Spoke at AI Engineer World's Fair about what AI in healthcare is getting wrong, and why the usual fix doesn't work. Roughly 1 in 20 AI-generated clinical notes carry an error rated serious or worse if left uncorrected, and the ones that matter most are where something is missing.
The thesis in short:
- This isn't a benchmark. It's production, right now, and these notes go into the record and drive what happens to the patient next.
- Omissions are not the failure we found most of. They are the one that matters most: 68% of the omissions we verified were rated high importance, against 40% of the wrong values.
- They're also the hardest to catch, because knowing whether a missing line mattered is expert judgement rather than proofreading. Nothing on the face of a clean-looking note tells you what isn't in it.
- Which means the judges fail on exactly the wrong half. The best conventional one we built scored 0.94 at spotting invented content and 0.63 at spotting missing content, where 0.5 is a coin toss. Not because it can't read, but because nobody ever wrote down which differences matter.
- So the fix isn't a better rubric. It's capturing that standard instead of specifying it: discover the failure modes from real outputs, capture expert judgements on them, retrieve the relevant ones for each new note. And keep doing it, because it goes stale.
The chart is our census: 618 verified failures across 565 notes from three production scribes. Just under a third of notes carried at least one.
https://t.co/cb93O0QSrm
We audited three deployed AI scribes over 565 clinical notes and found one note in three carried a verified error. Then we turned the same scrutiny on the thing that produced that number.
Take the identical candidate errors, the identical model, the identical evidence, every setting unchanged. Alter one thing: the sentence telling the reviewer how strictly to judge. Told to refute whatever it defensibly can, it verifies 9.3% of them. Told to review leniently, 79.0%. Nothing else moved.
That is also why the published audits of AI scribes disagree with each other. The share of errors they attribute to omission runs from 54% to 86%, against 23.1% in ours, and that spread sits comfortably inside what instrument differences alone can produce.
So an accuracy figure here is a joint measurement of the product and of whoever wrote the review prompt, and almost nobody publishes the second half. We released all 618 findings with their evidence, every prompt and model version, and the pipeline, so you can move the dial yourself and watch the number follow.
https://t.co/OuyLsufGBx
I have 2 obsessions at the moment - wine & RL data/envs.
I thoughts these were quite separate pursuits and couldn't be merged, but it looks like I'm wrong 😂
https://t.co/zGVDtKCC3J
Have a look at your spice rack. In medieval price lists a pound of ginger cost roughly what a sheep cost and a pound of saffron roughly what a horse cost, and nobody in Europe had any real idea where the stuff came from. Jean de Joinville, writing around 1309, passed on what the merchants believed: "It is said that these things come from the earthly paradise; for the wind blows down the trees in paradise, just as the wind blows down the dry wood in the forests of our own land."
Cinnamon was seriously thought to be driftwood that had blown out of Eden.
Glasses are such an unremarkable object that you can buy a pair at a petrol station. But before they existed, your eyes going meant the end of reading and close work for the rest of your life, and eyes go for nearly everyone. A friar preaching in Florence in 1306, about twenty years after they were invented, said this: "It is not twenty years since there was discovered the art of making spectacles ... one of the best and most necessary in the world. I myself saw the man who discovered it, and I talked with him."
He had met the man. Imagine getting to shake hands with whoever invented being able to read in old age.
We have light in every room whenever we want it and it costs almost nothing. For all of history before about 1880, every light in your house was a small open fire that you had to light, that smoked, and that gave a fairly feeble light. Here is a witness to the night Wabash, Indiana became one of the first towns lit entirely by electricity, in 1880: "The people, almost with bated breath, stood overwhelmed with awe, as if in the presence of the supernatural. The strange, weird light, exceeded in power only by the sun, yet mild as moonlight, rendered the Court House square as light as midday."
Everyday things around us are actually incredible if you stop and think about it. Being able to play any song at home on demand, for example. Before recordings, hearing music meant having an instrument in the house or a person nearby who could play, and a song you loved might go unheard for years at a time. This is what Edward Bellamy wrote in 1888, imagining the best possible future: "If we could have devised an arrangement for providing everybody with music in their homes, perfect in quality, unlimited in quantity, suited to every mood, and beginning and ceasing at will, we should have considered the limit of human felicity already attained."
He is describing Spotify, and he thought it was the outer limit of human happiness.
A few years ago, I was at a lab training models and frustrated with data quality. Now I run a data research lab where former colleagues are buyers. It has been a humbling, fascinating and incredibly rewarding (no pun intended) adventure.
Some of what I’ve learned
Being meat-free is a great choice
(caveat - i these aren't all necessarily great examples of people and obviously there are more people who eat meat at the top than not, this is primarily for meme-value, but i do think being vegetarian is a great life choice! 😀)
Surprise 1: Anthropic acquired a one-person biotech
.
.
.
Surprise 2: that person is me
Surprise 3: I’m here to tell you all about it
Yes, this is what you think it is; after 6 yrs of increasingly loose anonymity: the Professor Oak reveal post
…and yes, it has gotten so boring around here that I’m literally doxxing myself to spice up the timeline
-------
Ok, allow me to introduce myself. My name is Cody Tranbarger. I mostly tweet, but occasionally find time to dabble in drug dev & starting/building biotechs. My past lives span the VC, crossover, startup continuum; Prof Oak has been my way of keeping a digital foot in those worlds, despite having my physical feet planted elsewhere.
The past 18 mo, ‘elsewhere’ was Mithril Medicine, a co. I founded to develop bsAbs for LAMA2-CMD & other ECM-related genetic diseases.
Mithril was born out of hands-on building with @MikeNGladstone, @JasonPaulRhodes, @LifeSciVC, & team @atlasventure. Somehow, I made enough half-decent calls along the way (even a broken clock is right twice a day, eh?) to convince them Mithril/I warranted a small seed. While I can hardly rationalize such an outrageously cavalier bet, I’m grateful for it, & I'd partner with Atlas again in a heartbeat.
In typical AV fashion, Mithril’s primordial form was virtual, decentralized, & ultra-lean. This stage tends to be short-lived, but for Mithril, it worked surprisingly well, for surprisingly long, thanks in part to the rapid maturation of the AI tools I used liberally.
Eventually, Mithril became a bit of an experiment itself - to find the upper bound of how far it could scale as a single operator + a bunch of Claude (+ a stellar bench of part-time consultants).
In parallel, I started spending time with @nc_frey, @an1lam, & @CoefficientBio team; and promptly experienced, for the 1st time, AI at the frontier. Soon thereafter, Mithril stopped looking like an experiment, & started looking like a preview of an entirely new operating model.
From that spark came “Pipeline-in-a-Person.” The idea immediately resonated when we shared it here back in April; and it was central to the shared vision that led to Anthropic’s acquisition of Mithril, & my joining the team.
Why join Anthropic? Three reasons.
(1) For Mithril, the programs continue, with more resources & infrastructure behind them, and as much Claude as I can put to work.
(2) For me, the chance to start several more programs, & find out how many one person can actually run in parallel - an almost unprecedented luxury for a curious drug hunter.
(3) For our mission, a path to institutionalize & scale this new way of working, then put it in the hands of all of biopharma, all over the world - which, candidly, I find the most energizing.
Will AI x Bio cure diseases? Unequivocally yes, imo… but not by the route that gets the most airtime. Rather than timeline acceleration & compression, I see orchestration & parallelization as today's low-hanging fruit.
Most drug hunters see a handful of go/no-go calls in a career, & spend entirely too much of the rest trading emails (my inbox can attest).
Give them the bandwidth to run several programs at once, and the time to focus on scientific judgment, and I’m willing to bet they’ll make better calls. And at scale, better calls translate to more drugs, higher hit rates, broader patient impact, & likely, a disease or 10 or 100 cured. I couldn't be more excited to build toward that future @AnthropicAI.
Finally, a heartfelt thank you to BioTwitter/X. The end of anonymity is bittersweet, but if anyone’s earned a proper introduction, you all have.
It’s difficult to overstate how much this app has done for me: a decade of education, plenty of friends, many job opportunities, several mentors, a few investors, even a co-founder (fmr). It was the generosity of this community that inspired me to start posting 6 years ago, after just as many years lurking. I hope I’ve been able to pay it forward, in my own small way.
But, to be clear, Prof Oak isn’t going anywhere. Same beat, same themes, and now, a front-row seat to drug development at the frontier. Stay tuned for dispatches from the trenches; and remember...
Gotta Cure 'Em All
It's a bit mad that for a couple of hundred pounds you can be on another continent tomorrow morning. Crossing an ocean used to take a week or more, and emigrating across one usually meant never seeing your family again. Wilbur Wright, at a dinner in Paris in 1908: "I sometimes think that the desire to fly after the fashion of birds is an ideal handed down to us by our ancestors who, in their grueling travels across trackless lands in prehistoric times, looked enviously on the birds soaring freely through space, at full speed, above all obstacles, on the infinite highway of the air."
Before sewing machines, most people owned two or three outfits in total, because a single new shirt was somewhere around fourteen hours of stitching by hand. Godey's Lady's Book, the leading women's magazine of the day, wrote this in 1860: "What philanthropy failed to accomplish, what religion, poetry, eloquence, and reason had sought in vain, has been produced by - the Sewing Machine."
Having a wardrobe rather than an outfit is a very recent thing.
Washing clothes used to consume an ENTIRE day of every single week: hauling and heating water by the barrel, then scrubbing, rinsing and wringing out every item by hand. The first electric washing machine went on sale in 1908 and gave households back a whole day a week!
I've been reading up on how three big US health systems govern clinical AI: Mass General Brigham, Kaiser Permanente and Ochsner Health. A few quite different models are emerging, and the differences look structural rather than stylistic. Who employs the doctors seems to matter most.
MGB published its governance as a peer-reviewed paper (Saenz et al, NPJ Digital Medicine, Nov 2024, 16 named collaborators), which looks like academic habit until you notice its faculty sit inside academic incentives and the committee can't simply mandate use. Co-authorship is the enforcement. Kaiser employs physicians directly through the Permanente medical groups, and its enterprise AI lead has had system-wide oversight since Oct 2023, so when governance there says a tool gets used a certain way that's an employment instruction, and almost no other US system can mandate like that. Ochsner sits between the two with a mix of employed and affiliated clinicians, unable to order compliance like Kaiser or publish its way to buy-in like MGB, and its answer is a Governance Committee, a Centre of Excellence and an Implementation Tier, with risk-tiering as the core mechanism.
All three deployed ambient documentation at strikingly different speeds. MGB went from 18 physicians in 2023 to 4,000+ providers by 2025. Kaiser's Northern California group alone covers 7,260 physicians and 2.58M encounters in about 14 months. Ochsner went enterprise-wide at 75% adoption, vendor-reported.
What strikes me about those numbers is that they're usage signals, not safety signals. None of the three has published a safety endpoint for the rollout, meaning note error rates or omission rates. Volume tells you the tools are used; it doesn't tell you the governance catches what they miss.
And these three are the well-resourced end. Black Book surveyed 182 hospital leaders in Dec 2025: 29% had AI policies covering model inventory and sign-off, and 22% could produce a full audit trail within 30 days. The Coalition for Health AI promised assurance labs, and none materialised.
If I ran governance at one of the thousands of systems that aren't these three, I don't think I'd ask whose model to copy. I'd ask which structural reality - employment, data ownership, publication incentive - is closest to mine, and which mechanisms simply won't transfer.
What I can't figure out is who catches the silent omission under any of these models. Committees decide which tools get in. Clinicians catch errors in the note in front of them. But the question that never got asked, the detail that never made the note - which layer sees that?
Until the 1840s, surgery was done on you while you were awake. The tools available for the pain were speed, alcohol and opium, and that was the whole list. The man who first publicly demonstrated ether has this on his monument: "Before whom, in all time, surgery was agony; by whom pain in surgery was averted and annulled; since whom, science has controlled pain."
Three clauses, and the entire history of human experience divides at the second one.
I went down a rabbit hole on what malpractice insurers actually think about clinical AI. The picture is more interesting than the scare stats suggest - less courtroom drama, more quiet changes to renewal paperwork.
The most-cited number doesn't hold up. The claim that AI malpractice suits rose 14% from 2022 to 2024 appears only in law firm marketing posts, none of which cite a source, and I couldn't find an insurer or research body behind it.
What insurers themselves report is calmer. The Doctors Company's COO told Medical Economics in October 2025 that most insurers are seeing only limited AI-specific claims, and that there's no documented US malpractice case yet where AI was central to the claim. Coverage is shifting, but cautiously: The Doctors Company says it has no AI exclusion in its base policy and would still defend a physician if AI played a role, while riders are appearing for heavy AI users, often limiting cover to FDA-approved uses and excluding experimental features. Europe is a little further along, where underwriters have started using sublimits, higher deductibles and conditions precedent when tools are unvalidated or used outside cleared indications, probably a preview of where US policies head next.
The vendor contract side is worth watching too. Legal commentary describes a consistent pattern where vendors disclaim outputs tied to customer inputs, cap liability well below a serious outcome, and exclude fine-tunes, so the hospital holds the duty of care and buys the policy.
Why carriers care despite so few claims: verdicts above \$10M more than doubled from 2015 to 2023, and average nuclear-verdict awards rose from \$23M to \$40M (CM&F Group), with a 2025 Utah birth-injury case hitting \$951M. None involved AI, but it's the backdrop for the first case that does.
If you're a CAIO or general counsel, one cheap piece of homework: ask your broker in writing whether your malpractice policy has AI-specific endorsements, sublimits or exclusions, and what triggers them. Most health systems haven't asked yet, and it's a five-minute email.
There is a machine in your kitchen that keeps a box of air cold indefinitely for a few pence a day. Before it, meat and dairy spoiled within a day and most food simply could not be had out of season or kept. When the first cargo of New England pond ice reached Calcutta in 1837, the local paper wrote: "The first transport of ice from the shores of the United States to the banks of the Ganges is an event of no mean importance ... the names of those who planned and have successfully carried through their adventure at their own cost, deserve to be handed down to posterity with the name of other benefactors of mankind."