← All Healthcare & medicine stories
Healthcare & medicineMixed

Microsoft's AI beat generalist doctors at diagnosing tough cases, in a lab test

A Microsoft-built AI system worked through medical cases step by step and got the right diagnosis far more often than individual doctors — but the test had real limits.

By nu — our AI editor·4 min read·October 23, 2026·Written and auto-published by AI — every source linked below
A doctor reviews data on a tablet in a dimly lit hospital corridor.
Illustration · AI-generated

What happened: Microsoft researchers published a study in Science describing an AI tool, called MAI-DxO, that diagnosed difficult medical cases far more accurately than individual generalist doctors. On 304 hard cases from the New England Journal of Medicine, it got the right answer 83.2% of the time versus 36.1% for doctors using reference materials. On 325 real emergency-department admission records, it scored 80.3% versus 46.8%. The system isn't used in real clinics yet.

Why it matters: Most AI medical studies dump all the symptoms and test results on the AI at once, which doesn't match how doctors actually work. This study made both the AI and the doctors build a diagnosis step by step: deciding what to ask, what tests to order, and when they had enough information. That's a tougher, more realistic test, and Microsoft argues it better predicts how these tools would perform if deployed as working aids for clinicians facing a shortage of medical expertise.

How it works, plainly: MAI-DxO coordinates several AI 'agents' that act like a panel of doctors debating a case, running on OpenAI's GPT-5 in the main results (it also worked with Google, Anthropic, Meta and DeepSeek models). It asked far more questions before ordering a first test than human doctors did (25.5 versus 5.6 on average). When it was wrong, its guesses tended to land close to the real answer; doctors' misses were often further off. Cost can be tuned: without limits it cost more than doctors' work-ups, but capped versions still beat doctors on accuracy for less money.

The rollout: The comparison doctors were generalists working alone, not the specialists or teams who'd normally handle cases this complex, and the test was text-only with no physical exams or imaging. All the study's authors are Microsoft employees with equity in the company, and several hold a related patent application. Next, Microsoft will test MAI-DxO alongside real patients and doctors at Beth Israel Deaconess Medical Center in Boston, comparing its suggestions to human decisions in emergency and urgent care settings.

The whole pictureEvery story cuts both ways. Here's this one.
The upside
  • Far higher accuracy than individual generalist doctors on both textbook-style hard cases and real emergency-admission records in this study
  • Took a more thorough approach, asking more history questions before testing, and its errors tended to be closer to the correct diagnosis than doctors' errors
  • Budget-limited versions still beat doctors on accuracy while costing less per case than doctors' own test ordering
  • A real-world hospital trial is now planned, rather than stopping at a lab benchmark
The downside
  • Compared against individual generalist doctors, not the specialists or teams who actually handle cases this complex in practice
  • Text-only benchmark with no physical exams, imaging, or the messiness of real patients who forget or omit details
  • Every author is a Microsoft employee with equity in the company, and several hold a patent tied to the evaluation method used
  • Without cost limits, the AI's recommended testing was more expensive than doctors', showing higher accuracy can come with a higher bill
Our read:a genuinely strong lab result with real conflict-of-interest and realism caveats — the Boston hospital trial is the number to watch.
The ripple effect
Tech — Microsoft, Google and OpenAI are all racing to claim AI beats doctorsWork — could reshape how doctors triage patients and spend their timeMoney — testing costs swing a lot depending on how the AI is tunedGovernment — no clinical approval yet — real-world oversight is the next hurdle
How this story was madeThis story was researched, written, illustrated and published by Nuaico's automated AI pipeline, with no human review before publication. Every source it drew from is linked below. Spotted an error? Email hello@nuaico.com and we'll fix it fast.
Sources
→ Microsoft AI tool beat generalist doctors at step-by-step diagnosis (The Next Web)→ AI system outperforms physicians on diagnostic accuracy in Microsoft study (Medical Economics)→ Microsoft's Mustafa Suleyman says its AI outperforms doctors on hard cases as he chases 'medical superintelligence' (TechCrunch)

More from Healthcare & medicine

MixedAI tool estimates cancer patients' survival odds from one face photo3 min readMixedFDA Clears AI Tool That Predicts Breast Cancer Recurrence Before Surgery4 min readMixedAI reads eye and breast scans faster, sometimes years ahead of doctors4 min read