Can AI Diagnose Patients Better Than Doctors and What the Microsoft Study Really Reveals

Few headlines grab attention like an AI beating doctors, and this one has serious science behind it. On October 22, Microsoft's AI team published results in Science showing that its MAI Diagnostic Orchestrator, or MAI-DxO, solved difficult cases far more often than generalist physicians, a finding Reuters reported on October 23.

Rather than answering a tidy multiple-choice question, the system behaves like a small clinical panel. It decides what to ask, orders tests, weighs the results and revises its working diagnosis as new evidence arrives, which is much closer to how a real work-up unfolds. For patients, it raises a simple question: could faster, smarter diagnosis be closer than we thought?

The headline figures come from a peer-reviewed study that tested the system on 304 challenging case reports from the New England Journal of Medicine, with the AI running on OpenAI's GPT-5. The comparison group was generalist doctors, and the gap was wide:

• 83.2% accuracy for MAI-DxO, against 36.1% for generalist doctors who had textbooks and internet access
• 80.3% versus 46.8% on 325 emergency department admissions
• Lower estimated testing costs, although those figures come from case-based estimates rather than actual hospital bills

Before anyone declares the stethoscope obsolete, the limits deserve attention. The cases were historical and text-only, so neither the AI nor the doctors could perform physical examinations or review imaging.

The physicians were generalists, not the specialist teams who usually handle such puzzles, and early analysis of the paper notes that much of the gain seems to come from stronger underlying models rather than the orchestration layer alone. Dominic King, a corporate vice president in Microsoft's health unit, told Reuters that medical AI has mostly been tested with exams, adding, "No patient ever came to see me with a nice summary."

That is precisely why sequential testing matters. For hospitals, the business case is clear: fewer unnecessary tests and faster triage could ease cost pressure, and Microsoft is pitching this as a step toward medical superintelligence. Reaction across the industry has mixed praise for the peer-reviewed rigor with calls for trials in live clinics.

The next real test arrives soon, with researchers planning to work with Beth Israel Deaconess Medical Center in Boston to see whether the system can help triage patients in emergency department waiting rooms. Until that evidence lands, the safest reading is that MAI-DxO is a powerful decision-support prototype, not a replacement for clinicians.