Microsoft's AI Diagnostic Tool: Medical Superintelligence Claims

A new AI system from Microsoft diagnosed complex medical cases correctly at more than double the rate of generalist physicians in a recent simulation. The results, published in Science, were described by Mustafa Suleyman, the head of Microsoft's AI division, as a step toward building "medical superintelligence." The system, named MAI-DxO, was created to help address a growing gap between the demand for medical expertise and the number of available experts.

The system operates by coordinating multiple AI agents that work together like a panel of human doctors. They are designed to question a patient, order necessary tests, and then debate a diagnosis. The study compared this process to the work of individual primary care doctors and general internists. It's important to note that the test cases, taken from the New England Journal of Medicine, are of a complexity usually handled by specialists or teams of physicians, who were not part of this comparison.

In the simulation, Microsoft's system was correct 83% of the time, while the doctors were correct 36%. The study also found that much of this performance gain came from more powerful underlying AI models, with Microsoft's system providing additional improvements on top of them. While Microsoft tested its system with models from several companies, its best results were achieved using OpenAI's GPT-5 and GPT-5.5 models. The system's performance was also evaluated based on the estimated cost of running it for each case.

Looking ahead, Microsoft is preparing to test its diagnostic tool in real-world clinical settings. Through a new partnership with Beth Israel Deaconess Medical Center in Boston, the company will evaluate the system's clinical reasoning capabilities. Patients will receive standard medical treatment, and consenting participants will also use the AI system in parallel. Microsoft and its partners will then compare the AI's recommendations with those made by the human doctors.

The publication of this study is part of a broader strategy from Microsoft to move beyond single-task benchmarks and develop AI that can integrate with clinicians and healthcare systems. The company is also working with the Mayo Clinic to train a foundation model for healthcare, with a release planned for next year. Suleyman stated that in 2027, the company will be "on the verge of a coding style breakthrough for medical records," enabling highly accurate predictions about how a medical record might unfold.

Microsoft is not the only company pursuing this goal, as Google and OpenAI have also made similar claims about their AI's medical capabilities. However, Microsoft is committing significant resources to this area. The company's leadership has stated that healthcare models will get "an order of magnitude better in capability within the next 24 to 36 months" as they are used in real-world applications to identify and patch gaps. Suleyman has affirmed that Microsoft aims to be the provider of the best medical clinical decision support products in the coming years.