Microsoft's Diagnostic AI Doubles Doctors' Accuracy in Step-by-Step Simulation

Microsoft's Diagnostic AI Doubles Doctors' Accuracy in Step-by-Step Simulation

A simulation that works like a real clinic visit

Microsoft's AI diagnostic system has beaten generalist physicians by a wide margin in a simulated diagnosis challenge. The system, called MAI-DxO, diagnosed 83.2 percent of complex cases correctly, compared with 36.1 percent for participating doctors; on emergency-department cases it scored 80.3 percent versus 46.8 percent for physicians. The peer-reviewed findings were published in the journal Science on October 22, 2026.

No more multiple choice: the AI must ask its own questions

Most medical AI evaluations have relied on multiple-choice tests where the model sees the full case up front. Microsoft's new Sequential Diagnosis Benchmark, or SDBench, instead starts the AI with only a brief chief complaint. The system must request history, choose exams and tests, and converge on a diagnosis step by step. Dominic King, a corporate vice president in Microsoft AI's health unit, told Reuters that earlier tests missed how doctors actually work. Microsoft says it will open-source SDBench along with its gatekeeper, judge, and cost-estimation code.

A panel of virtual doctors that argue with each other

MAI-DxO is not a single language model but an orchestration layer: multiple AI agents act as a panel of virtual physicians, each proposing diagnoses, selecting questions or tests, and challenging the emerging view as new information arrives. Microsoft tested MAI-DxO with models from OpenAI, Google, Anthropic, Meta, and DeepSeek, and its best results came with OpenAI's GPT-5 and GPT-5.5. The AI asked an average of 25.5 questions before ordering its first test, compared with 5.6 for physicians.

Microsoft calls it a step toward medical superintelligence

Microsoft AI CEO Mustafa Suleyman described the work as a step toward "medical superintelligence," telling The Verge the company aims to build something "indistinguishable from human performance" within a couple of years. Microsoft's official blog presents SDBench as a test of sequential reasoning and MAI-DxO as its leading system on that benchmark. The company plans a prospective study with Beth Israel Deaconess Medical Center in Boston to test the system in emergency-department triage, with patients interacting directly with it.

A cheaper configuration beats doctors for under a third of the cost

The paper also compares estimated diagnostic costs. The unconstrained MAI-DxO configuration cost an estimated 5,431 dollars per case, versus 4,725 dollars for physicians. But a budgeted configuration reached 79.6 percent accuracy at about 1,303 dollars per case, beating the doctors' 36.1 percent at less than a third of their estimated cost; an intermediate setup hit 80.9 percent at 2,241 dollars. The orchestrator lifted base models by an average of 7.9 percentage points on the complex cases and 4.4 points on emergency cases.

Generalists gain a demanding assistant

The near-term impact falls on generalist and emergency medicine, where systematic questioning and test selection matter most. The caveat is large: SDBench measures benchmark performance in a text simulation, not clinical outcomes, and earlier research has shown general-purpose chatbots performing poorly in realistic diagnostic workflows. Success in a structured simulation does not automatically transfer to a hospital ward.

A live trial in an emergency waiting room

The next test is real patients. The planned Beth Israel Deaconess collaboration would evaluate the system at triaging patients in emergency waiting rooms. With SDBench open-sourced, rival teams can compete on the same yardstick. Whether the simulation's twofold advantage survives contact with the clinic will depend on that prospective evidence.