A Microsoft AI diagnostic tool described in a study published in Science was more than twice as accurate as general physicians at working through more than 300 unusual and confusing medical cases, according to a wire-service report on the research published on 22 and 23 October 2026. The system treats diagnosis as a sequence of decisions, asking questions and ordering tests step by step, rather than producing an answer from a fixed set of symptoms and results.
Key takeaways on the Microsoft AI diagnostic tool
- The study, published in Science, tested the system on more than 300 highly unusual medical cases, according to the report.
- The report says it was more than twice as accurate as general physicians, who used textbooks and the internet, at working up patients and reaching a diagnosis.
- It also outperformed physicians on scenarios drawn from a large database of emergency-room admissions, but the report gives no figures for that result.
- Microsoft is working with Beth Israel Deaconess Medical Center in Boston to test the Microsoft AI diagnostic tool for triaging patients in emergency department waiting rooms.
- None of this shows the system is ready to replace doctors or that it performs equally well in routine clinical practice.
What the Microsoft AI diagnostic tool does differently
The report describes an orchestrator that directs the diagnostic process and uses a simulated panel of clinical experts working together on a case. It decides what to ask about the patient’s history, which examination findings matter and which tests to order. It then reviews the results, orders further tests if needed and keeps updating its working hypotheses and plan.
That differs from most existing medical AI, which the report says typically examines all symptoms, laboratory results and imaging together to produce one diagnosis. Dominic King, corporate vice president of the health unit at Microsoft AI, is quoted saying that AI has become very good at getting a 100% pass rate on multiple-choice exams, but that no patient ever came to see him with a nice summary.
What the study measured
According to the report, the system was tested on more than 300 highly unusual cases and compared with general physicians who could use textbooks and the internet. The only comparison given is that the system was more than twice as accurate. The report does not say how many physicians took part, what accuracy percentages were recorded or how correctness was defined, so we cannot put numbers on either side.
The report also says the system did better than physicians on emergency-room scenarios, and that diagnostic work-ups cost less when the orchestrator designed them than when human doctors did. Harsha Nori, head of AI at Microsoft AI Health, is quoted saying the team is also measuring how judiciously and carefully the system goes through a work-up. No cost figures are given.
General background: from static questions to sequential reasoning
Earlier evaluations of medical AI often relied on multiple-choice questions or fixed case summaries, which do not capture how clinicians gather information over time. In 2025 Microsoft Research published a preprint on “sequential diagnosis” with a system called MAI-DxO, which is available on the Microsoft Research site. The wire report on the Science study does not name the tool, so we have not confirmed that it is the same system or that the figures match.
Our analysis: promising, but a long way from the bedside
In our assessment, the Microsoft AI diagnostic tool is a meaningful step because it tests a more realistic task than a quiz. Unusual, complex cases are exactly where a broad, tireless differential can help. But they are not typical of everyday care, and a strong result on difficult cases does not predict performance on common presentations, where a human’s speed and judgement are hard to beat.
The planned waiting-room evaluation at Beth Israel Deaconess Medical Center is the more important test, because triage involves real patients, time pressure and safety stakes. We treat the current results as research findings, not evidence of clinical benefit. See also our healthcare IT coverage.
What is not disclosed about the Microsoft AI diagnostic tool
- Exact accuracy percentages for the system and the physicians.
- How many physicians and cases were compared, and how correctness was judged.
- Figures for the emergency-room scenarios and for the cost comparison.
- The underlying language models and whether the tool has a product name.
- The design, timetable and endpoints of the waiting-room evaluation.
Risks and caveats for the Microsoft AI diagnostic tool
Results come from retrospective or simulated cases, not live patients, and the report gives no safety data. The comparison group used textbooks and the internet, which may differ from how clinicians actually work. A tool that performs well in a study can still fail in settings with missing data, unusual patients or different workflows. Software that supports clinical decisions can also fall under regulatory oversight; the US FDA explains its approach to clinical decision support software. This is news reporting and not medical advice.
What to watch next
- Publication details of the Microsoft AI diagnostic tool, including accuracy tables, in the Science paper.
- Design and results of the Beth Israel Deaconess emergency-department triage evaluation.
- Independent replication on cases the developers did not select.
- Any regulatory pathway Microsoft describes for clinical use.
FAQ
What is the Microsoft AI diagnostic tool?
It is an AI system from Microsoft AI that diagnoses step by step, asking questions and ordering tests with a simulated panel of experts, according to the report.
Can the Microsoft AI diagnostic tool replace doctors?
No such claim is supported. The results are from difficult test cases, and the planned emergency-department evaluation is a next step.
How accurate was it?
The report says it was more than twice as accurate as general physicians on more than 300 unusual cases, without giving exact percentages.
How we reported this: we relied on a wire-service report on the Science study and on Microsoft Research’s earlier public material, attributing each claim. We have not read the Science paper itself (Science). Last updated 23 October 2026. This is news reporting and not advice.