Microsoft's AI Beat Doctors 83% to 36% on the Hardest Cases
Microsoft's AI Beat Doctors 83% to 36% on the Hardest Cases By Daniel Rao · October 23, 2026 "No patient ever came to see me with a nice summary." That's Dr. Dominic King, a trained physician who now runs health work at Microsoft AI, explaining why the medical-AI benchmark that matters most was never the exam. Medical AI has spent years collecting 100% pass rates on multiple-choice tests where every fact is handed over upfront — neat patient histories, tidy lab panels, one correct answer waiting at the bottom. Real medicine never looks like that. A patient walks in with a story that doesn't add up, and the doctor has to decide what to ask, what to order, and when to change their mind. So Microsoft built a test that works like that instead. And then its AI played doctor — step by step, hypothesis by hypothesis — and got the hardest cases right 83% of the time. The actual doctors it was tested against: 36%. The study landed Thursday in the journal Science , ...