Microsoft researchers say they have built an artificial intelligence tool that diagnosed complex medical cases with more than twice the accuracy of generalist physicians. The results were published on Thursday (22) in the journal Science, according to reports by Brazilian outlets G1 and UOL.
What sets this study apart is the format. Instead of answering a multiple choice exam with a tidy patient summary, the AI works through the case step by step, much like a doctor on shift: it decides what to ask about the patient's history, picks which tests to order, reads the results and, if needed, requests new ones. A system called the "orchestrator" coordinates the process the way an attending physician might gather several specialists around a table, updating its working hypotheses as new information comes in.
In the tests, the tool was given more than 300 rare, hard-to-diagnose cases and compared against generalists who were allowed to consult books and the internet. According to the researchers, it also outperformed doctors in simulations built from records in a large database of emergency room visits, and the diagnostic workups it ordered cost less than those of the human physicians.
The criticism of older benchmarks comes from Microsoft itself. "AI got very good at acing 100% of those questions," said Dominic King, corporate vice president of health at Microsoft AI, as quoted by G1. "But of course that doesn't match the real world. No patient has ever walked into my office with a neatly organized summary." Harsha Nori, head of AI at Microsoft AI Health, said the study also measures "the care and judgment" with which the system runs an investigation. The next step is a trial in which the tool will triage patients in emergency room waiting areas, in partnership with Beth Israel Deaconess Medical Center in Boston.
What we still don't know
Before retiring the stethoscope, some caveats. The 300 cases came from written records rather than real patients in an exam room, and the comparison doctors worked alone, with no colleagues to consult, which is exactly the kind of teamwork the "orchestrator" simulates. The study was also run by Microsoft itself, which graded its own product, and the cost savings come from simulation rather than a working hospital. Independent validation and any regulatory approval for use on patients are still missing.
The incentives matter too: Microsoft is racing Google and OpenAI for the healthcare AI market, where it sells cloud computing and tools like Copilot to hospitals, and a paper in Science doubles as a shop window. The practical takeaway: watch the Boston triage trial. If the tool performs well with real patients, the story starts to affect your next doctor's visit. Until then, treat the "twice as accurate" figure for what it is: a promising lab result that still has to prove itself in the field.