28 May 2026 · LinkedIn
In 1959, two researchers named Robert Ledley and Lee Lusted published a paper in Science arguing that medicine was a problem of reasoning under uncertainty, and that this reasoning could in principle be automated.
For more than half a century, that argument sat as an aspiration rather than a result. Bayesian systems, rules engines, symptom checkers - each generation came closer, none arrived. The New England Journal of Medicine's clinicopathological conference cases became the goalpost: the hardest reasoning tasks in the profession, the ones that test whether a doctor can think.
Last week, Science published the paper showing the goalpost has been cleared.
Peter Brodeur and colleagues evaluated OpenAI's o1 model against hundreds of physicians across diagnosis, differential generation, and management decisions. The model outperformed the human baselines. It outperformed every previous generation of AI clinical support. The work appeared in Science - the venue with arguably the highest evidentiary bar in the world for a claim of this kind.
The careful reading matters. The tasks were structured and text-based; the comparisons were blinded; nobody is claiming an LLM is yet safe to deploy unsupervised in an emergency room. Several reviewers have pointed out, correctly, that average accuracy says little about how errors distribute across elderly patients, non-English speakers, or atypical presentations.
But the headline survives the caveats. On the core cognitive work of the profession - taking a presentation, generating a differential, choosing a management pathway - the machine is now better than the median doctor on the cases the profession has long held up as the test.
Medicine joins coding as a domain where the capability question is settled. What remains is distribution, liability, and whether health systems restructure around what is now possible or pretend it isn't.
The training pipeline for a physician in the West runs roughly a decade. The capability that pipeline is supposed to produce has just been matched by a system that did not exist three years ago.
That asymmetry hasn't seriously been reckoned with yet.
Join the conversation on LinkedIn →
← All writing