Microsoft’s AI Doctor Panel Outscored Generalists on Hard Cases
Most medical AI is tested like a student taking a multiple-choice exam.
Real diagnosis starts with a few facts, and the doctor has to decide what to ask and which test to order next.
That early work also shapes who ends up in an operating room, so diagnostic quality reaches every surgical decision.
Microsoft researchers published a study in Science on October 22, 2026 that tested an AI system built to work the way a doctor does, one step at a time.
Five Voices, One Decision
The system is called the Microsoft AI Diagnostic Orchestrator, or MAI-DxO.
A single language model plays five roles that debate each step of a case.
One role tracks the likely diagnoses, one chooses tests, one argues against the leading idea, one watches cost, and one checks for mistakes.
After the debate, the panel decides whether to ask a question, order a test, or commit to a diagnosis.
The design sits on top of different AI models, and it improved accuracy for most of the models tested.
How the Test Worked
The researchers converted 304 New England Journal of Medicine case records into step-by-step encounters.
They added 325 emergency department admissions drawn from a hospital records dataset.
Each case began with a summary, and every other detail had to be requested by asking a question or ordering a test.
Paired with GPT-5, the orchestrator reached 83.2% accuracy on the journal cases, while generalist physicians reached 36.1%.
On the emergency admissions, it reached 80.3%, against 46.8% for the physicians.
The physicians could use clinical reference tools such as UpToDate, but not AI.
How It Caught Scurvy
Standard AI models tended to lock onto their first impression.
The challenger role forced the orchestrator to look for evidence against its own leading idea.
One patient had months of falls, leg weakness, easy bruising, bleeding gums, and tooth loss after gastric bypass surgery.
A standard GPT-5 setup pursued nerve and autonomic disorders with a sprawling work-up and never reached the correct diagnosis, which was scurvy.
The orchestrator connected the gum bleeding, tooth loss, and diet, then ordered a single plasma vitamin C test.
Its estimated testing cost was $1,161, compared with $8,806 for the standard setup.
What the Study Cannot Tell You
The authors are Microsoft researchers, and the results come from a simulation rather than real patients.
An AI model played the patient record, and it generated realistic findings for tests that the original cases never reported.
The physicians were individual generalists rather than specialists or care teams, and they worked through a 56-case subset of the journal cases.
The authors state that the results do not yet establish clinical effectiveness in real-world decision support.
Cost also needs a careful reading, because the standard setting that scored 83.2% averaged an estimated $5,430 per journal case against $4,725 for the physicians.
The budget-limited versions cost less, at $2,241 for 80.9% accuracy and $1,302 for 79.6%.
Why Surgeons Should Watch
Diagnosis sits upstream of every operation, so tools that sharpen the work-up could change who reaches the operating room and how prepared the team is.
The lasting idea is the method, which makes a team challenge its first answer before it spends money on the next test.
Testing with real patients in real clinics is the step that will show whether the panel approach survives outside a benchmark.
Related reading
The 2026 Surgical Technology Intelligence Report covers which technologies have actually reached
the operating room.
