Microsoft AI finds the right diagnosis more often than general practitioners in a simulation

A doctor who has to make a difficult diagnosis starts with a few symptoms and little information. Step by step, the doctor asks questions, orders tests and rules out possibilities. Every test costs time and money. Microsoft has built an AI system that mimics exactly this process, and is now presenting the results in the scientific journal Science.

The system is called MAI-DxO. Microsoft tested it on 304 complicated cases from the medical journal The New England Journal of Medicine, which publishes one such difficult case with its final diagnosis every week. According to Microsoft, the system found the right diagnosis in 83 per cent of the cases. General practitioners and general internists who were given the same cases reached 36 per cent when they were allowed to use books and the internet, and 22 per cent without any aids. On emergency department cases, from a database of a hospital in Boston, the system scored 80 per cent against 47 per cent for doctors. It is also said to do this with fewer tests and therefore lower costs.

Microsoft already showed a first version of this research in June 2025, as a preprint. It has now been reviewed by other scientists and published in Science.

A panel of digital doctors

MAI-DxO is not a separate AI model, but a way of making existing language models work together. The system has several AI assistants act as a team of doctors. They ask questions, request tests and discuss the diagnosis. For each test, they weigh up whether it yields enough for what it costs.

To measure this, Microsoft also built a new test. In it, the participant, human or AI, first gets only a small part of the story and has to decide which information to ask for next. Microsoft is making this test public for some of the cases, so that other researchers can use it too.

The underlying model does a lot of the work

The study also shows where most of the progress comes from. Tech site TechCrunch points out that the language models themselves are already very good. Without Microsoft's system, OpenAI's GPT-5 already found the right diagnosis in 78.9 per cent of the difficult cases, at about 8,400 dollars (about 7,500 euros) in tests per case. With MAI-DxO, that rose by about 4 percentage points and the costs fell by about 3,000 dollars (about 2,700 euros) per case. With the newer GPT-5.5, the gain was larger: from 72 to almost 87 per cent. On the emergency department cases, GPT-5.5 scored about 85 per cent with and without the system.

The system also helped a lot with a smaller and cheaper model, Google's Gemini 2.5 Flash: from 46 to almost 69 per cent. Even so, the difference between the small model and the large ones remained bigger than what MAI-DxO added.

The doctors in the comparison were general practitioners and general internists working alone. In practice, such complicated cases usually go to specialists or a team of doctors, and they did not take part. The cases were also written-up cases, not real patients in the consulting room.

Next step: real patients

That step is coming now. Microsoft is going to work with Beth Israel Deaconess Medical Center in Boston. Patients there simply receive normal care, and those who give permission also talk to MAI-DxO. The system's advice is then compared with the doctors' judgement. According to Dominic King, who leads the health division at Microsoft AI, the technology is ready for such a test in practice.

Microsoft is also working with the American Mayo Clinic on its own large AI model for healthcare, which is due to appear next year. In a second article, in the journal BMJ Digital Health and AI, the company describes what is needed for what it calls "medical superintelligence": AI that connects knowledge and decisions across all of healthcare and eventually becomes better than the strongest medical teams of today.

What does this mean for the road to the singularity?

This research shows that AI, in one of the hardest thinking tasks in medicine, making a diagnosis, already scores better than individual doctors in a simulated situation. What stands out is that the biggest jumps come from ever better language models: every new model automatically makes such a system stronger. A big advantage of such models is that they have knowledge from all specialisms at once. A general practitioner cannot know everything about rare heart, skin and metabolic diseases at the same time, but an AI model can make those connections. The trial in Boston has to show whether that also works in the consulting room. If it does, expert diagnosis could become available much more widely and cheaply, also in places where there are few specialists.

What is the technological singularity? Read the explanation →