Microsoft’s diagnostic AI, MAI-DxO, scored about 83% on tough medical cases that stumped doctors, who managed 36%.
Story Snapshot
- Microsoft tested MAI-DxO on 304 complex New England Journal of Medicine cases; it hit about 83% accuracy.
- Generalist physicians using reference materials scored about 36% on the same cases, a big gap.
- The results come from a study that Microsoft says was published in Science and formalizes a new benchmark.
- Real-world care remains the prize; prior reviews show mixed AI gains outside curated tests.
What Microsoft’s Study Actually Measured
Microsoft built a step-by-step test called Sequential Diagnosis Benchmark to mimic a clinic visit. The benchmark turns 304 New England Journal of Medicine clinicopathological cases into interactive files. A diagnostician asks questions, orders tests with costs, and commits to a final answer. Microsoft reports that its system, MAI-DxO, got the correct diagnosis in about 83% of these complex cases. Generalist physicians given reference materials landed close to 36% on the same set.
The company frames this as a strict, head-to-head trial of process, not trivia. The system had to reason across multiple steps. It needed to decide which facts matter, which lab to order, and when to stop. That structure reflects the grind of real diagnostic work, where ordering the wrong test can send a case down a rabbit hole. The big result is the margin: MAI-DxO outperformed one-to-one, on the same problems, using the same information rules.
The Numbers Behind 83 Versus 36
Microsoft and several summaries place MAI-DxO near 83% across the 304 complex cases. Reports cite physicians at about 36% with access to textbooks and online sources, and lower without those tools. The gap’s scale matters more than a single point estimate. Doubling a doctor’s accuracy would be big; a gap this wide on hard cases suggests a tool that could change front-line triage, consult timing, and test ordering, if results carry over to real care.
Microsoft has introduced MAI-DxO, a diagnostic AI that outperformed human physicians in complex medical simulations. Utilizing a chain-of-debate reasoning method, the tool correctly diagnosed 83.2 percent of cases in a recent study published in Science. Azat TV pic.twitter.com/OYSA4bmH4T
— Azat TV (@azattelevision) November 2, 2026
These cases are famous for being tough. The New England Journal of Medicine clinicopathological records often feature rare disease twists, misleading symptoms, and confounders. The benchmark pushes the system to work like a resident on call. It rewards correct early hits but also tracks costs and evidence. That design encourages discipline. It also means the system did not just guess the final label; it had to build a path to it and stop at the right time.
Where This Fits In Medical AI Progress
Medical artificial intelligence has shone on curated tests before, then hit turbulence in real clinics. A 2025 review of 83 studies found average diagnostic accuracy of about 52% for generative systems and no overall edge over physicians, with expert doctors still ahead in many tasks. That history makes an 83-to-36 split notable. It hints that newer systems that plan, check, and reason step by step may be crossing a quality line that older models could not reach on complex differential diagnosis.
Hospitals move when tools prove safe, reliable, and useful at the bedside. Microsoft’s team says the Science paper formalizes the benchmark, and coverage points to early hospital studies next. The smart bet is a phased role: decision support first, second opinions next, and autonomy last. That path matches common sense and conservative values—pilot, measure, improve, and only then scale. Patients want speed and accuracy; clinicians want control and accountability. The playbook should deliver both.
What Sensible Adoption Looks Like
Health systems should test MAI-DxO in controlled pilots with clear guardrails. Start in settings with diagnostic bottlenecks, like night coverage or rural clinics. Track misses, near-misses, costs, and time to correct diagnosis. Keep a human in charge of final calls and consent. Publish the results. If the 83% edge over 36% shows up in live charts, not just journals, hospitals will not need a sales pitch. They will see fewer blind alleys and faster right answers where it counts.
Sources:
foxnews.com, microsoft.ai, wheresthe.ai, aiweekly.co, arxiv.org
