Where did the AI go to medical school?

I’m going to start with my friend Edward.He’s part of the reason I’m a doctor.I’ve known Edward for 30 years. He runs a printing shop across from where I used to live in medical school.And well, our friend at the time had a spectacular collection of photocopied medical texts and notes that were a godsend to me. It was hard to get the few copies of the books at Gerstein library, and at the time, paying $3K for books a year was tight for me.So Edward essentially saved my medical career with his quasi-legal collection of textbooks. It was reading this pirated library that made me into a doctor.And hey—guess what—a quasi-legal pirated series of texts is EXACTLY how the AIs learned medicine as well.Where did the AI go to medical school?It sounds like a funny question. But where did the the training data for LLMs come from regarding medicine?The Pile is the first fully documented open training corpus for LLMs—800 GB of training data that included PubMed Central and roughly 30 million PubMed abstracts. And Wikipedia. Lots of Wikipedia.Meta’s LLaMA disclosure was mostly web crawl but Wikipedia was passed through training twice while the crawl got one pass. GPT-3: English Wikipedia was about 0.6% of the raw tokens but 3% of the sampling mix, a roughly five-fold overweight. And when the Washington Post audited C4, the “clean web” dataset behind several models (formally audited here), wikipedia.org was the second largest source among 15 million websites, behind only Google’s patent database. The labs stopped disclosing their mixes after 2023, so the clearest recipes we have are the early ones, and the early ones all say the same thing: Wikipedia was deliberately over-served and over-weighted by the AIs.But don’t get me wrong, the AI also read medical textbooks (but, like me, not through the bookstore). Books3 and Library Genesis are “shadow” libraries holding millions of pirated books including a large share of the medical canon, and were the likely source of training—this is now a matter of court record (Bartz v. Anthropic ended in a $1.5 billion settlement, the largest in the history of American copyright). Then after the books—the AIs studied for exams with question banks like MedQA (12,723 licensing-exam questions) and MedMCQA (nearly 194,000 more).And, at the very end, physicians were hired to grade and tune: OpenAI’s HealthBench alone enlisted 262 physicians to write 48,562 rubric criteria. One Nature paper title summarizes the final synthesis: “Large language models encode clinical knowledge.” They do indeed.Notably, one element of what was almost entirely missing: patients. The major open clinical datasets are contractually walled off from consumer AI, so I guess I would view this as the equivalent of systematically excluding patients. A systematic review in JAMA last year found only 5% of healthcare LLM evaluations used real patient data. Run against 2,400 real abdominal-pain cases, models that aced the United States Medical Licensing Examination (USMLE) part 1 ended up underperformed against physicians! The AI learned medicine the way a brilliant student would if they just studied in the library and never when to the hospital to see a patient.Now who exactly were the faculty?Meet the facultyThe “teachers” are the ones who wrote the open content. And I think those great teachers were the people that edited the Wikipedia content that formed the earliest open corpora. And MAN that Wikipedia corpus is huge. When Heilman and West counted (I will get to Heilman in a moment), Wikipedia’s medical content ran to 155,000 articles across more than 255 languages, supported by 950,000 references, viewed 4.88 billion times in 2013 alone! And the core community maintaining all of it: fewer than 300 people, half of them healthcare providers.I think this is the greatest medical education faculty that has ever existed.In February 2004, a Dutch physician newly arrived in Britain began editing Wikipedia as User:Jfdwolff. A project to build medical articles—WikiProject Medicine—literally began on his user page. Today Dr. Jacob de Wolff is a consultant in acute medicine in London, and the encyclopedia’s featured articles on coeliac disease, meningitis, dengue fever and thyrotoxic periodic paralysis, plus core pages on diabetic ketoacidosis, Guillain-Barré syndrome and heparin-induced thrombocytopenia, are substantially his. His editing philosophy is the whole thing in one line: “you cannot build an encyclopedia on poor sources.” His articles are brilliant. I’ve loved reading them.My fellow Canadian Dr. James Heilman approached the project differently. On a slow night shift in Saskatchewan in 2008, he read Wikipedia’s obesity article, found it riddled with errors, and realized he could simply fix it: “I sort of became hooked from there.” By 2016 he was editing 60 hours a week around emergency shifts in Cranbrook, British Columbia. He built the translation task force that pushed core medical articles into dozens of languages, packaged offline medical Wikipedia for low-connectivity settings, and when psychologists filed complaints against his licence for posting the public-domain Rorschach inkblots, he held the line: “This information is encyclopedic.” (The inkblots are still up). My favourite thing about James—he caught commercial textbooks plagiarizing Wikipedia (so say what you want about piracy—the publishers were copying the volunteers before the machines were).And the byline goes on. The pathologist is Michael Bonert, User:Nephron, who has released roughly 2,500 of his own micrographs into Wikimedia Commons; if you have ever pulled up a histology image on Wikipedia, there is a fair chance you were looking through his microscope. When a multimodal model learns what a biopsy looks like, there is a fair chance it did too.Fifteen years ago we all wrote an article together in the Journal of Medical Internet Research called “Wikipedia: a key tool for global public health promotion.” I argue how key the tool is has increased an order of magnitude now.The greatest of medical educatorsWikipedia was punching above its weight even before AIs. Pharmacologists in Munich compared Wikipedia’s drug information against standard textbooks and found 99.7% accuracy. Wikipedia sits in the top 10 Google results for 71-85% of health queries. More than half of physicians admit consulting it when taking care of patients. A McMaster scoping review calls it the most frequently visited health information resource on the internet.Osler’s Principles and Practice of Medicine was the world’s most significant medical textbook for over 40 years. We named half of medicine after him. On the only metric that ultimately defines a teacher, minds reached times understanding transferred, I think Dr. Jacob de Wolff and Dr. James Heilman belong in the conversation with Dr. Osler. They taught everyone on Earth with a browser, for free, and almost no one learned their names.And everything those volunteers wrote got read again, by the AI models, more times than its share of the data would justify: two to five times over in every documented corpus. When a frontier model explains coeliac disease at 2 a.m. to a frightened parent, some real fraction of that fluency is a Dutch registrar’s prose from 2004. When GPT-4 cleared the USMLE passing threshold by 20 points with NO medical fine-tuning, generic pretraining alone sufficed, and the preclinical faculty of that pretraining was, in meaningful part, WikiProject Medicine.So, where did the AI go to medical school?A few places: Pubmed Central.Pirated books.And then Wikipedia, where it was taught by a small number of dedicated health professions volunteers.And then polished themselves with exam banks.And to be honest, that is not unlike medical students!

Where did the AI go to medical school?

Register for free access or login.

Related Topics