Microsoft is making significant advances in healthcare through an artificial intelligence system that has demonstrated more than twice the diagnostic accuracy of human doctors in a study involving complex medical cases. According to research published in the journal Science, Microsoft AI CEO Mustafa Suleyman believes the findings represent an important step towards what he describes as “medical superintelligence”.
How does the AI doctor work?
The system, called MAI-DxO, or Microsoft AI Diagnostic Orchestrator, functions differently from a conventional chatbot. It operates like a virtual team of doctors, using multiple AI agents that work together to ask questions, request laboratory tests and discuss symptoms before arriving at a diagnosis.
In a simulation involving challenging medical cases published in the New England Journal of Medicine, MAI-DxO reached the correct diagnosis in 83% of cases. By comparison, general physicians equipped with medical reference tools achieved an accuracy rate of just 36%.
However, the comparison has a limitation: cases of this complexity would typically be handled by teams of medical specialists, who were not included in the study. Nevertheless, Dominic King, Microsoft's Vice President of Health, described the findings as highly significant and encouraging, highlighting the shortage of medical experts as one of healthcare's biggest challenges.
The role of GPT-5 and GPT-5.5
Microsoft's findings add to similar claims from companies such as Google and OpenAI that advanced AI systems can match or outperform human doctors in certain diagnostic tasks.
The study found that the underlying AI models played a major role in determining performance. Microsoft evaluated systems from OpenAI, Google, Meta and other companies, with OpenAI's GPT-5 and the newer GPT-5.5 delivering the strongest results.
The DxO system also helped reduce costs. In one example, GPT-5 alone achieved 78.9% accuracy on a complex case at a cost of approximately $8,400. Incorporating Microsoft's DxO system reduced the cost by around $3,000 while slightly improving diagnostic accuracy.
When paired with GPT-5.5, DxO increased accuracy from 72% to nearly 87% on complex cases while keeping costs relatively low.
The system also improved the performance of less powerful models. Google's Gemini 2.5 Flash achieved an accuracy rate of only 46.4% when operating independently, but its performance rose to nearly 69% when used with Microsoft's DxO framework.
However, the system did not produce substantial improvements in accuracy or cost when advanced models were tested on fast-paced emergency department cases.
What happens next?
Suleyman, who previously co-founded Google's DeepMind, has spent much of his career exploring how AI can address medical challenges. He expects healthcare AI systems to improve considerably over the next two to three years.
Microsoft is already collaborating with the Mayo Clinic to develop a specialised AI model for healthcare, with a release expected next year. Suleyman told TechCrunch that by 2027, doctors could have access to highly accurate AI predictions about how a patient's medical history and condition might develop over time.
Before introducing DxO into routine patient care, Microsoft is testing the system in a real-world clinical setting through a partnership with Beth Israel Deaconess Medical Center in Boston.
During the trial, patients will continue receiving standard treatment from human doctors while the AI operates in the background without influencing their care. Researchers will subsequently compare the system's recommendations with the decisions made by the medical professionals.
Suleyman said the technology was something people around the world wanted and pledged that Microsoft would work to introduce it safely into hospitals globally.
