Microsoft AI Diagnoses Tough Cases Better

Microsoft says its new AI tool can diagnose difficult medical cases more accurately than doctors who are not specialists. In a simulation run by the company, the AI was right 83% of the time on hard cases. Primary care doctors and general internists, working on their own with reference materials, were correct 36% of the time.

The system is called MAI-DxO, often shortened to DxO. It works like a team of doctors. Multiple AI agents question a patient, order tests, and debate the correct diagnosis. The results of this study were published Thursday in the journal Science.

The study's own data indicates that much of the improvement comes from more powerful underlying AI models. Microsoft's system adds gains on top of those models. Microsoft tested DxO with models from several companies, but its best results came with OpenAI's GPT-5 and GPT-5.5 models.

There are important limits to the findings. The cases in the simulation are so complex they are usually handled by specialists or physician teams, who were not part of the comparison. The AI's cost is also a factor. On its own, GPT-5 scored 78.9% on complex cases at about $8,400 per case. DxO reduced that cost but only raised accuracy by about four points.

On emergency department cases, the AI showed little improvement. For those cases, GPT-5.5 with and without DxO both scored about 85% at a similar cost. The AI helps most on complex cases and with less capable AI models, though the gap between weaker and stronger models was larger than the boost the system provided.

Mustafa Suleyman, Microsoft AI's CEO, called the result a step toward 'medical superintelligence.' He said the company is working with the Mayo Clinic to train a foundation model for healthcare, with a release planned for next year. He believes that model will enable the company to 'make really enormous leaps here.'

Suleyman has spent his career trying to use AI to solve medical issues. He was previously a co-founder of DeepMind, which was acquired by Google in 2014. He joined Microsoft to lead its AI unit in 2024. He expects healthcare models to get 'an order of magnitude better in capability within the next 24 to 36 months.'

Through a new partnership with Beth Israel Deaconess Medical Center in Boston, Microsoft will evaluate the clinical reasoning capabilities of DxO across different hospital settings. Patients will get standard treatment. Those who consent will also engage with DxO, and the partners will compare its recommendations against human doctors' decisions.

Microsoft's study coincides with a manifesto about what it will take to build medical superintelligence, which is defined as 'the capacity to surpass the strongest contemporary human-led health systems in medical reasoning and coordination across the full arc of care.' Microsoft hopes to move beyond benchmarking individual tasks and into integrating with clinicians and healthcare systems for better outcomes.