Glossary
October 22, 2026

Microsoft's AI Solves Difficult Written Medical Cases More Than Twice as Often as Family Doctors, in a Study Published in the Journal Science

Microsoft says its AI program worked out the right diagnosis in 83% of very hard medical cases, compared with 36% for family doctors and hospital generalists given the same cases, and the study now appears in Science, one of the world's leading science journals. The cases were written records, not real patients, and the doctors were not specialists. A test with real patients in a Boston hospital is planned.

Microsoft published a study today in the journal Science testing whether an AI program can do one of medicine's hardest jobs: figuring out what is wrong with a patient when the answer isn't obvious.

The test was built to feel like a real doctor's visit rather than a multiple-choice exam. The AI, or a doctor, starts with only a short description of a patient. From there they have to ask questions and order tests one at a time, and each answer is revealed only when they ask for it. Every test has a price, so the study tracks both whether the final diagnosis is right and what it would have cost to get there.

Microsoft's program, called MAI-DxO, works like a panel of doctors talking a case through. Several copies of an AI program take different roles. One suggests possible diagnoses, another picks which test to order next, and another questions the group's reasoning. The program can run on top of AI programs from different companies. The best results came from ones made by OpenAI, the maker of ChatGPT.

On 304 especially difficult cases taken from the New England Journal of Medicine, where doctors publish puzzling cases as teaching exercises, Microsoft says its program reached the right diagnosis 83% of the time. Family doctors and hospital generalists working on the same cases got 36% right when they could use books and the internet, and 22% without them. On 325 more ordinary cases from emergency-room records, the program was right 80% of the time and the doctors 47%. Microsoft says its program's choice of tests also cost less.

The limits matter as much as the numbers. These were written case files, not people. Neither the AI nor the doctors could examine anyone or look at scans. The doctors were generalists, while cases this hard usually go to specialists or teams, who weren't tested. The costs are estimates, not real medical bills. TechCrunch, reviewing the results, points out that much of the improvement seems to come from the newer AI programs underneath rather than from Microsoft's panel design. On the emergency-room cases, adding the panel did little beyond what the underlying program managed alone.

Microsoft's own announcement says that "the true test is performance in real clinical environments." It plans a study at Beth Israel Deaconess Medical Center in Boston in which real patients will use the program and its suggestions will be compared with their doctors' judgments. It hasn't said when.

Sources

  1. Microsoft AI — Building medical superintelligence (announcement of the Science study)microsoft.ai · primary
  2. arXiv — Sequential Diagnosis with Language Models (the team's earlier preprint of the study)arxiv.org
  3. TechCrunch — Microsoft's Mustafa Suleyman says its AI outperforms doctors on hard casestechcrunch.com
  4. The Next Web — Microsoft AI tool beat generalist doctors at step-by-step diagnosisthenextweb.com
Follow the timeline
Get an email when new entries are added.
© 2026 Sugarpine