Microsoft MAI-DxO beats doctors at diagnosis, study finds

Generalist physicians given the same cases got 36%.That's the headline from a 2026 study by Microsoft AI researchers, peer-reviewed and published in Science, on the Microsoft AI Diagnostic Orchestrator, or MAI-DxO.

Before you read it as "AI beats doctors," look at how the test was built.The team ran the model on 304 clinicopathological conference cases from the New England Journal of Medicine, plus 325 emergency-department-to-admission encounters from a database called MIMIC-IV-ED.

In the NEJM cases, physicians could consult reference materials.In the emergency cases, MAI-DxO scored 80.3 percent against 46.8% for doctors.

That's roughly 1.7 times as accurate, not double.Earlier studies fed the AI a complete case in one prompt.

Here, MAI-DxO started from a patient's initial complaint, formed a working diagnosis, chose follow-up questions and ordered tests while weighing cost.The authors say previous work hadn't tested diagnosis as an active, sequential process that weighs accuracy against cost.

Microsoft says the model reached its answers at half to a third of the doctors' testing costs.The "patient" in this study was an AI.

A gatekeeper agent answered the model's questions from the underlying case record.

Human doctors had already done the legwork of assembling those detailed files.The tests ran in a controlled benchmark, so their usefulness in a real hospital is up for debate.

And medicine involves more than analysing symptoms.

Some of the hardest work is getting patients to open up about things they didn't think were worth mentioning.Microsoft AI CEO Mustafa Suleyman is bullish.

He told The Verge that "extremely impressive deployments" of such systems could arrive in production next year.

Microsoft.