke volume. Red blood cell mass is high at baseline, and splenic contraction during exercise releases a stored erythrocyte reserve, acutely boosting oxygen-carrying capacity in a way that amounts to endogenous blood transfusion. Metabolic flexibility is extraordinary: sled dogs oxidize fat at rates that would be impossible in humans, shifting to near-complete fat dependence within the first day of sustained effort and maintaining that state for days without the glycogen depletion that floors human performance (”hitting the wall”).

Their rhabdomyolysis resistance is worth its own note: Dogs running the Iditarod accumulate muscle damage markers that would indicate severe injury in a human, yet somehow recover between legs and finish the race. The mechanisms (enhanced heat shock protein expression, differences in membrane repair kinetics, local anti-inflammatory adaptations in type I fibers) are not fully elucidated, but may someday lead to clinically relevant insights.

The Pronghorn: Running from a Ghost

The pronghorn antelope (Antilocapra americana) is the fastest land mammal in North America and, by most accounts, the second fastest in the world. It can run 11 km in 10 minutes, sustain an average speed of 65 km/h for distances that far exceed the cheetah’s sprint range, and hit peak speeds above 100 km/h.

Its VO₂max, measured by Lindstedt and colleagues in a landmark 1991 Nature paper using an inclined treadmill protocol, was 306 ml/kg/min, approximately 3× that of comparably sized mammals. Several dramatic adapations make this possible. Normally lung volume scales linearly with the mass of an animal. The pronghorns TLC is about 6-10 liters, which is massive for an antelope weighing 40-60kg (roughly 200% what you would expect for allometric scaling) and the trachea is disproportionately wide to minimize resistance. The pronghorns heart is also oversized relative to mass-matched ungulates. Hemoglobin concentration is high (50% hematocrit), and the capillary density in locomotor muscle is dense even by elite-athlete standards.

The leading hypothesis for why the pronghorn is so overbuilt for speed relative to anything currently living in North America is that it is not, in fact, overbuilt; it evolved to outrun an extinct predator. The Pleistocene North American ecosystem included Miracinonyx, a cheetah-like felid with comparable sprint performance, alongside American lions, dire wolves, and short-faced bears with sprint capacities that no extant North American predator matches. The pronghorn’s aerobic excess is a phylogenetic ghost, the survival adaptations to a evolutionary pressure that no longer exists.

Bats: The Ceiling of Mammalian Physiology

With over 1,400 recognized species representing roughly 21% of all mammal species worldwide, bats (order Chiroptera) are the second most speciated mammalian order after rodents. Flight is an extraordinarily successful ecological strategy, though unlike birds, bats are largely invisible and inaudible to human senses despite their ubiquity.

Flight is also expensive. Bats must sustain very high metabolic rates during flight. The result is that bats hold the mammalian VO₂max record among measured species: values range from around 200 to 400+ ml/kg/min, with highly active nectar-feeding species like Glossophaga soricina (the Pallas’s long-tongued bat) approaching 400 ml/kg/min during hovering.

Bats have the smallest alveoli of any mammal with the largest the alveolar surface area along with the thinnest blood gas barrier distance, suggesting that they have optimized lung geometry for maximal diffusion.

To circulate this oxygen, bats possess relatively larger hearts than comparably sized mammals, with increased mitochondrial and vascular densities and prominent perivascular adipocytes to store energy. They have wider T-tubules enabling faster excitation contraction coupling, with heart rates up to 900 bpm.

Regulating blood flow is crucial, and bats have a highly specialized peripheral vascular system. The wing arteries branch into muscularized arterioles that can regulate flow and maintain arteriovenous pressure differential. The wing also contains contractile muscularized venoules that act as peripheral pumps to facilitate blood return. The fact that bats sleep upside down adds additional hemodynamic adaptation: some bats have arterial valves to in the aorta to avoid backwards flow when upside down). Finally bats have two vena cave with muscular zones around them to regulate venous return. These adaptations allow the bats’ wings to function simultaneously as a propulsive surface, a radiator, and a high-flow capillary bed suspended at the end of a long, narrow vascular circuit operating at extreme pressure and flow gradients.

Additionally bats have a very high hematocrit; hematocrit values above 70% have been recorded in Tadarida brasiliensis and Miniopterus minor, which is the highest measured in any mammal. To compensate for the added viscosity, bat RBCs are tiny: just 35-45 fL. More numerous, but much smaller RBCs are more costly to synthesize but enables very high oxygen capacity without excessive viscosity.

Finally, the cristae in bats’ mitochondria are especially convoluted, enabling oxygen consumption at a higher rate compared to other mammals. (Convergently, this optimization is shared across all fliers including bats, birds, and insects…)

In short, every aspect of a bats geometry is optimized to maximize surface area.

Yet despite these incredible adaptations the bat’s VO2max values fall short of comparably sized birds. Two factors appear to limit bats specifically. First, the wing membrane must remain tensioned across elongated finger bones rather than across a fused, keratinized structure, which imposes constraints on the musculoskeletal mechanics of the thorax. Second, and more intriguingly, bats may be thermally limited, the aerobic activity of flight generates enormous heat loads, and mammalian thermoregulation mechanisms are less efficient at dissipating that heat than avian mechanisms.

This thermal hypothesis points toward one of the genuinely unexplained puzzles in comparative biology: why do bats and birds divide the day with such striking consistency, with bats overwhelming nocturnal and birds overwhelming diurnal? The temperature differential between day and night, and its interaction with the heat budget of active flight, may be part of the answer. Cooler ambient temperatures at night may allow bats to operate closer to their aerobic ceiling without overheating.

In case you were wondering how to measure a flying bat’s VO2max: a closed cycle wind-tunnel is used. Bat Wind-tunnel by AirflowSciences.

Hummingbirds: The Vertebrate Limit

If bats defined the mammalian performance ceiling, hummingbirds define the overall vertebrate limit. Hovering flight, the hummingbird’s defining behavior, is the most metabolically expensive sustained activity known in vertebrates. A hovering ruby-throated hummingbird (Archilochus colubris) has a mass-specific VO₂ (650-1,090 ml/kg/min) of that exceeds every other measured vertebrate, with a heart rate approaching 1,200 beats per minute.

The adaptations that make this possible are densely layered. The avian respiratory system is categorically superior to the mammalian one, and this matters enormously at the high end of aerobic performance. Birds use a parabronchial lung with unidirectional airflow and cross-current gas exchange, achieving oxygen extraction efficiencies of 40–50% per breath, compared to the mammalian tidal lung’s ~25%. This cross current architecture with continuous unidirectional airflow is far more efficient. This is an improvement over the mammalian tidal lung, which may be rate-limiting at peak performance: elite human athletes can develop exercise-induced arterial hypoxemia at maximal effort, suggesting that pulmonary diffusing capacity is a genuine bottleneck in mammalian performance that birds have largely circumvented.

The hummingbird’s metabolic demands are prodigious: It consumes up to double its mass each day and it’s blood glucose reach up to 740 mg/dL to provide energy for the exercising muscle. (Understanding how the hummingbird avoids glycosylation damage from this high glucose may someday lead to new therapies for humans with diabetes…)

The geometry of everything in the hummingbird’s oxygen cascade is optimized: The hummingbird heart is about the size of a pencil eraser, but this is proportionally enormous being about 2.5% of the birds mass (by comparison the human heart is just 0.5%). Capillary density in flight muscle is extreme. And the mitochondrial volume fraction in flight muscle approaches 30–35% of total fiber volume. This is near the theoretical limit: The hummingbirds muscle is so dense that adding more mitochondria would displaces sarcomeres and reduces force-generating capacity. Likewise, adding more capillaries would reduce the contractile cross-section. The hummingbird has found the geometric optimum under the avian body plan constraints. To go any higher, you need a fundamentally different architecture.

Bumblebees: When the Lungs Are the Problem

There is a story, variably attributed NASA, Boeing, or other universities, that aeronautical engineers looked at the bumblebee and concluded that aerodynamically it is impossible to fly. The story is apocryphal, and obviously bumblebees can fly, but the metabolic bioenergetics of bumblebee flight genuinely does look like a math error.

A bumblebee beats its wings at 200+ times per second and achieves one of the highest mass-specific VO2max of any organism: 900-1200 ml/kg/min. Other bee species, such as the Euglossine bee hold the actual record, with VO2max as high as 2,567 ml/kg/min. (but I just love that fake bee/NASA anecdote…)

How does a bee achieve this massive increase over other fliers like hummingbirds and bats?

Simple: insects do not have lungs. They breathe through spiracles - ten pairs of valve-like openings along the thorax and abdomen - that connect to a branching network of air tubes called tracheae. These tracheae deliver oxygen directly to flight muscle cells, bypassing the entire hemoglobin/cardiac output bottleneck that constrains every vertebrate on this list. There is no diffusion barrier imposed by alveolar membrane, no hemoglobin saturation curve to navigate, no cardiac output ceiling. Oxygen moves from air to mitochondria through a progressively narrowing system of tubes by a combination of convection and diffusion.

The consequence is that the bumblebee can devote a remarkable fraction of its body mass to flight muscle, which consumes approximately 90% of total oxygen during flight. The thorax is essentially a metabolic engine wrapped around wings.

This architecture works because of scale. Direct tracheal diffusion is practical at 0.1 grams and impractical at any larger size. Diffusion distance scales with the square of body radius, while oxygen demand scales with volume. A bumblebee-sized vertebrate with tracheal breathing would suffocate; a bee-sized vertebrate with lungs would be wasting structural mass on infrastructure that diffusion could handle directly. The bee sits at the scale where you can simply route the tubes to the cells and skip the middle steps. Bees are living example of the engineering maxim that “the best part is no part.”

The bumblebee VO₂max sits near the theoretical physical ceiling; the rate at which oxygen can diffuse down a concentration gradient through tissue is set by Fick’s law.

Conclusion: The Struggle for Surface Area

The cube-square law is the physical constraint that makes this struggle necessary in the first place. As an organism grows, its volume increases as the cube of its linear dimensions while surface area increases only as the square. Oxygen demand scales cubically with volume. Oxygen delivery scales squared with the surface area of lung, capillary, and mitochondrial membrane across which diffusion occurs.

This mismatch is the fundamental challenge evolution has invented novel solutions to overcome. Every adaptation we’ve discussed is a clever geometric solution to a surface-area problem: The avian parabronchial lung increases the effective respiratory surface per unit volume compared to the mammalian lung. The dense capillary beds of the pronghorn and hummingbird muscle maximize the surface across which oxygen can diffuse into cells. The densely folded inner mitocondrial membranes of of bat cardiomyocytes, are themselves a surface-area solution, increasing the area available for oxidative phosphorylation within each mitochondrion. The bumblebee’s tracheal system is the logical endpoint: eliminate the centralized gas-exchange organ entirely and route the tubes directly to the cells, so that the surface for oxygen delivery is distributed throughout the tissue that needs it.

Every order of magnitude increase in body size makes the surface-to-volume ratio worse, and every organism above a certain size has had to solve that problem with increasingly elaborate architecture. As Haldane said “Comparative anatomy is largely the story of the struggle to increase surface in proportion to volume.”

This week, social media lit up with alarm. Politicians, prominent physicians, and health policy commentators posted breathlessly about a crisis: doctors are abandoning medicine in droves. The evidence? Two new papers on physician attrition from clinical practice.

Clearly, none of these commentators bothered to read the Methods.

Both papers examined physicians who had already left clinical practice and described their characteristics: a portrait of departures, not a prophecy of collapse. And yet a single Kaplan-Meier plot generated some truly spectacular misreadings. Among the more memorable: that “half of all doctors quit within 10 years,” or that no one is “still work[ing] in clinical medicine within 25 years.” These are not at all what the data show.

So what does the data actually say?

The papers describe approximately 5% annual attrition from clinical medicine. Let’s put that number in context before deciding whether to panic.

Across professional jobs broadly, annual turnover runs 10–18%. Among registered nurses, the workforce physicians are constantly compared to, it’s around 16%. In the tech industry, you’re looking at 15–20%. Five percent is not a crisis. Five percent is, if anything, surprisingly low for any knowledge profession in the modern economy.

More importantly, attrition from clinical medicine should not be zero:

It is entirely normal and healthy for physicians to move from clinical practice into leadership, research, and medical education roles. Healthcare systems need physician executives, clinical trialists, and department chairs.

It is equally normal for physicians to move into biopharma and tech, industries that desperately need people who understand how drugs work and how prescribers think. How can we hope to develop AI solutions to make physician jobs better without physicians driving tech development? (I know something about this because I’ve spent years doing exactly this...)

Some physicians leave clinical practice to write. Anton Chekhov was a practicing physician. So were Arthur Conan Doyle, Michael Crichton, Siddhartha Mukherjee, and Oliver Sacks. We do not typically describe their career arcs as a workforce tragedy.

Physicians also go into public health, government, and, as at least one of the more hysterical posters about this data has done, into politics. If anything, we need more physicians in government advocating for sensible health policies.

Then there’s retirement. Physicians aged 65 and older currently make up 20% of the clinical workforce, with another 22% are between 55 and 64. A substantial share of annual departures simply reflect people reaching the end of long careers. It is also normal - and morally appropriate - for physicians to leave clinical practice to care for young children or aging parents. We should not build a workforce model that treats caregiving as a leak to be plugged.

Finally, some people leave because clinical medicine isn’t right for them, and that’s okay too. It’s a shame to watch someone exit after the enormous investment of medical school and residency. It is a bigger shame to watch someone stay in a job that makes them miserable because they feel bound by sunk costs. One of the great things about clinical training is it can be used in non-clinical contexts.

Five percent is fine.

But is everything fine?

No. Physician attrition from clinical practice has increased over the past decade, and not all of the increase is benign. Some of it reflects genuine growth in legitimate alternative opportunities: biopharma and med tech are booming, and the workforce is aging. But some of it reflects burnout, and the data here are harder to dismiss.

Physician burnout hit a peak of 53% in 2022, spiking to a record 62.8% during the COVID-19 pandemic. The good news is that it has been declining since: 41.9% of physicians reported at least one burnout symptom in 2025, down from 43.2% in 2024 and 48.2% in 2023. The bad news is that it is uneven. Some specialties (emergency medicine, urological surgery, and hematology/oncology) are all approaching or exceeding 50% burnout rates, meaning nearly half of physicians in some of the highest-demand specialties are running on fumes.

And burnout isn’t just a wellness metric, it translates directly into departures. A 2024 MGMA survey found that 27% of medical groups had a physician leave or retire early that year due to burnout. Physicians working with incompletely staffed teams were more than twice as likely to report burnout, and understaffing itself predicts intent to cut hours or leave entirely, creating a vicious cycle as departures beget departures.

The cost of burnout-driven departures is estimated at $4 billion annually in the United States.

Perhaps most telling is how the reasons for leaving have shifted. The study with the misunderstood Kaplan-Meyer plot found that early departures in 2008 were more likely to be driven by personal health issues, malpractice insurance premiums, and lack of professional satisfaction. The updated data show a shift toward burnout, chronic workplace stress, administrative burden, and unrealistic patient expectations. Burnout and the resulting departures are less about individual circumstances and more about systemic conditions.

That’s the story worth telling. Not “doctors are quitting at rates comparable to baristas and therefore civilization is ending,” but “a measurable and growing fraction of physician departures are burnout-driven, the conditions driving burnout are structural and identifiable, and we know what they are: administrative burden, understaffing, flat reimbursement, and the erosion of professional autonomy.”

The misread Kaplan-Meier curve doesn’t help us understand any of that. If anything, the panic it generated crowds out the more important conversation with noise.

So why not automate it?

Robots are well suited for repetitive mechanical tasks, able to perform the identical motion with perfect tempo. It’s a compelling idea. Currently available devices include the Lund University Cardiopulmonary Assist System (LUCAS, Stryker Medical), the AutoPulse (ZOLL Medical Corporation), and the Thumper (Michigan Instruments). And yet, despite decades of development and clinical use, mechanical CPR devices haven’t improved outcomes in any clinical trial.

Not only that, 6 observational studies and 2 RCTs have found worse neurological outcomes with mechanical CPR:

The observational evidence is consistent and damning. The largest comes from the CARES national registry: Buckler and Rossano analyzed over 80,000 cardiac arrests and found that mechanical CPR was associated with lower neurologically favorable survival and strikingly, EMS agencies that never used mechanical CPR at all had the best outcomes. Liao et al reached similar conclusions using the Utah CARES registry, finding a relative risk of just 0.41 for functional survival after propensity-score adjustment in patients receiving mechanical CPR. Hayashida et al prospectively followed over 6,500 patients across 67 Japanese hospitals and found that mechanical CPR in the emergency department was independently associated with lower ROSC, lower hospital admission rates, and lower survival to discharge. Putzer et al found a 30-day survival rate of just 5.1% with LUCAS versus 16.3% with manual CPR in an unselected Austrian OHCA population. Kim et al, drawing on a nationwide South Korean registry of over 20,000 patients, found that every mechanical device tested was associated with lower ROSC, with AutoPulse specifically linked to significantly worse survival to discharge.

It’s tempting to attribute these findings to confounding: perhaps sicker patients or longer codes are selectively getting the machine? But the randomized data tells the same story. The ASPIRE trial was stopped early after mechanical CPR patients had nearly half the survival to discharge of manual CPR patients (5.8% vs 9.9%). PARAMEDIC, the largest RCT, found no improvement in 30-day survival and lower rates of favorable neurological outcome at 3 months. Among survivors who were followed long-term, MMSE scores were modestly but significantly lower in the LUCAS group, a signal that even patients who survived may have paid a cognitive price for mechanical CPR.

We should also consider the counterfactual. If mechanical CPR was really better, wouldn’t we expect to see improved outcomes with it somewhere? And yet, despite decades of use and dozens of studies - both randomized and observational - none has demonstrated a compelling benefit to mechanical CPR.

So why doesn’t the tireless machine achieve better outcomes than the fallible, fatiguable human?

Interruptions Kill

Even brief interruptions in compressions markedly reduce coronary perfusion pressure. Edelson et al. found that a 10-second interruption of CPR prior to defibrillation reduces the rate of ROSC by almost 50%. (This is why charging the defibrillator before the rhythm check matters so much - you minimize the pause between “stop compressions” and “shock.”) Every second of no-flow time is a second the brain isn’t being perfused, and it takes time to restore perfusion pressure after each interruption.

Putting on a mechanical CPR device takes time - time during which no compressions are being delivered.

A 2007 study found that the mean time to set up the AutoPulse™ was 4.7 minutes, though it’s unclear how long compressions were stopped for. A 2012 study by Yost et al. found that compressions were paused a median of 32.5 seconds (IQR 25–61 seconds) to apply the LUCAS device. That means more than a quarter of patients had no compressions for over a minute while the team strapped on the machine.

Maybe it’s a training problem? Steffen et al. trained medical students to rapidly apply a mechanical CPR device, which reduced the deployment time from a median of 37.0 seconds (IQR: 29.0–42.0) in the control group to 26.0 seconds (IQR: 20.0–30.0) after training. That’s still a long interruption. And in a follow-up assessment four months later, the time had crept back up to 34.5 seconds (IQR: 24.0–45.8), suggesting that the benefit of training is both minimal and transient.

The Machine Isn’t Even Better at the Mechanical Part

Here’s the counterintuitive finding: the machine frequently fails at the one thing it was built to do.

A 2011 manikin study compared conventional versus mechanical CPR as performed by 21 ambulance crews. Only 58% of compressions delivered by the LUCAS device were adequate, compared to 88% with manual CPR. And only 12 of the 21 crews (57%) correctly applied the mandatory stabilization strap.

Why? Mechanical CPR devices frequently become displaced during a resuscitation. A device that has shifted even a few centimeters may be compressing the upper abdomen instead of the sternum. Malpositioned CPR is worse than tired CPR.

There’s also a subtler problem. Human rescuers naturally vary the force they deliver, which allows adequate chest recoil between compressions. Mechanical devices deliver metronomic compression, which sounds like a feature, but may actually impair venous return if there isn’t sufficient recoil time. The heart needs to refill between compressions; if it can’t, cardiac output drops regardless of how fast or consistently you’re pushing.

This was demonstrated in a pediatric cardiac arrest study that randomized piglets to either mechanical or conventional CPR. Survival was significantly higher in the conventional CPR group (50% vs. 15%, p = 0.016), suggesting humans may be doing something mechanically beneficial.

Mechanical CPR Causes More Injuries

A meta-analysis of 11 trials found that mechanical CPR was more likely to cause injuries than manual CPR. Rib fractures, sternal fractures, liver lacerations, and in the case of a displaced device, injury to abdominal organs. I’ve personally seen this: a malpositioned mechanical CPR device doing vigorous, regular compressions directly over a pâtéd liver.

Beyond displacement and injury, there’s the specter of device failure itself. A 2025 study examined 255 failures of the LUCAS device reported to the FDA. These included fractures, soft tissue injuries, and injury to internal organs. Device failures occurred in 54% of reports, in many cases leading to prolonged interruptions in CPR. This is a failure mode totally unique to mechanical CPR; even tired humans don’t suddenly freeze and prevent compressions from being performed.

Where Mechanical CPR Does Make Sense

There is a role for mechanical CPR. There are situations where human CPR cannot be safely or practically delivered: on a sled coming off a ski slope, in a helicopter, or under fluoroscopy in the cath lab. In those austere or logistically constrained settings, an imperfect device beats no compressions at all.

The key word is substitute, not supplement. Mechanical CPR should be used where humans cannot do the job, not as a default upgrade over humans who can.

Are These Machines Worth the Cost?

The LUCAS 3 costs approximately $15,000 per unit. Add training, maintenance, and replacement parts, and the total cost of ownership is substantially higher. Most EMS systems have multiple units across multiple vehicles.

Now ask a different question: how much community CPR training could be purchased for the same investment? Bystander CPR performed by imperfect, sweaty humans demonstrably saves lives: It roughly doubles survival rates from out-of-hospital cardiac arrest. Training programs cost on the order of $30–50 per person.

Every dollar spent on a device that demonstrably doesn’t improve outcomes is a dollar not spent on community CPR training, which does.

Clinician Educator Tracks (CETs) have become very en vogue in residency and fellowship programs. Walk through the websites of academic medicine departments and you’ll find them everywhere: structured pathways for residents who want to become the next great medical educator, complete with projects, portfolios, and certificates.

The impulse behind CETs was good. For decades, great teachers in academic medicine were invisible. The system rewarded researchers and high-volume clinicians. Teaching was something you did in the gaps, unpaid and unrecognized. CETs were born from a legitimate grievance: we should recognize great teachers and value their time.

But something went wrong. A reasonable idea, recognizing and reward excellent educators, metastasized into a different beast entirely: that we can create great educators by putting residents through a track. That we can identify, at the PGY-2 or PGY-3 level, who is going to be a great clinical teacher and accelerate their development. That’s a flawed premise, and it’s worth examining why.

Great Teachers Are Born, Not Trained

Some people are genuinely great at explaining things.

The latin word educere means “to pull out.” Great teachers are intuitively great at getting students to engage and make connections. They have a feel for where someone is confused, a gift for analogy, an instinct for when to slow down and when to push. These are not skills that transfer easily from a curriculum. You either have that quality or you develop it slowly over years of practice and feedback, not from a a few seminars in your early years of training.

More fundamentally: to be a great clinical teacher, you have to be a great clinician first. This is the part that CETs gloss over. You can’t decide to be clinically excellent. You can’t will someone into being a top-5% clinician by enrolling them in a “pathway.” Clinical excellence is earned over thousands of patient encounters, through pattern recognition that takes years to build, through developing judgment that only comes from having been wrong in ways that mattered.

If you’re not good at airway management, how will you teach airway? If your ultrasound is mediocre, what are you passing on to your learners? If your physical diagnosis is sloppy, you’re not teaching physical diagnosis, you’re teaching sloppy physical diagnosis. The content of clinical teaching is clinical excellence.

So when a resident in year two identifies as an “educator” and channels their energy into a teaching track, the question worth asking is: what exactly are they going to teach? On what foundation?

Of course anyone can teach. Learning and sharing your learning with others is the essence of medical education. In this field, genuine curiosity is a super power. If you dedicate time to learning about a topic, that effort does make you a good teacher - at least on that topic. If you are the type of person who generally channels your energy into learning, teaching is a natural outgrowth. This process does not depend on CETs, nor is it restricted to people involved in CETs. Curiosity is an innate characteristic, not a skill taught in CETs.

Confusing the Artifacts of Education with Education Itself

There’s a pattern in CETs that’s worth naming. Residents in these tracks spend a lot of time making PowerPoints, designing curricula, and drafting learning objectives. The output looks like education. But education is not PowerPoint, and possessing a sufficient number of slide decks does not make someone a good educator. These are the artifacts of teaching, not the thing itself. Confusing the work product with the work is a trap, and CETs make it easy to fall into it.

There’s an analogy in surgery. If you put a medical student in a suturing workshop and had them practice knot-tying and wound closure, they’d get better at suturing. But you wouldn’t call them surgeons. The technical artifact - a closed wound - is not the same as the judgment, decision-making, and pattern recognition that make a surgeon. Education works the same way.

Formal pathways close doors for everyone else

Another issue with CETs is that residents/fellows who do not participate may be viewed as “unqualified” to be educators.

I remember when I was fleeing a failed physician-scientist pathway, I briefly explored pivoting to the educator track. I was told that I “lacked a teaching portfolio.”

The irony is that I had hundreds of teaching files - including many of my OnePagers, handouts, and POCUS teaching - but the content wasn’t formatted into the proper curricula model demanded of the CET.

In retrospect I am immensely grateful: leaving academic medicine was one of the best decisions of my life. (More on this in a future blog perhaps). It is a bit ironic that the content that would largely define my professional career “wasn’t right” to be an educator in a formal CET.

When I think back on my co-residents, the best educators were not the ones in those tracks. They were generally brilliant people who were also good educators. They did not consider “educator” to be their primary identity, but rather as one part of their job as physicians.

The Track Analogy

Consider what it would mean to have a “famous surgeon track” in high school. A student decides, at 16, that they’re going to be a great surgeon. They read about surgical technique. They practice on simulators. They write papers about operative approaches. And then they apply to medical school with a certificate in pre-surgical excellence.

You immediately see the absurdity. The ambition is great, but the differentiation is occurring before there’s anything to differentiate on. The identity precedes the substance.

That’s what happens when an intern joins an educator track: the label comes before the expertise. The identity - “I identify as an educator” - is conferred before there’s a clinical foundation deep enough to build on.

What Does the Evidence Say?

If we take CETs seriously as an intervention, we should ask: does formal educator training actually make better teachers?

The best data comes from K-12 education, where this question has been studied rigorously. The landmark work from Eric Hanushek and colleagues shows that holding a master’s degree in education, the most common advanced teaching credential, produces essentially no measurable improvement in student achievement. Large-scale studies from The New Teacher Project (TNTP) found that certified teachers didn’t outperform alternatively-certified or uncertified teachers in the first few years, which is precisely when you’d expect the gap to be largest if the credential mattered. The strongest evidence comes from randomized evaluations of Teach For America, where corps members with minimal formal education training performed comparably or better than certified peers in the same schools, particularly in math.

The CET literature in medicine is far less rigorous and far more self-congratulatory. A systematic review found only 36 unique curricula described in the literature across all of GME from 2008 to 2021. Most reported outcomes focused on learner satisfaction or self-reported behavior change. None had a control group. None described objective evidence of improved knowledge or skills as a result of clinician-educator curricula. The evidence base consists almost entirely of participants saying they found the programs valuable.

What Should We Do Instead?

There is value in teaching people to be better educators. Pedagogy is a real discipline. You can learn to give better feedback, structure a teaching session more effectively, and use deliberate practice in clinical instruction. These things matter at the margins.

But the sequence matters enormously. Pedagogical training has the most value when it’s delivered to people who already have deep clinical expertise: faculty who’ve earned their content knowledge and now want to be more intentional about how they transmit it. That’s the population who can actually use it.

The right model is this: identify faculty who are clinically excellent and demonstrably great teachers because your residents tell you, because you’ve watched them teach, because their learners thrive. Recognize them formally. Give them protected time. Fund their educational work. Have them teach everyone else how to be a better educator. Then the best of those will rise further. That’s how you build an educator faculty.

What you don’t do is ask residents to declare a teaching identity before they’ve built the clinical foundation that clinical teaching requires. The answer to undervaluing great educators was never to manufacture them earlier while using CET credentialism to gate-keep teaching positions. It was to recognize and protect the ones who had already become great.

Focus on the results of teaching rather than the artifacts of teaching. The deliverable is not a portfolio - it’s learners who understand a topic.

Imagine a time traveler from 120 years ago dropped suddenly into the present. The world would be baffling: how we communicate, travel, work, and live would be nearly unrecognizable. Smartphones. Commercial aviation. The internet.

But there’s one arena that our hypothetical time traveler might find reassuringly familiar: medical education.

Superficially, many aspects of medical education look very similar today as in the past. In some cases, the same topics are taught in the very same rooms.

Thomas Eakins captures the vibe of 19th century medical education perfectly in his painting depicting a surgery by Dr. David Agnew. Dr. Agnew is undoubtedly one of the greats of surgery, but if we examine Eakins’ painting closely we can see that Agnew’s lecture may be a bit soporific. In his audience we see one man checking his watch, several with eyes closed, two quietly cuddling, and one who has gone full horizontal.

We can - and must - do better than this.

The Traditional Toolkit, and Its Limitations

To understand why change is needed, we first need to honestly assess our existing tools. Consider the lecture, from the Latin lectura, meaning reading. This format was developed in the medieval university, when books were scarce and literacy was low. One person reading to a group was a pragmatic solution since it solved both problems. But over the centuries, as books became plentiful and literacy commonplace, the lecture format persisted, particularly in medical education.

The lecture is timely and can be interactive, but attendance is capped by room size, and learning requires showing up at a specific place and time. The content, more often than not, dies in the room.

Review articles and textbooks may reach a broader audience - though not always - but they come with two serious problems. First, cost. Textbooks are expensive, paywalled reviews are expensive, and this cost is prohibitive to learners around the world who aren’t fortunate enough to train at well-resourced institutions. Second, latency. The gap between writing and publication is typically several years. In fast-moving fields like oncology, the content can be out of date before the first printing.

Journal club is perhaps the most underrated traditional format: interactive, timely, and generative. But it’s locked to a single institution, available only to those with the time and resources to attend. Like a lecture it is synchronous, requiring learners to attend at a particular time.

Each of these formats has real strengths. But taken together, they share a common failure: they reach too few people, cost too much, and update too slowly.

The Modern Alternatives

Here’s the good news: for every traditional format, there’s a modern equivalent that addresses these limitations directly.

The core insight is that the content hasn’t changed, the vessel has.

Think about what this means in practice. Instead of (just) giving that great lecture to a few dozen people in a room, you record it and post it as a podcast or YouTube video. If the residents in your program love it, I guarantee that a hundred times more people want to hear it. Instead of writing a paywalled review article, summarize it in a tweetorial or an infographic. Instead of one person at a time asking questions in office hours, hundreds learn from a public thread.

And instead of spending time writing a textbook chapter that will be obsolete on arrival — you decline that invitation.

Why the Modern Alternatives Are Genuinely Better

This isn’t just a difference in format. Social media enhanced medical education is better, on nearly every dimension that matters.

• Scale. The most popular medical podcasts - including mine - have tens of thousand listens per episode and have been listened to over a million times. I could give the same lecture to a new audience every day for a decade and not reach nearly so many learners.

• Asynchrony. Learners can engage anywhere, anytime - not just in a lecture hall at a particular hour. This is not a minor convenience. For a resident post-call, a physician in a rural clinic, or a medical student in a country without a major academic center, asynchrony is the difference between access and no access.

• Interactivity. Not only can you reach a massive audience, the audience can reach you for comments, questions, pushback, and discussion. The conversation flows both ways.

• Latency. There are no publication delays. This makes social media-enhanced education ideal for fast-evolving topics, exactly the ones where traditional formats fail most egregiously.

• Cost. With a few notable exceptions, most of this content is free, with no paywall, tuition, $200 textbook, or conference registration fee.

• Expertise, democratized. Perhaps most importantly: learners can now learn about a topic from the world expert, not just from whoever at their institution happened to be tapped to prepare grand rounds. A trainee in a small rural critical access hospital has access to the same voices as a resident at a prestigious urban academic medical center.

• A new path for early-career educators. Social media allows people early in their careers to rise based on the quality of their ideas and their teaching, not solely on institutional affiliation or academic rank. The barriers to entry are a microphone and expertise. (This is certainly true of me.)

To put it simply: social media-enhanced medical education has the potential to reach a larger audience, with shorter latency, greater interactivity, and at lower cost compared to traditional didactic methods.

Social Media as a Stage for Professional Discourse

There’s another dimension worth highlighting, and one that’s easy to overlook.

A few years ago, Nature polled researchers on why they use different social media platforms. The results were striking. Unlike Facebook or LinkedIn, researchers use Twitter specifically to find peers, discover new papers, follow discussions, and engage in debate. It’s a professional commons.

This matters enormously for education, because one of the most valuable things a learner can do is watch experts think.

As a trainee, I rarely got to see my attendings reason through uncertainty in real-time. If I was lucky, I might witness a genuine professional debate once every few months at a particularly animated conference or journal club. But on social media, I get to watch smart colleagues engage in thoughtful, public discussions almost every day. I get to see how a new study can change someone’s mind. I get to see how experts disagree: respectfully, rigorously, and with evidence. This discourse is also a window into an important concept in medical education and the practice of medicine: epistemic humility - the virtue of acknowledging one’s own knowledge, understanding, and bias. Watching - or participating - in real-time as medical concepts are discussed, and as the limits of our collective understanding are reached, may be one of the highest value exercises that a clinical can offer or experience publicly.

Consider instead of the lecture, the older The Socratic method: an argumentative dialogue between individuals. In contrast to the passive lecture, Socrates’ audience leaned forward, amplified his gestures, and were clearly unafraid to express strong agreement or disagreement. People loudly expressing opinions in a public forum? That sounds a lot like social media!

On the morning of June 4, 1993, ground crew at Miami International Airport made a discovery that shouldn’t have been possible. Tucked into the wheel well of a Douglas DC-8 that had just completed the two-thousand-mile flight from Bogotá, Colombia, was a seventeen-year-old boy. He was covered in frost. His extremities showed signs of frostbite. But he was alive, even though physiologically this should have been impossible. 5 hours earlier, Juan Carlos Guzmán Betancur had hidden near the runway threshold in Bogotá, and climbed into the space behind the right main landing gear strut, a space just four feet wide by three and a half feet deep, barely large enough to accommodate a small adult pressed against hydraulic lines. When the gear retracted after liftoff, metal covers closed over the opening, sealing him inside. As the aircraft climbed to 35,000 feet, the temperature and oxygen levels dropped precipitously. This should have killed him three different ways:

• Hypoxia: At 35,000 ft, PO₂ is ~37 mmHg. The threshold for sustaining consciousness is ~60 mmHg. Pilots exposed to this pressure have a time of useful consciousness of 15-30 seconds. When elite mountaineers summit Everest (which is somewhat lower at 29,029 feet) they do so at the absolute limits of human physiology, and require weeks to acclimatize. When these climbers arrive at the summit, their hypoxic ventilatory response has increased respiratory rate to defend alveolar PO₂; erythropoietin has stimulated red cell production; 2,3-DPG levels have shifted the oxyhemoglobin dissociation curve. Even with weeks of acclimatization, they can only survive by massively hyperventilating; driving arterial CO₂ down to ~8 mmHg, near the physiological floor for sustainable breathing effort. The stowaway in a DC-8 wheel well was neither an elite athlete nor was he acclimatized, yet somehow he survived an oxygen tension much lower than the summit of Everest!

• Hypothermia: The ambient temperature at 35,000 feet is –55°C (-67°F), colder that a cold winter day in Antarctica. Accidental hypothermia often produces cardiac arrhythmia and death below ~28°C core temperature. The coldest recorded accidental hypothermia survivor had a core temp of 13.7°C, and only survived due to ECMO! The DC-8 stowaway received no life support, but somehow survived prolonged exposure to a profoundly colder environment!

• Decompression sickness (DCS) and nitrogen gas embolism: The drop in barometric pressure causes dissolved nitrogen to evolves from tissues. This is the same pathology that kills scuba divers who ascend from high pressures too quickly. Sudden exposure to low pressure at altitudes above 20,000 feet can also cause DCS. The risk of DCS increases with altitude, with the risk approaching 100% above 30,000 feet. At 35,000 feet barometric pressure is ~20% of that experienced at the Earth’s surface (179 mmHg compared to 760 mmHg at sea level). Without a pressure suit at 35,000 feet, the nitrogen should have bubbled out of his blood, yet somehow he survived without developing DCS or gas emboli.

Juan was lucky but this was not a miracle. Between 1947 and 1993, the New York Times documented ten separate incidents of stowaways who concealed themselves in the unpressurized landing gear bays of commercial aircraft on international flights. Of these 10 flights: Five stowaways survived and six died (one flight carried two stowaways where one died and another survived). A larger FAA report, identified 113 cases of wheel-well stowaways, which included 27 survivors. Clearly stowing away on a flight is suicidally dangerous, but somehow dozens of people have survived it. How, physiologically, does a human being survive conditions that are incompatible with life?

That question has vexed physiologists for decades. The answer is that the three physiological derangements (the hypoxia, low barometric pressure, and hypothermia) experienced at 35,000 feet may interact in a way that is actually protective. The putative mechanism involves several steps:

Step 1: Hypoxia disables thermoregulation

The preoptic anterior hypothalamus functions as the brain’s thermostat and is exquisitely sensitive to hypoxia. As PaO₂ falls during ascent, this structure loses the ability to defend core temperature. The body becomes poikilothermic: temperature tracks the environment rather than being regulated against it. The fact that most of the survivors were thin and lightly clothed may also facilitate this the rapid drop in temperature, though the warmth of hydraulic fluid and wheels, may blunt the drop in body temp slightly, providing a “grace period” keeping them alive while still concious.

Step 2: Hypothermia suppresses metabolic demand

Core temperature falls toward ambient. This would be catastrophic under normal circumstances but oxygen consumption falls with lower temperature. In ectotherms (animals that do not tightly regulate their body temperature), oxygen consumption falls proportionally to temperature. This concept, termed the the Q10 temperature coefficient is 2.2 for cerebral metabolic rate of oxygen (CMRO₂), means that oxygen consumption is roughy halved for every 10°C drop in temperature.

Thus, when core temperature reaches ~27°C (the threshold for unconsciousness) brain oxygen consumption is suppressed to ~45% of baseline. Deep hypothermic circulatory arrest (DHCA) is a technique used in the operating room that exploits this principle to perform aortic arch repairs, cooling the brain as low at 15°C to prevent anoxic brain injury.

Step 3: The ODC left-shift compounds the benefit

In addition to reducing metabolism, hypothermia also shifts the oxyhemoglobin dissociation curve to the left. The left shift increases hemoglobins affinity for oxygen, enhancing oxygen loading in the lungs despite the lower atmospheric oxygen pressure. This is the same result as seen with increasing 2,3-BPG in people acclimitized to altitude, except the effect is larger: at 0°C, the affinity of hemoglobin is 22 times greater than at 37°C. This means that despite the profoundly low atmospheric oxygen pressure, the body’s oxygen-carrying capacity is being partially rescued by the fall in temperature. The hemoglobin loads up what little oxygen is available and holds onto it tightly.

Step 4: Solubility of gasses increases with hypothermia

Normally, exposure to hypobaric conditions would cause decompression sickness as the nitrogen dissolved in blood comes out of solution. The risk of DCS is significant when exposed suddenly to altitudes above 30,000 feet. Fortunately however, Henry’s Law, the principle governing gas solubility in liquids, states that solubility increases with decreasing temperature; Cold liquids hold more dissolved gas than warm ones. Compared to 37C. Water at 0C dissolves approximately 3x more oxygen, and 2x more nitrogen. The cold may retaining dissolved nitrogen in solution, rather than allowing it to nucleate into bubbles, causing DCS. This explains one of the more puzzling features of the survival cases: the survivors presented primarily with frostbite and hypoxic sequelae, not the joint pain, cutaneous mottling, or neurological deficits that would be expected from severe altitude DCS.

Step 5: Rewarming and repressurization is gradual

As the aircraft descends, ambient pressure rises and temperature slowly increases. The stowaway, still unconscious or barely conscious, begins to rewarm. Reoxygenation occurs. However the rewarming phase introduces its own lethal hazards. Cold myocardium is electrically unstable; the dysrhythmias that hypothermia suppresses during cooling can manifest explosively on rewarming. In clinical DHCA management, rewarming rate is controlled, antiarrhythmic agents are on hand, and defibrillators are ready. In a wheel well, none of this is available. The heart is rewarming at whatever rate the ambient temperature recovery dictates, with no pharmacological support.

Simultaneously, the nitrogen solubility advantage that cold may have conferred on the way up is being reversed. Rising temperature reduces gas solubility, potentially allowing previously-dissolved nitrogen to come out of solution in tissues that are now being reperfused as cardiac output recovers. The combination of reperfusion and nitrogen evolution is the precise mechanism of decompression sickness.While the cold may have prevented DCS by increasing gas solubility, as the aircraft descends and core temperature rises on rewarming, nitrogen solubility decreases again. This means that DCS can occur during rewarming. Something analogous is seen in the OR. In deep hypothermic circulatory arrest (DHCA), blood and tissues dissolve higher amounts of gases (nitrogen/oxygen). Rapid rewarming after DHCA causes these gases to come out of solution faster than they can be eliminated, forming intravascular bubbles. Because the wheel well re-pressurizes faster than it warms, the risk of DCS appears to be lessened. The descent profile of a commercial aircraft is not the same as a controlled decompression schedule, but apparently the change in balanced to prevent DCS.

And then there is the most immediate hazard. As the aircraft approaches landing altitude and the gear is lowered, any stowaway who has recovered enough consciousness to be mobile and now risks falling from the open wheel well.

Who Survives and Why

In the NYT dataset, five stow-aways survived and six who did not. The reasons why illuminate just how narrow the physiologic window is. According the the FAA report, fatal outcomes among wheel-well stowaways cluster into three categories: those who fell from the aircraft before or during gear lowering (mechanical), those found frozen (overcooling past the survivable range), and those who likely failed to cool fast enough before irreversible hypoxic brain injury. The FAA report proposes a few factors that may influence survival

• Age and body composition: Survivors tended to be “youthful, thin, non-exercising individuals” who may be less prone to DCS (lower body lipid content for nitrogen partitioning) and acheive faster peripheral cooling; less thermogenic muscle mass means faster entry into the protective hypothermic state before hypoxic injury occurs.

• Warm hydraulic fluid and tire friction: heat from the tires and warm hydraulic fluid in lines likely provides some conductive and radiant heat. These sources of warmth are “progressively diminished as flight at altitude proceeds” meaning there may be a thermal grace period during ascent. This may actually be crucial: it slows the rate of cooling during the most vulnerable phase when consciousness is still present and the hypothermic protective effect hasn’t fully engaged.

• Lightly clad: In most reports, survivors came from tropical countries and were described as “lightly clad.” Counterintuitively, poor insulation may accelerate cooling toward the protective hypothermic state faster once at altitude, rather than prolonging a dangerous intermediate temperature range.

• Luck: stowing away in a wheel well is profoundly dangerous; the FAA report identified a 76% fatality rate. The July 26, 1969 DC-8 flight from Havana to Madrid illustrates how razor thin the margins are. On this flight, two stowaways climbed into the same wheel well. A seventeen-year-old boy was found alive at Madrid. His companion did not survive. There were exposed to identical conditions (temperature, pressure, aircraft) individual physiological variation was sufficient to determine survival.

The people who attempt this dangerous mode of transport are desperate, not irrational. Nietzsche writes “He who has a why to live for can bear almost any how.” Wheel well stowaways are fleeing poverty, political instability, or family separation. 17 Year old Juan Carlos was trying to reach an aunt in Miami. Armando Socarras Ramírez, the 1969 Havana–Madrid survivor, was fleeing Castro’s Cuba. The recent January 2025 JetBlue case involved an 18-year-old and a 16-year-old from the Dominican Republic, found dead in the landing gear at Fort Lauderdale. The dataset keeps growing because the conditions that produce it haven’t changed.

Is Survival a Three Stooges Syndrome?

Wheel-well stowaways are simultaneous exposed to three potentially fatal insults: hypoxia, hypothermia, and hypobaric pressure. Their survival seems physiologically impossible, but it is exactly because these insults are co-occurring that they can survive.

In one episode of The Simpsons, doctors inform Mr. Burns that he is host to every disease known to medicine, but that each is so perfectly balanced against the others that none can touch him. Like the Three Stooges entering a room, Mr Burns’ many diseases can’t “get through the door” to harm him.

It is precisely because of the concomitant exposure to multiple extreme conditions at once that makes wheel-well stowaway survival possible. Juan Carlos survived his 5 hour physiologic ordeal not despite being simultaneously hypoxic, hypothermic, and hyperbaric, but because it occurred simultaneously. Mr. Burns would understand.

Scientific publishing is among the strangest markets in the world. Researchers produce a scientific manuscript, submit it to a scientific journal, and pay to publish it. Other researchers review it for free. Universities then pay billions of dollars to access it. The companies in the middle (Elsevier, Springer, Wiley) make margins above 38%, rivaling even the most lucrative of tech companies. In return, these middle companies provide little more than a submission portal, some typesetting, and a brand name. The system was designed for a world of printing presses and physical journals. It has no business surviving in 2026.

The dysfunction runs deeper than cost. The structure of scientific publishing actively distorts the science itself: incentivizing sensational findings over careful ones, punishing negative results, creating years of unnecessary delay between discovery and dissemination, and concentrating gatekeeping power in a small number of editors whose decisions are almost entirely unaccountable. A medical treatment that could save lives may sit in a submission queue for eighteen months while its authors navigate sequential rejections.

Reform efforts have been halting and incomplete. Open access mandates transferred cost from universities to authors without reducing it. Preprint servers like arXiv and medRxiv helped, but remain disconnected from the formal review and publication process. SciHub has made accessing articles easier - by illegally bypassing paywalls - however, the authors and reviewers remain uncompensated.

No one has redesigned the system from scratch. This is proposal for how we could.

The Problems

I. Sequential Submission Wastes Years
Authors submit to one journal at a time. Rejection triggers a new submission cycle, with new reviewers who have never seen the paper. A paper rejected by three journals before acceptance may have consumed years and a dozen reviewer-weeks, producing nothing but delay. A 2021 systematic review of biomedical publications found the mean time from submission to publication ranged from 91 to 639 days at different journals.

We need to speed up the submission/review process.

II. Publishers Extract Value Without Creating It
The labor (writing, reviewing, editing) is donated gratis by the research community. A 2020 study estimated that a typical researcher spends 100 hours per year on uncompensated peer review. In aggregate researchers contribute 100 million hours per year, totaling billions of dollars in uncompensated labor. Publishers charge institutions for access to this donated labor. APC (article processing charge) models added a second revenue stream: now authors also must pay to publish open access. Science charges authors $5,450 for open access, NEJM $11,690, Nature $12,850.

We need to compensate the people doing the work: authors and reviewers.

III. Reproducibility Is Optional
Most published papers do not include the underlying data or analysis code. Scientific data described as “available on request” is, in practice, frequently inaccessible, with studies indicating that up to 93% of authors who use this phrase either do not respond or decline to share their data. Independent replication, where it is attempted, frequently fails. The incentive structure of journals rewards novelty over rigor; negative and null results have no viable path to publication.

We need to ensure that scientific data is available for re-analysis and that negative studies are published.

IV. Reviewer and Editorial Power Is Unaccountable
A desk rejection requires no justification and generates no record. Editors can suppress inconvenient findings, protect preferred paradigms, or simply neglect submissions indefinitely. None of this is visible. The latter, a “pocket veto” can delay resubmission for months. Reviewer accountability is similarly absent; a reviewer can accept an assignment and silently ignore it for months.

Fair review is also important; a 1991 article analyzed reviewers who peer reviewed more than 25 times and found that 1.5% of were assassins (consistently negative) whereas 1.2% were zealots (consistently positive). Neither assassins nor zealots make good reviewers.

Conflicts of interest are often hidden; a 2021 study found that 45% of authors do not disclose conflicts, and author COI are only typically elicited after acceptance. Editors and reviews may not be required to disclose COI.

We need to transparently hold editors and reviewers accountable.

V. Incentives Are Misaligned at Every Level
Impact factor rewards journals for publishing citable work, not important work. Researchers are evaluated on publication count and journal prestige, not on contribution to knowledge. Prolific reviewers receive no formal recognition. The system optimizes for metrics that are gameable proxies of quality (e.g. citation count) rather than quality itself.

We need a robust system that recognizes the quality of work done by both authors and reviewers.

VI. Format Balkanization Is a Hidden Tax
Each journal has its own reference format, word limits, figure specifications, and submission portal. Reformatting a rejected manuscript for resubmission is a significant time cost with zero scientific value added. A 2019 study estimated the average researcher spend 14 hours per manuscript, or 52 hours per year reformatting.

We need a modern system for efficient layout and formatting.

A Ten Point Plan to A Fix Scientific Journals

1. Simultaneous Submission to Multiple Journals
The overall acceptance rate for scientific articles is about 30%. This means that the typical manuscript is reviewed by 2-3 reviewers 3-4 times prior to publication, generating 6-12 reviews of the same work. The reviews are sequential, duplicative, and often subjective.

Instead imagine if authors submitted a manuscript once and designated up to ten journals in ranked order of preference. Peer review occurs once, by a single larger assembled panel (4-5 reviewers), with results shared across all the designated journals. Journals compete to accept the work; authors are not forced to guess which journal will find their work acceptable before a single reviewer has read it. This avoids “wasting” repeat peer reviews with each resubmission. It provides a less subjective assessment of the work because more reviews are obtained, so an outlier review is more obvious. It also enhances the overall efficiency of peer review because the typical paper gets just 4-5 reviews instead of 6-12.

Submitting to multiple journals at once creates a bias to action by editors. The author’s pre-specified rank list governs by default, but authors retain the right to override it if their top-ranked journal has not responded within six weeks of all other reviews being complete. This creates a structural incentive for editorial speed without mandating it.

2. Automatic Preprint Upload on Submission
Every submission is automatically deposited to a discipline-appropriate preprint server at the time of submission (e.g. arXiv, bioRxiv, etc). Science enters the public record the moment it is ready for scrutiny, not the moment a journal decides it is worthy. The preprint is versioned; the accepted manuscript supersedes it, but the record of submission date is permanent and citable.

3. Mandatory Data, Code, and Reproducibility Standards
Every submitted manuscript must include all data and analysis code necessary to reproduce its results. Every figure and table must link directly to the code that generated it. Limited exceptions apply (patient privacy, national security, proprietary clinical assays) but these must be explicitly declared and approved, not assumed.

This is the single most important structural change for the quality of the scientific record. Without it, peer review is evaluation of a summary rather than evaluation of the actual science.

4. Standardized Submission Formats
A small set of submission templates covers the major article types: randomized trial, observational study, methods paper, systematic review, case series, brief communication. All journals on the platform accept all standard formats. Layout is rendered from a single source file using LaTex. Authors format once for their article type, not once per journal submission. Discipline-specific consortia could define additional format variants within the standard framework but the overall goal is reduce the number of competing formats.

The submission, peer review, editorial workflow, and hosting all run on a single open-source software platform. A shared platform eliminates the redundant cost of every publisher maintaining proprietary submission portals and editorial systems, money currently extracted from institutional subscription fees. It also creates a powerful migration incentive: publishers who adopt the platform gain dramatically streamlined workflows, better submission tracking, and access to the shared reviewer pool. The software is a public good; improvements made by any consortium benefit all of them. Hosting costs are pooled, providing economies of scale.

5. Online-Only Publication
There is no print edition. This reduces cost, latency, and landfill waste. There are no page limits imposed by the cost of ink. Supplementary materials are not second-class content buried in inaccessible PDFs — they are part of the article. Figures are full resolution. Video and interactive data visualization are first-class outputs. Individual articles (or complete journals) can be printed on demand, but the era of printed and mailed science is over, saving hundreds of thousands of tons of paper waste.

6. University Subscription Fees Fund the Ecosystem
Institutions pay tiered access fees based on size and research output, replacing both journal subscription fees and APCs. This pool funds platform operations, editorial infrastructure, and reviewer compensation. Authors pay nothing to submit. Readers at subscribing institutions pay nothing to read. A public access tier makes all articles freely readable after twelve months. Reviewers can use accumulated credits to access articles.

7. A Credit System Compensates Reviewers and Authors
Subscription revenue is distributed through a structured credit system. Reviewers earn credits for completed reviews, weighted by quality ratings from editors and authors. Authors of highly impactful work earn additional credits through a revenue-sharing mechanism: for papers exceeding a field-normalized citation threshold (e.g. the top 1% within discipline, with self-citations excluded) a share of subscription revenue is allocated proportionally.

This rewards genuinely transformative work (a paper cited thousands of times by independent researchers) without rewarding incremental work in high-volume fields. An important design consequence: editors have a direct financial incentive not to miss a potentially important paper. The cost of a wrongful desk rejection is now visible and quantifiable.

8. Full Editorial and Reviewer Transparency
Every journal’s acceptance rate, desk-rejection rate, median time-to-decision, and reviewer completion rate are public, updated in real time. Every editor’s decision record is public; not the content of decisions, but the pattern: how often they desk-reject, how often accepted papers are later retracted, how often they passed on papers that subsequently became highly cited elsewhere.

Reviewers are similarly tracked: assignment acceptance rate, completion rate, time-to-review, and quality ratings (by editors). Chronic non-completers lose eligibility to earn review credits. There are no more “pocket vetoes.” A reviewer who accepts an assignment and misses the deadline is flagged; the editor must reassign within a defined window. Conversely, prolific and punctual reviewers get public recognition for the quality work that they do, with visible stats that can influence promotion.

In addition to the stats, the public profile for each editor/reviewer shows their conflicts of interest, which is updated continuously. The current publishing paradigm requires a COI statement from authors only, and only at the time of publication. In addition to ties to industry, the profile shows a timeline of grants, patents, etc providing a complete assessment of a reviewer’s potential conflicts. In this new system, everyone involved in the manuscript, including authors, reviewers, and editors, would have a profile that includes any and all COIs. This page could integrate with other databases such as NIH funding, clincaltrials.gov, and Sunshine Act, providing automatic updates and additional transparency.

9. A Structured Post-Publication Review Process
The historical failure of post-publication critique has been an incentive problem, not a cultural one.

Post-publication peer review is integrated into the platform, not bolted on. Registered researchers can submit formal methodological critiques, data re-analyses, or replication reports. These are reviewed, credited, and linked permanently to the original article. Authors have a defined right of response. Editors can escalate serious concerns to an independent panel empowered to recommend correction, retraction, or expressions of concern.

Post-publication review earns credits at rates comparable to pre-publication review, meaning that major methodological flaws that bypass peer review still get noted. With greater transparency, journals have a stronger incentive to address concerns about fraud and reproducibility.

10. Governance: A Consortium of Consortia
The platform is governed by a federation of discipline-specific consortia (e.g medicine and health sciences, physical sciences, life sciences, social sciences, etc) with subgroups for more specific topics (genetics, randomized controlled trials, topology, etc). Each consortium sets field-specific norms: format variants, citation thresholds for revenue sharing, reviewer qualification standards.

Above them sits a small central body, responsible only for platform infrastructure, cross-disciplinary standards (authorship, data sharing, retraction protocols), and antitrust protection. This central body has no editorial power. It cannot be acquired by a commercial publisher. Its charter requires supermajority approval from member consortia to amend.

This structure guarantees that the model can be flexible to changing needs, unlike the current system, while avoiding capture. There are precedents such as arXiv and ORCID.

How do we built this?

Change is difficult, and replacing an entrenched and ossified system like scientific publishing is especially hard. None of these 10 points is technically difficult, and several aspects (simultaneous submission, preprint servers, data repositories, open-source review software) already all exist in partial form. The obstacle is not engineering. It is the will do tear down the existing paid publication system and rebuild it thoughtfully and fairly. The question for the scientific community is whether the aggregate cost of the current system, measured in delayed treatments, un-replicable findings, wasted researcher hours, and institutional subscription fees, has finally exceeded the political cost of replacing it.

In his famous monologue, Hamlet marvels at what mankind has become. “What a piece of work is a man! how noble in reason! how infinite in faculty! in form and moving how express and admirable! in action how like an angel! in apprehension how like a god! the beauty of the world! the paragon of animals!” But what if Hamlet had the causation backwards? What if it is precisely our admirable “form and moving” - our bodies - that enabled our “noble reason” and “infinite faculties”? Specifically, could distance running be what made human cognition possible?

There are an estimated 6.5 million terrestrial animals. Roughly 2/3 of these are predators. But of the millions of predators on land, just 5 are classified as persistence pursuit predators — meaning instead of ambushing prey, they engage in prolonged pursuit, running their quarry to exhaustion over miles and hours. In this sense, we really are the paragon of animals.

The argument is that that running and intelligence are not merely correlated. The same biological adaptations that made us exceptional endurance hunters created the physical substrate that a large, metabolically expensive brain requires; running built the hardware that enabled our superior wetware.

The Body Built for Running

Bramble & Lieberman (2004) identified approximately 26 features in Homo anatomy specifically adapted for endurance running: the nuchal ligament — a thick band at the back of the neck that stabilizes the head during a running stride, enlarged gluteus maximus, shortened toes, arched foot, narrow waist, and semicircular canal enlargement for gaze stabilization. These features appear in Homo erectus — the same lineage showing significant encephalization — the evolutionary process by which our brains grew disproportionately large relative to our body size.

The timing is significant. These skeletal changes appear in the fossil record roughly 1.5 to 2 million years ago, in concert with the first substantial increases in brain size. This is unlikely to be mere coincidence. The Endurance Running Hypothesis posits that the many of the skeletal adaptations that differentiate Homo from other primates were based on running. What if these anatomical changes enabled the cognitive changes?

Running Built the Brain’s Infrastructure

The human brain is metabolically extraordinary, consuming approximately 20% of resting cardiac output despite being just 2% of body mass. Sustaining an organ this hungry requires an exceptional circulatory system. The human cardiovascular system has several unusual features: relatively large heart-to-body ratio, high stroke volume capacity, and crucially the ability to maintain some cerebral perfusion even during near-maximal exertion. These were perhaps initially adaptations for our unusual hunting style (e.g., keeping tissues oxygenated during a two hour pursuit), but they may have become essential for our enlarging brains to function.

The eccrine sweating system and loss of body hair that enable persistence hunting also protect a large brain from heat damage. Neurons are fragile at high temperatures; the brain begins to malfunction above roughly 104°F (40°C). Most large mammals solve this problem with panting, but that limits simultaneous breathing during sustained exertion. Humans solved the cooling problem differently. The selective brain cooling mechanism, via the cavernous sinus and cooled venous blood returning from facial evaporation, is unusually well developed in humans. It is plausible this thermoregulatory system was selected for distance running, but incidentally enabled a large brain to operate safely for sustained periods in the hot African savanna.

Effective hunting also improved the availability of calories and the quality of food. Being able to take down large ungulates (antelope, wildebeest, and early horses) provided a stable supply of nutrition, including nutrients and micronutrients essential for brain development. Hunting provided early humans with dense animal fat, critical for brain lipid composition: DHA, arachidonic acid, and organ meats with micronutrients (B12, zinc, iron) that are limiting for neural development. A brain that cannot be fed cannot grow. Running gave us the means to feed it.

The Hunt Demanded More Than Legs

The relationship between running and cognition was almost certainly not one-directional. The demands of persistence hunting likely created selection pressure for greater intelligence in its own right. Persistence hunting is cognitively demanding in ways that pure ambush hunting is not. It requires tracking and interpreting spoor over hours, mental modeling of prey behavior and terrain, sustained goal-directed attention over long timeframes, prospective memory and planning (“the animal will head toward water when it overheats”), and coordination within a group. Perhaps the need to convey these more complex ideas was even the genesis of language.

The anthropologist Louis Liebenberg has argued extensively that tracking is one of the most cognitively complex tasks hunter-gatherers perform, functionally equivalent to scientific reasoning: forming hypotheses about an unseen animal’s behavior, testing them against evidence in the environment, and revising them in real time. If persistence hunting created selection pressure for these cognitive capacities, brain size could follow as a consequence. If running built the hardware, hunting refined the software.

The Company We Keep

Our choice of best friends supports the hypothesis that persistence pursuit was important. Perhaps it is not a coincidence that the wolf, the first species we domesticated, is another of the 5 persistence pursuit predators, and has a VO2max among the highest in the animal kingdom. (Sled dogs, a close relative of wolves, have a VO2max of 175-300 ml/kg/min!)

Dogs are a profound outlier in the history of domestication. They were the first species we domesticated, and by a lot! Studies suggest dogs were domesticated 15,000 to 40,000 years ago, at least 5,000 and perhaps as much as 30,000 years before the goat. Dog domestication precedes the agricultural revolution by at least 5,000 years and possibly much more, occurring during the Upper Paleolithic, by hunter-gatherers, before any other animal or plant. This strongly supports the hypothesis that dog domestication was driven by hunting utility rather than pastoral or agricultural need. We did not recruit the wolf because we needed a companion around a fire. We recruited it because it could run and keep up with us.

In choosing the wolf, we reveal what we understood ourselves to be: persistence hunters, defined by the chase. We recognized a kindred physiology and made it our companion.

Too Good to Last

Indeed, it is likely that early humans were too good at hunting. In the Late Pleistocene, the majority of megafauna (animals over 44 kg, roughly the size of a large dog) went extinct in a remarkably short period of time. Roughly 50,000 years ago, there was an extinction of megafauna in Australia and Europe. 13,000 years ago in North America, over 70% of megafauna — including mammoths, mastodons, and saber-toothed cats — vanished in less than 3,000 years. Each of these extinctions coincided with the arrival of humans.

This is the dark epilogue to the persistence hunting story. The same cognitive and physical capacities that evolution had spent millions of years refining turned out to be so effective that we systematically eliminated the megafauna that had shaped them. We outran the evolutionary niche that had built us. The prey disappeared; eventually, so did the hunter’s niche. Within a few thousand years of the North American extinctions, the agricultural revolution had begun. Running had made us intelligent enough to eliminate our own ecological role.

The Paragon of Animals

So what made us the paragon of animals? Not, as Hamlet implies, some ineffable divine spark. Rather, the physiologic adaptations that enabled our endurance played a critical role in the growth of our intelligence. The cardiovascular system that could sustain a two-hour pursuit also delivered oxygen to an enlarging brain. The thermoregulatory system that kept a hunter cool on the savanna also protected neurons from heat damage. The caloric windfall from taking down large ungulates provided the raw materials for neural tissue. And the cognitive demands of the hunt itself created selection pressure for the very faculties Hamlet was celebrating. Put another way: we did not become intelligent and then become hunters; we became intelligent because we hunted and because our bodies were built to run.

Ninety Nine Reasons and Pitch Ain’t One

One conspicuous example of a mistranslation turned myth is when students are instructed to have a patient repeat “ninety nine” while feeling for tactile fremitus. The principle is that alveoli filled with fluid conduct sound from the larger airways better than alveoli filled with air. By having the patient phonate a diphthong sound, low frequency (< 80 Hz) sound is produced, which produces the vibrations of fremitus. In the era before chest radiographs and ultrasound this may have been a useful way to identify consolidation within the lung.

The problem is that “ninety nine” doesn’t work well because it isn’t a diphthong. Imitation of the act of physical exam without understanding the reason explains this long perpetuated misperception. As described by William Dock:

When our medical ancestors studied in Austria or Germany, they observed that physicians asked patients to say neun und neunzig to evoke fremitus over the thorax. When they came home they taught their patients to say ninety-nine, thus translating literally, but not phonetically, what they had heard. This was a serious error, since their teachers would have asked patients to say: nein nein, if that was the sound they had wanted… Neun und neunzig is pronounced noyn unt noynzig and the ‘oy’ is what it takes to evoke a palpable, low-pitched vibration, most effectively transmitted from the larynx to the rib-cage. “Nein, nein” and “ninety nine” are high pitched sounds, useless for evoking fremitus.

If you want to elicit tactile fremitus, the better way to produce the desired low frequencies is by having the patient say an English diphthong like “bougie”, “toy”, “boy”, or “boat”. If they know the song “99 Luftballons”, you could always have them sing in German. Or just use POCUS.

The fremitus story is almost charming in its origins: a translation error made in good faith, propagated by imitation; a ritual performed because its practitioners didn’t understand the rationale.

Shoe Covers: A Solution in Search of a Problem

If you work in a hospital chances are you’ve put on shoe covers. This is a well intentioned practice largely predicated on a misunderstanding about the history of medicine. In order to understand this myth we need to go back to the early days of modern surgery.

The early 20th century was a dynamic time in surgery. The advent of sterile technique, effective anesthesia, and the development of new surgical tools and techniques advanced the field rapidly. Despite these improvements the operating rooms of the early to mid twentieth century posed some unique risks for patients and staff. One complication in particular was greatly feared: in the words of one witness, “Death by anesthetic explosion is given heavy emphasis... not only because it is the most dramatic of all surgical fatalities, but also because it is so unusual.” The author went on to recommend that “Conductive shoes should be worn by all hospital personnel.”

The problem was that friction between dissimilar materials led to the accumulation of static electricity, which could not be grounded if the person was wearing non-conductive footwear, such as typical rubber soled shoes. The remedy was simple, specialized conductive shoes (worn with cotton socks while standing on special conducting floors) permitted the accumulated static charge to be dissipated harmlessly.

There was a considerable industry that emerged, and several products were developed promising to mitigate the risk of OR fires due to static electricity. In the words of one inventor:

“Everyone has experienced electrical shocks after scuffing across a rug, or touching a door handle of a moving car, etc. These sparks are only minor unpleasantries in most instances, but in certain environments they are capable of setting off severe and disastrous explosions. For example, although the problem has had very little publicity and laymen are not generally aware of it, doctors know that the danger of explosion is always present in a surgical operating room because anesthetic gases, such as ether, are highly explosive. The slightest electrical spark can and has caused tragic explosions. Thus, one object of my invention is to promote safety in operating rooms by preventing these sparks from occurring.” [Pence, US Patent 3335506, 1967]

Along with specialized conducting shoes, a panoply of other inventions appeared, including charge dissipating floors, specialized beds, clothing, and other devices. Low tech approaches were also advocated, including the practice of placing wet towels around one’s shoes.

Fortunately, the flammable anesthetic era passed swiftly. The invention of the triad of halothane (1956), enflurane (1966) and isofluorane (1979) as several intravenous agents, fentanyl (1960), etomidate (1964), and propofol (1977), effectively ended the use of flammable ethers.

Soon, in an effort to save money, the disposable shoe covers with grounding leads were replaced by non-grounding covers, which can be distinguished based on the absence of the black stripe:

Even though the risk of OR explosions was largely extinguished, physicians accustomed to the ritual of donning shoe covers invented an alternative raison d’être for them: to protect their shoes from contamination and thereby reduce infection.

Unfortunately, this alternate rationale, now long accepted, is unfounded. While coverings may offer limited shoe protection, they do not, in fact, prevent surgical site infections. In fact, shoe covers may increase the risk of infections, as the act of donning or doffing over shoes can lead to hand contamination. In one study, in which the hands of 18 volunteers were examined before and after putting on over shoes, demonstrated new colonization of the hands with organisms likely obtained from the shoes. Furthermore, the idea of protecting oneself from contamination may be similarly untrue. A study by Weightman and Banfield found that, when, over a 5 day period, they collected 1387 discarded single use over-shoes, only 29% were intact when filled with water, and of the 103 with visible blood on them, 32% had blood on the inside.

So now, decades after the demise of volatile anesthetic, can we put this myth to rest? Recently, several professional organizations have advocated for shoes that remain in the surgical areas instead of donning and doffing shoe covers.

In this case the ritual persisted, kept alive by a post-hoc rationalization. As medicine advances, we must look critically on the rituals we practice and why they came about in the first place.

Radial Arterial Lines: The Finger Found It First

In virtually every ICU and operating room in the world, when a patient needs an arterial line, the needle goes into the radial artery. Ask a clinician why, and they will tell you about safety, specifically about collateral circulation, and a bedside test called the Allen’s test, which is supposed to confirm that the ulnar artery can supply the hand if the radial occludes. The story is tidy and physiologically credible. It is also largely retrospective and probably untrue.

The actual reason we use the radial artery is simpler and more honest: a finger can find it. The radial pulse sits superficially at the wrist, easy to palpate against the distal radius, and in the pre-ultrasound era this made it the obvious target. The ulnar artery runs deeper, beneath the flexor carpi ulnaris, and is harder to locate by touch alone. Palpation was the technology we had, and we chose the artery that yielded to it.

That original reason made complete sense in its era. The problem is that ultrasound arrived decades ago, and with it the ability to visualize and cannulate any vessel we choose. The palpability rationale quietly expired, but the preference for the radial did not.

Instead, it acquired a new justification: the Allen’s test. This test (pressing on both radial and ulnar arteries until the hand blanches, releasing the ulnar, and timing the return of color) was originally devised in 1929 by Edgar Van Nuys Allen to diagnose thromboangiitis obliterans (Buerger’s disease). It was not designed as a pre-cannulation safety screen. It was adopted in that role decades later on the basis of plausible reasoning rather than evidence.

The evidence, when it arrived, was not kind to the test. The landmark Slogoff et al. study in Anesthesiology followed 1,699 patients who underwent radial arterial cannulation and found that although partial or complete radial artery occlusion occurred in more than 25% of patients after decannulation, no ischemic damage to the hand or disability occurred in any patient. Ischemia is vanishingly rare regardless of Allen’s test result. A systematic review found the test has a sensitivity of 54.5% and a positive predictive value of just 53%, barely better than a coin flip for the outcome it is supposed to predict! The RADAR trial subsequently enrolled over 200 patients with abnormal Allen’s tests and found no ischemic complications compared with patients with normal results.

The anatomical picture adds another layer. Teaching holds that the ulnar is the larger artery and therefore cannulating the radial is the more conservative choice. But this depends on where you measure. At the brachial bifurcation the ulnar is indeed larger. By the time both vessels reach the wrist, the radial has given off fewer branches and is actually larger in most people — one cadaveric study found the radial 26–28% wider at the wrist. In clinical practice, though, the size relationship varies considerably between individuals: one ultrasound study of 167 hospitalized patients found the right radial was not the largest forearm artery in 46% of cases. In pediatric patients, the ulnar was larger in 60% of those studied.

What this means practically is that we are defaulting to a vessel without looking, using a screening test that doesn’t predict the complication it screens for, based on a safety rationale that was invented to replace a palpability rationale that no longer applies. The ultrasound probe, now standard of care for central venous access, could just as easily tell us which wrist vessel to use in a given patient. We need to use it.

A recent article by in Pediatric Anesthesia puts it simply: “Historically, the ulnar artery has rarely been considered for arterial cannulation as it is less easily palpated than the radial artery. With the current routine use of ultrasound in pediatric patients, the ulnar is as accessible as the radial and could be viewed as an equivalent site for cannulation.” The radial artery was chosen because a finger could find it. Now with the help of ultrasound, we have better fingers.

NPO after midnight

If you have ever had surgery, you were almost certainly told not to eat or drink anything after midnight the night before. This instruction is so universal — delivered by surgeons, nurses, pre-operative phone lines, and hospital websites — that it feels like a law of nature. It is, in fact, a rule of thumb derived from a single retrospective case series published in 1946, applied to a population it was never designed to cover, and perpetuated for decades after the evidence clearly supported a more nuanced approach.

In 1946, the obstetrician Curtis Mendelson described 66 cases of pulmonary aspiration of gastric contents during general anesthesia in 44,016 pregnant women (an incidence of 0.15%). Two patients died, both from obstruction caused by aspirated solid food. Among the 40 patients who aspirated liquid stomach contents, there were no deaths; they developed transient symptoms that resolved. Mendelson recommended withholding oral nutrition during labor and made no distinction between solid food and clear liquids.

This was a reasonable recommendation for his population of women undergoing emergency obstetric anesthesia, often with full stomachs, without modern airway management. But this specific recommendation for a high-risk group morphed into a universal rule for all surgical patients. By the 1960s, “NPO after midnight” was the global standard, applied to elective outpatient procedures in healthy adults as readily as to emergency surgery in laboring women.

The irony is that the pre-Mendelson era had it more right. Joseph Lister wrote in 1883: “While it is desirable that there should be no solid matter in the stomach when chloroform is administered, it will be found very salutary to give a cup of tea or beef-tea about two hours previously.” Nineteenth-century clinicians distinguished between solids and liquids. The post-Mendelson consensus collapsed that distinction and overcorrected.

Decades of subsequent research established that gastric emptying of clear liquids is largely complete within 1–2 hours, and that prolonged fasting does not meaningfully reduce gastric volume or acidity compared with allowing clear liquids until 2 hours preoperatively. The American Society of Anesthesiologists formally codified this in its 1999 guidelines as the “2-4-6-8 rule” allowing clear liquids up to 2 hours before surgery, breast milk 4 hours, non-human milk and light meals 6 hours, fatty meals 8 hours. Every major anesthesia society has since adopted similar guidance.

The midnight rule persists anyway. One study found that the median duration of hospital NPO orders was 12.8 hours - roughly 2 missed meals - and that nearly a quarter of those orders were avoidable by the investigators’ criteria. The consequences are not trivial: hunger, thirst, hypoglycemia, patient distress, and the metabolic stress of prolonged fasting before an already physiologically demanding urgery. Prolonged preoperative fasting is known to worsen the endocrine-metabolic response to surgical trauma, increasing insulin resistance and protein catabolism.

The midnight rule endures not because the evidence supports it, but because it is simple to communicate (”NPO p MN”), impossible to get wrong in the direction of feeding too close to surgery, and deeply embedded in pre-operative nursing workflows. How many preventable cases of hypoglycemia are experienced every year due to an old, catchy, but medically incorrect slogan?

The Routine Preoperative Chest X-Ray: Screening for a Different Era

For most of the twentieth century, patients admitted for elective surgery received a chest radiograph as a matter of routine. The rationale was entirely reasonable at the time: tuberculosis was endemic, patients might never have been previously imaged, and the pre-operative period was a natural opportunity to screen for unsuspected cardiopulmonary pathology that might alter anesthetic or surgical management. In that era, a routine chest X-ray was a useful instrument.

The problem is that the era changed. TB prevalence collapsed 100x between 1900 and 1980. Today, most patients presenting for elective surgery have had prior imaging, primary care contact, and documented medical histories. The pre-operative assessment now includes detailed history-taking, physical examination, and, for appropriate patients, targeted investigation. The routine chest X-ray became vestigial.

The evidence is unambiguous. A systematic review found that postoperative pulmonary complication rates were essentially identical between patients who received preoperative chest X-rays (12.8%) and those who did not (16%), a non-significant difference that suggests the radiograph adds nothing to surgical risk stratification in unselected patients. The AAFP, citing ACR Appropriateness Criteria, concluded that only 2% of routine preoperative chest X-rays lead to a change in management.

Current guidelines do not recommend routine pre-operative chest radiographs, except in specific circumstances (patients over 70, those with known or suspected cardiopulmonary disease, smokers, and patients scheduled for major thoracic or cardiac procedures). And yet the practice persists, particularly in lower-resource settings and in hospitals where pre-operative protocols have not been recently revised.

A recent study found that as many as 74% of pre-operative radiographs were inappropriate. So why do we keep doing them? To quote Tevya “Tradition!”

The Trendelenburg Position: The Man Who Changed His Mind, and Nobody Noticed

Placing a patient in Trendelenburg position (patient supine, head down, feet elevated) is one of medicine’s most durable reflexes. A patient becomes hypotensive; a nurse or paramedic tilts the bed. The reasoning is intuitive: gravity will drain blood from the legs into the central circulation, increasing venous return and (hopefully) cardiac output. It is the hemodynamic equivalent of squeezing the toothpase tube from the bottom.

The position was described by the German surgeon Friedrich Trendelenburg in the 1870s for an entirely different purpose: improving surgical access to pelvic organs by allowing the bowel to fall away under gravity. It had nothing to do with shock or hemodynamic resuscitation. Its career as a treatment for hypotension began during World War I, when the American physiologist Walter Cannon — one of the most eminent scientists in the country, chair of physiology at Harvard Medical School and president of the American Physiological Society — advocated for it as a field treatment for traumatic shock on the basis that it would augment venous return.

Less known is that within a decade, Cannon reversed his recommendation. He concluded that the evidence did not support the position as a resuscitative measure. The practice did not notice. It continued to propagate through textbooks, medical and nursing and paramedic training as though the retraction had never occurred.

The subsequent evidence has consistently confirmed Cannon’s revised view. The hemodynamic effects of Trendelenberg are small, transient, and accompanied by significant risks. Studies show at best a transient increases in cardiac output of just 6-9%, lasting only 3–5 minutes in acutely ill patients, with no sustained benefit in blood pressure.

A review of the literature concluded that current data “do not reveal any beneficial or sustained changes in systolic blood pressure or cardiac output.” A 2004 review in the Canadian Journal of Emergency Medicine carried the blunt title: “Myth: The Trendelenburg position improves circulation in cases of shock.”

In specific patient populations, the Trendelnburg position is actively harmful: it raises intracranial pressure, increases aspiration risk, worsens pulmonary mechanics (particularly in obese patients), and may worsen pulmonary edema in cardiogenic shock by increasing venous congestion. A 1995 review called the position “essentially useless for vascular resuscitation,” also described it as “potentially harmful in the presence of cardiac, pulmonary, ocular, and central nervous system pathology.”

Surveys continue to show that nurses and paramedics use the Trendelenburg position for hypotension in routine practice and continue to believe it to be beneficial. The man who popularized it for shock publicly recanted 103 years ago. The practice survived his retraction by more than a century!

Historic depiction of Trendelenberg position (source: Wikimedia Commons)

What These Cases Have in Common

These examples span a century of medicine and involve radically different clinical domains. In each case, a practice originated for a specific, historically cogent reason. But that reason expired, and still the practice continued.

How can we prevent ourselves from falling prey to these medical myths? The best remedy is understanding why we do what we do. Rather than being content with simplified explanations given in medical training, we should question the the actual rationale, including its historical context, the conditions under which it was established, and whether those conditions still obtain.

Medicine is, among other things, an act of transmission: each generation receives practices from the last and passes them to the next. While the transmission is usually faithful, it does not always preserve the understanding that made those practices make sense: the mechanism, the context, the original justification. When understanding is lost, the ritual continues. Instead of simply imitating we need to understand the reason for these practices.

Consider switchboard operators. In 1920, AT&T was the largest employer in the US, and over half the jobs were telephone operators. In the 1950 census the US had 151.3 million people. At that time, Bell Telephone employed 342,000 switchboard operators. Private switchboards employed roughly a million more. That means that roughly 1% of all people in the US were employed at telephone operators. Because these operators were disproportionately female, 1 in every 13 employed women worked as a telephone operator in 1950! Automation has replaced virtually all of these jobs. According to the Bureau of Labor Statistics, there are fewer than 600 people who work as telephone operators for the phone company today.

Telephone switchboard operators, Seattle 1952 (Source: Wikimedia Commons)

Similarly, in 1950 there were roughly 20,000 elevator operators in Manhattan, out of a population of 1.9 million, again yielding a prevalence of about 1%. Today fewer than 50 buildings in Manhattan have manual elevators. Aside from a few anachronistic operators in historic buildings, elevator operator is essentially an extinct occupation.

If mechanization and automation seemingly eliminated so many professions in the 19th and 20th century, will recent advances in artificial intelligence disrupt knowledge domains like medicine?

Medicine feels different. But is cognitive work like diagnosis, reasoning, and judgment really more defensible than barrel-making?

Every day we see headlines about breakthroughs by LLMs. AI demonstrates superior speed and often higher sensitivity in detecting specific, localized findings (like fractures or nodules) compared to human radiologists. GPT-5 achieved 95.2% average accuracy across all USMLE steps, and OpenEvidence’s LLM achieved a perfect score on a simulated USMLE Step 1 Exam. Models like Claude 3.7 Sonnet have shown up to 98%–100% accuracy in common diagnostic scenarios. The FDA has approved multiple autonomous AI systems for diabetic retinopathy screening that operate without physician oversight, achieving sensitivity above 90%.

I’ve done a lot of work in the area and I’m well aware of what AI can - and can’t - do.

Some aspects of our jobs can - and will - be automated. For example AI scribes already save a few minutes per patient in clinic. AI could dramatically streamline chart review and note writing. It wouldn’t surprise me if several parts of our job are easier in the coming years.

But will that put us out of job? I doubt it. Being a doctor, particularly an intensivist, requires a diverse combination of talents beyond knowledge: communication skills, procedural skills, and trustworthiness.

LLMs function well at knowledge tasks but may not be so great at gathering a history from medics, family, etc. LLMs are great at following guidelines but not so good at handling tough edge cases. They are great at making rare diagnoses from a perfect case history but fail inexplicably and dangerously at more routine tasks. Furthermore, they are not going to convince IR to come into the hospital at night. And no way is the field of robotics anywhere close enough to able to climb over a bunch of wires to the head of the bed and intubate during a code. Finally trust is foundational in medicine, and I doubt we’re at the point of trusting AI with life or death.

AI does exceptionally well at tasks with fast and verifiable outcomes, which is why they excel at writing code. If the code compiles and runs as expected it was correct; if not try again. Note writing is analogous to coding, and LLMs can undoubtedly produce clinical documentation faster than any human. If code fails a unit test the LLM knows immediately and precisely why. But misdiagnosis is rarely so obvious, and practicing medicine is not simply producing a string of tokens. Determining the correct diagnosis on a standard test is fast and verifiable, but in real life the diagnostic process requires iterative steps in the face of incomplete information. AI does not perform nearly as well in these situations.

The remarkable LLM performance on multiple choice USMLE exams is not seen on more realistic clinical prompts. AI radiologists consistently demonstrate a pattern of higher sensitivity but lower specificity: a tendency to overcall. This suggests that AI would be more suitable for a first pass read than the final interpretation.

The limitations of medical robots are even more profound.

In an example of Moravec’s paradox, robots are surprisingly adroit at “hard” tasks that need to be done precisely the same way each time, like doing a backflip. On the other hand they are surprisingly clumsy at tasks that require interaction with objects. Humans process haptic feedback easily and intuitively, while robots struggle with even simple tasks on non-rigid materials like folding clothes.

Procedures require haptic feedback and intuition about how pliable biological materials will respond. Humans take years or decades to master these skills, after a lifetime of interacting with soft objects in the real world. Robots - lacking embodied cognition - have no such experience. Unlike the vast corpus of medical text to train on, there is no haptic database for training medical robots. I suspect that we will need thousands of tele-presence surgeons (humans operating remotely) to create the millions of hours of training data needed to train truly autonomous robot surgeons. Finally, if we consider how autonomous driving has been gradually deployed, I’d expect many years of telepresence surgeons followed by more years of robots acting as first assist before any robot is acting autonomously.

Dramatic failure of an allegedly autonomous - but likely teleoperated - Optimus robot, December 2025

Finally, consider trust and communication.

One of the first lessons taught in medical school is “never lie.” It is far better to say “I don’t know” than to make up an answer that might be wrong. Cutting edge LLMs hallucinate approximately 2% of the time when asked to summarize data, but far more often when asked complex queries. A recent study found that LLMs hallucinated up to 58% of the time when asked to summarize federal court cases. A 2025 study embedded fabricated clinical details (fake lab value, physical sign, or diagnosis) into 300 physician-validated vignettes and tested six major LLMs. Hallucination rates ranged from 50% to 82%; even after mitigation prompts, the best-performing model (GPT-4o) still elaborated on false clinical content 23% of the time. Even a single digit hallucination rate seems fundamentally incompatible with the trust placed in medical professionals: would you trust a doctor who lied to you a quarter of the time?

Most people consider face-to-face human interactions essential to convey empathy and to react to emotions. Conveying information and listening to concerns is a core competency in medicine.

This is not to say that LLMs cannot convey (simulated) empathy effectively. In text-based exchanges, LLMs may actually do better than humans. But this advantage goes away decisively when we consider our preference for face to face interactions, particularly for breaking bad news. Aside from a tiny cohort of San Fransisco tech bros, I suspect very few people would want to be told they have a brain tumor by a chatbot. The point isn’t that AI cannot seem empathetic in text. It’s that the encounter most central to medicine - delivering difficult news - is not a text exchange. It requires presence, accountability, and a human being who will be there the next day.

Let’s return to history, and examine two more jobs where there’s been significant automation: banks and aviation.

You might have predicted that the invention of the ATM in 1967 would have markedly reduced bank teller jobs. In 1950 there were approximately 150,000 bank tellers. This doubled to about 300,000 in 1970, and doubled again to 600,000 in 2010. Despite the ubiquity of ATMs, there are far more bank workers today than in 1950. They don’t simply hand out money to people. They do more complex tasks: setup accounts, transfer funds, apply for loans, and provide advice. Physicians can probably expect a similar trend towards more time spent on higher level tasks.

Actor Reg Varney demonstrates using the first ATM at Barclays Bank, London June 1967 (Source: Wikimedia Commons)

Similarly, after the first autopilot was invented in 1912 you might expect pilots would be losing jobs en masse a century later. In fact, pilot salaries are at historic highs. Automation has made flying much safer, but we still demand two responsible licensed humans to supervise it. I suspect medicine will mirror aviation: safer, faster, more efficient care delivered by human experts supervising sophisticated automation.

Lawrence Sperry, inventor of the autopilot, climbs out of the cockpit to demonstrate his invention. June 1914 (Source: Reddit)

Let’s reconsider some some of those historical precedents. How many of those jobs have actually gone away? When I asked Claude to analyze the 204 English occupational surnames listed on Wikipedia, it found that just 22% were extinct. 33% still exist, and 45% have evolved. Even after centuries of automation, technology is more likely to gradually change jobs than to replace them. (Anyone who watches The Office knows there are still candle makers.)

Similarly, telephone operator went from a common occupation to a rare one, but the transition was gradual, taking place over decades. The technology for mechanical call switching was invented in the 1880s but wasn’t implemented until the 1920s, and the phasing out of switchboard operators wasn’t complete until around 1980. I suspect if AI eliminates any jobs in healthcare, the transition will be similarly gradual. Furthermore, connecting calls manually may have gone away, but the market for making calls has never been larger. Today over 3.6 million people are employed in call centers in the US, roughly 2.5% of the total US labor force! Will medicine undergo a similar transition? Will we spend less time “connecting calls” and more time talking to people?

Finally, if you are worried about AI taking physician jobs, ask yourself what jobs are truly safe? If AI is taking jobs is medicine that require so much skill, training, and trust, what other “knowledge” professions will go by the wayside first? Lawyers, programmers, accountants, engineers, etc. If a robot professional can do our job, it can surely do almost every other. I’ll worry about AI replacing doctors when those other professions are going extinct.

I am optimistic about how AI will improve medicine, but I am profoundly skeptical that it will lead to major reductions in physician jobs in the coming decades. I think it’s far more likely that AI will put an end to PCPs writing clinic notes at midnight than that AI will put an end to PCPs. I think it’s far more likely that the monotonous aspects of our job will be automated. I suspect more emphasis will be placed on skills like complex reasoning, hand eye coordination, and interpersonal communication, and less on rote memorization. Future doctors will be like pilots today: compensated more and delivering better, safer care with the help of AI.