Microsoft has developed an AI diagnostic tool that solved complex medical cases more than twice as accurately as generalist doctors in a study published in Science.

Unlike many medical AI systems, it works through a case step by step, deciding which questions to ask and which tests to order before reaching a diagnosis.

The tool, called the MAI Diagnostic Orchestrator (MAI-DxO), simulates a panel of clinical experts, Reuters reported.

It gathers information, reviews test results, and revises its working diagnosis as new evidence emerges. Many medical AI systems instead assess symptoms, laboratory results, and scans together.

In a peer-reviewed study published in Science, MAI-DxO, running on OpenAI's GPT-5, reached the correct diagnosis in 83.2% of 304 challenging cases published in the New England Journal of Medicine. Generalist doctors, who had access to textbooks and the internet, scored 36.1%.

The system also outperformed doctors on 325 emergency department admissions, scoring 80.3% compared with 46.8% for the human group.

Researchers said its diagnostic work-ups cost less, too, although those figures are estimates based on the cases rather than actual medical bills.

Dominic King, a corporate vice president in Microsoft AI's health unit, told Reuters that medical AI has largely been evaluated using multiple-choice tests.

“AI has become very good at getting a 100% pass rate on those,” King said. “No patient ever came to see me with a nice summary.”

Harsha Nori, head of AI at Microsoft AI Health, said the researchers also measured how effectively the system selected tests, rather than looking only at whether it reached the right answer.

“We're not just measuring the accuracy of the systems,” Nori told Reuters.

The results come with important limitations. The doctors in the comparison were generalists, not specialists, and the cases were presented as text.

Neither the AI nor the doctors had access to physical examinations or medical imaging in the study.

MAI-DxO can also work with different AI models. An earlier preprint, published in June 2025, reported improvements using models from Google, Anthropic, Meta, and other companies.

The system is not yet being used in clinical practice. Microsoft's researchers plan to work with Beth Israel Deaconess Medical Center in Boston to test whether it can help triage patients in emergency department waiting rooms.