Microsoft AI has published its diagnostic system, the MAI Diagnostic Orchestrator (MAI-DxO), in the journal Science. Rather than giving one instant answer, it runs several AI agents that act like a panel of doctors: they question a simulated patient, order tests and debate the diagnosis step by step.
On complex cases drawn from the New England Journal of Medicine, Microsoft reports 83% accuracy, against 22% for physicians working without reference resources and 36% with them. On emergency department cases from the MIMIC-IV database, it reports 80% against 47%. Microsoft says the gains held with newer models such as GPT-5.5, and that one cost-optimised setup reached 81% on the hard cases at less than half the cost of comparable systems.
Why it matters: this is a peer-reviewed test of AI working through a diagnosis the way clinicians do, not just answering exam questions. The caveats are real. The doctors were generalists, not the specialists who usually handle such cases, and TechCrunch notes that much of the gain comes from the underlying models.
Microsoft itself says “the true test is performance in real clinical environments”, and plans a prospective study with patients at Beth Israel Deaconess Medical Center in Boston.
Sources: Microsoft AI, Building medical superintelligence; the paper in Science