The best model found 78% of the facts in a patient's file.
Stanford researchers put nine models through 508 questions about real patient records, checked by clinicians. The best run recalled 78% of the needed facts and the weakest reached 43%. Scores dropped further whenever an answer had to be pieced together from several visits.
Why it mattersHospitals are wiring these assistants into the record systems doctors read before treating you. A line missed in a long file is not a typo but a decision made on half the facts.
Read in the app — free, in 9 languages
Related stories
A search engine for doctors raised $250 million to make drugs.2026-09-26Claude pushed a physics calculation past the record a human set.
2026-09-26AI tuned a virus to infect bacteria a million times more effectively.
2026-09-25Insurers say hospital AI added $942 million to their bills.
2026-09-25AI read miners' X-rays better than the certified specialists.
2026-09-25