Microsoft AI Diagnostic Orchestrator Beats Doctors in Tests
By Ganesh Karki
Key facts
- Section
- Tech & AI
- Country
- USA
- Published
- 18 Nov 2026
Microsoft AI Diagnostic Orchestrator scored 83.2 percent on hard diagnosis cases in a peer-reviewed study, against 36.1 percent for generalist physicians using reference materials. The paper, published in Science on 22 October 2026, turns a much-discussed 2025 demonstration into validated research. The system is still not a product, and nobody has used it on a live patient yet.
Microsoft AI Diagnostic Orchestrator at a glance
| Detail | Information |
|---|---|
| Published | Science, 22 October 2026 |
| System | MAI-DxO, running on OpenAI GPT-5 |
| Hard cases | 83.2% on 304 New England Journal of Medicine cases |
| Physicians | 36.1% with references, 22.3% without |
| Emergency cases | 80.3% on 325 cases, against 46.8% for physicians |
| Benchmark | SDBench, open-sourced under an MIT licence |
| Status | Research only, no product or release date |
What happened
MAI-DxO is not a model. It is an orchestration layer that makes a frontier model work like a panel of doctors, asking questions one at a time and ordering tests in stages rather than answering from a full case file. Microsoft first showed it in June 2025 alongside a preprint, when it scored 85.5 percent against 20 percent for 21 physicians working without references. The peer-reviewed version adds the arm that was missing: doctors who could search and consult references scored 36.1 percent, narrowing the lead from about fourfold to about 2.3 times.
TechRepublic reports the study also tested 325 emergency department cases from the MIMIC-IV-ED database, where the system scored 80.3 percent against 46.8 percent for physicians. TensorFeed notes the system asked an average of 25.5 questions before ordering its first test, against 5.6 for physicians, and that Microsoft has announced a prospective study with Beth Israel Deaconess Medical Center in Boston.
Why it matters
Diagnosis is one of the highest-stakes uses of AI, and a peer-reviewed result carries weight that a company blog post does not. The study shows orchestration adds real value: plain GPT-5 with a basic prompt scored 78.9 percent on the same cases, so the orchestrator added about four points and cut roughly 3,000 dollars per case in estimated testing costs. It also matters as a contrast to the week's other health AI news, from the Apple health AI study for up to 200,000 people to ChatGPT healthcare agents, because this one is a benchmark result, not a product launch.
What is not confirmed
The cases were historical and text-only, with no physical exam, imaging or pathology slides, and every case had a confirmed diagnosis, so the study could not measure false-positive rates. The human comparison was individual generalists, not specialists or teams, and the authors say they would expect the gap to narrow against specialists. The cost figures are simulated diagnostic spending, not hospital savings. Microsoft has not said when the Beth Israel study will start or when, if ever, the system will reach clinical use.
Sources
Sources and references
Information checked: 18 November 2026
Related stories
SpaceX Falcon 9 NROL-136 Launch: What to Watch on 21 November
SpaceX Falcon 9 NROL-136 launch is scheduled for 21 November from Vandenberg, carrying reconnaissance satellites for the National Reconnaissance Office.
20 November 2026AO3 Returns After 17 Days Offline After Security Breach
AO3 returns after 17 days offline following a security incident, with some features still disabled, and went down again briefly on 19 November for maintenance.
20 November 2026ChatGPT for Accounting Launches With QuickBooks Link
ChatGPT for Accounting launched this week, connecting the chatbot to QuickBooks, Xero and NetSuite so finance teams can reconcile books and close the month.
20 November 2026Tesla Cybercab Rides Begin in Dallas With V1.5 Cars
Tesla Cybercab rides began in Dallas on 19 November, the second city after Austin to get the steering wheel-free robotaxi, with upgraded V1.5 vehicles.