Microsoft’s AI team taught a machine to practice medicine the slow way, and the results are hard to ignore.

In a study published Thursday in the journal Science, Microsoft researchers reported that their MAI-DxO diagnostic system correctly diagnosed 304 fiendishly complex medical cases 83% of the time. General physicians given the same cases managed 36% with textbooks and the internet, and just 22% without references. That’s better than a two-to-one edge for the machine.

This isn’t another model acing a multiple-choice exam. The whole point of the work is that the AI does what doctors actually do: ask questions, order tests, reconsider, and narrow things down.

How MAI-DxO works

Most medical AI takes the all-at-once approach. You feed it every symptom, lab result, and scan at once, and it spits out a diagnosis. That scores well on benchmarks but looks nothing like a clinic visit.

MAI-DxO flips that around. It starts with only the patient’s initial symptoms, then an “orchestrator” coordinates a simulated panel of clinical experts that debate the case. The AI first decides what to ask about the patient’s history and what matters in the physical exam. Based on what it learns, it picks which tests to order, studies the results, and updates its working hypotheses as it goes, the researchers told Reuters.

In other words, it replicates the sequential judgment a doctor uses when information arrives piece by piece — including the awkward fact that the first test ordered is often the wrong one.

Microsoft’s team also built a benchmark called SD Bench to evaluate this kind of step-by-step diagnosis, rather than the single-shot multiple-choice format most medical AI tests use.

What the medical diagnosis study actually found

The headline results come from 304 high-difficulty clinical cases originally published in the New England Journal of Medicine, each one the kind of diagnostic puzzle doctors write up precisely because it’s unusual.

MAI-DxO diagnosed 83.2% correctly. Physicians working without references hit 22.3%; with textbooks and the internet, 36.1%.

On a second evaluation using a major database of US emergency department admissions, the AI scored 80% against physicians’ 47%. And the AI did it while estimating diagnostic costs at one-half to one-third of what current systems rack up, in part by ordering fewer unnecessary tests.

Dominic King, the corporate vice president of Microsoft AI’s health unit, told Reuters that the multiple-choice tests the industry has leaned on don’t capture real practice. “AI has become very good at getting a 100% pass rate on those,” he said. “But of course, that’s not real-world. No patient ever came to see me with a nice summary.”

The fine print

Microsoft’s own framing deserves a read before anyone declares doctors obsolete.

TechCrunch notes that the comparison physicians were primary care doctors and general internists working alone — while cases this tricky would normally land in front of specialists or whole teams, who weren’t part of the study. The study’s data also suggests much of the improvement comes from more powerful base models, with Microsoft’s orchestration adding gains on top: the team’s best results came from running DxO on OpenAI’s GPT-5 and GPT-5.5, though it also tested models from Google, Anthropic, Meta, and DeepSeek.

That’s an important detail. It means MAI-DxO is partly a wrapper story — clever orchestration riding on someone else’s frontier model — and its future performance depends on foundation-model progress Microsoft doesn’t fully control.

Mustafa Suleyman, Microsoft AI’s CEO, called the results a step toward “medical superintelligence.” The phrase will get the headlines, but the engineering underneath is the more interesting story: a system that knows which question to ask next is a bigger deal than a system that knows every answer in advance.

What happens next

Microsoft says a large Boston hospital has started a verification study to test the tool with real patients, per Maeil Business News. Everything so far is simulation, and the leap from NEJM case files to a chaotic ER is where medical AI has stumbled before.

Expect Google and OpenAI to answer soon. TechCrunch points out that all three labs are now claiming their AI can match or outperform doctors, and Microsoft has made its bid the most detailed one yet by putting it through peer review in Science — the study first surfaced in June 2025 and has since been formally verified by outside experts.

The bigger question is clinical, not technical: whether hospitals will trust a panel of arguing AIs enough to change how diagnosis gets done.

Why this matters

Medicine has a supply problem. There aren’t enough specialists, they’re expensive, and the hardest cases are the ones most likely to get misdiagnosed. An AI that does the unglamorous middle of diagnosis — the questioning, the test-ordering, the revising — could genuinely change who gets good care and at what cost, if it ever makes it past the hospital pilot stage.

It also marks a shift in how AI labs pitch their work. The field spent years bragging about models that get perfect scores on exams doctors never take. Microsoft just made its case on a test that looks like the job itself. That is a harder test to pass and a more honest one — and it’s probably the standard every lab gets judged on from here.

FAQ

What is Microsoft’s MAI-DxO? It’s an AI diagnostic system from Microsoft AI’s health team that works through cases step by step, like a doctor: asking history questions, ordering tests, updating hypotheses, and reaching a diagnosis through a simulated panel of clinical experts. The peer-reviewed results were published in Science on October 22, 2026.

How accurate is Microsoft’s AI at medical diagnosis? On 304 highly difficult cases from the New England Journal of Medicine, MAI-DxO was correct 83.2% of the time. General practitioners scored 36.1% with textbooks and internet access, and 22.3% without references. On US emergency department admission data, it scored 80% versus physicians’ 47%.

Does this mean AI can replace doctors? No. The study compared the AI with generalists and internists working solo, while these complex cases are normally handled by specialists or physician teams who weren’t tested. Microsoft positions the tool as support for clinical judgment, and it hasn’t been used on real patients yet.

When will this AI diagnosis tool be used in hospitals? There’s no deployment date. A large Boston hospital has reportedly begun a verification study with real patients, but all published results so far come from simulations.

Sources: Reuters, TechCrunch, The Straits Times, ChosunBiz, Maeil Business News.