← Latest AI News

October 23, 2026 · Daily brief

This diagnostic tool beat doctors 83% to 36% on tough cases

Sovereignty angle
A tool that outscores a solo doctor on hard cases sounds like progress. But this was a simulation with generalists, not the specialist teams who handle these cases in real life. Microsoft tests it in a real ER next. Watch how it performs with actual patients before you change how you diagnose.

Microsoft's MAI-DxO diagnosed 83% of complex medical cases correctly in a simulation, versus 36% for generalist doctors. It works like a panel of specialists debating each case.

What changed

On October 22, Microsoft published a peer-reviewed study in the journal Science showing that its AI diagnostic system, MAI-DxO, correctly diagnosed 83% of complex medical cases from the New England Journal of Medicine. Generalist physicians working alone with reference materials got 36%, and 22% without references. On emergency department cases, MAI-DxO hit 80% accuracy versus 47% for doctors.

MAI-DxO runs multiple AI agents that act like a panel of doctors. Each agent proposes diagnoses, decides which tests to order, challenges the group's conclusions, and weighs the cost of each test. It uses a step-by-step process instead of seeing the whole case at once, which is closer to how real diagnosis works. The system performed best with OpenAI's GPT-5 and GPT-5.5 models. On GPT-5, it reduced estimated diagnostic costs from roughly $8,400 to $5,400 per case while adding about four percentage points of accuracy. Microsoft plans real-world tests at Beth Israel Deaconess Medical Center in Boston.

What you can do with it this week

If you run a medical clinic, keep an eye on Microsoft's Boston trial results, expected in 2027. MAI-DxO is not available for purchase yet, and there is no announced pricing. The current results come from simulations, not real patients, and the comparison was against generalists working alone, not the specialist teams that typically handle complex cases.

For now, the lesson is the panel method itself: getting several AI models or agents to argue before settling on an answer improves accuracy more than using a single response. You can apply this idea in other high-stakes work like legal research, financial analysis, or contract review. Run the same question through two or three AI tools, compare their answers, and ask one tool to critique another's response before you make a decision.

What to keep in your own hands

The study shows that much of MAI-DxO's advantage comes from the underlying GPT-5 model, not the orchestration layer. GPT-5 alone scored 78.9%, and the panel added about four points on top. On emergency cases, GPT-5.5 hit about 85% accuracy both with and without MAI-DxO, at similar costs. The panel helps most on complex cases and with weaker models.

No AI tool should make a diagnosis without a physician reviewing it, and Microsoft is not claiming otherwise. The system has not been tested on real patients, and simulated cases with known answers are easier than live clinical work. The doctors in the study were generalists, not specialists, and they worked alone, which is not how hospitals handle the hardest cases. Before you change any clinical workflow, wait for the results from the Boston trial with real patients in real emergency and urgent care settings. Even then, the AI's recommendation is input, not the final call. The physician still decides.