Microsoft’s diagnostic AI scores 83% in simulated medical cases
Microsoft says its MAI-DxO system correctly diagnosed 83% of 304 complex medical cases, compared with 36% for generalist doctors using reference materials. The study, published in Science, used existing cases rather than treating patients; doctors were assessed individually, and the researchers did not compare the system with specialist teams.
Bottom line — MAI-DxO remains a research system; Microsoft’s results need testing with real patients.
Go deeper 6
-
MAI-DxO uses multiple AI agents to ask questions, consider tests and challenge one another’s conclusions.
-
Doctors scored 22% without reference materials; in historical emergency cases, the AI scored about 80% versus 47% for doctors.
-
The AI was tested on 304 complex cases, while doctors assessed a separate set of 56 cases.
-
In one cost-conscious simulation, AI reached about 81% accuracy at an estimated €2,070 (US$2,241) per case; doctors averaged €4,361 (US$4,725). These were estimated work-up costs, not patient bills.
-
OpenAI’s GPT-5 scored about 79% on complex cases without Microsoft’s orchestration, rising to about 83% with it; GPT-5.5 rose from about 72% to nearly 87%.
-
The study’s authors work for Microsoft, and everyday clinical performance remains untested.