Microsoft’s MAI-DxO diagnostic system scored higher than doctors on two sets of cases in a study reported by Futurism. The results are striking, but they come from a controlled benchmark, not a trial of the system treating patients.
Microsoft AI Watch analysis
What happened
The study, which Futurism says was peer-reviewed and published in Science, tested the Microsoft AI Diagnostic Orchestrator on 304 clinicopathological cases from the New England Journal of Medicine and 325 emergency-department encounters. On the first set, the system achieved 83.2 per cent diagnostic accuracy, compared with 36.1 per cent for generalist physicians who could consult reference materials. On the emergency cases, it scored 80.3 per cent, against the doctors’ 46.8 per cent.
The system worked through cases sequentially, choosing follow-up questions and tests rather than receiving a complete list of symptoms at once. Futurism says the researchers used a “gatekeeper” AI agent to supply requested information from each case record. Microsoft also said the system reached its diagnoses at half or a third of the doctors’ estimated testing cost.
Why it matters
These results suggest that AI may be useful not only for matching symptoms to diagnoses, but for deciding what information to seek and which tests to order. That is a more demanding benchmark than handing a model a finished case file and asking for an answer.
But case records are not patients. The study’s simulated exchanges cannot fully reproduce how people describe symptoms, what they leave out or how a clinician responds. The cost comparison is Microsoft’s reported estimate, not evidence that deploying the system would halve the cost of healthcare.
Our read
This is an impressive research result, and one worth taking seriously without handing the stethoscope to a leaderboard. The numbers show how MAI-DxO performed on these test cases; they do not establish that it is ready to diagnose people in clinical practice. The next question is whether the approach holds up in real care, where the information is messier and the consequences are not benchmark points.
What to watch
- Whether further studies test the system with real patients and clinical workflows.
- How performance compares across different hospitals, patient groups and conditions.
- Whether independent research supports Microsoft’s reported accuracy and testing-cost results.
Discussion spark: What evidence should be required before a diagnostic AI moves from controlled case benchmarks into real clinical care?
Sources and evidence
- Source update (25 October 2026, 10:57 UTC)
not affiliated with or endorsed by Microsoft