Microsoft’s Mustafa Suleyman says its AI outperforms doctors on hard cases as he chases ‘medical superintelligence’

Microsoft published a study in Science showing its MAI-DxO system correctly diagnosed complex New England Journal of Medicine cases 83 percent of the time in a simulation, against 36 percent for primary care doctors and general internists working alone with reference materials. Microsoft AI chief Mustafa Suleyman called it a step toward 'medical superintelligence.'

The stakes are obvious. Microsoft's own health lead framed the problem as a growing mismatch between demand for medical expertise and supply. A tool that helps clinicians reason through hard cases could save lives and time. A tool that is overtrusted could do real harm.

We're encouraged, and we want to be precise about why. Peer review in Science is a higher bar than a company blog post. The system works like a panel of AI agents that question a patient, order tests and argue toward a diagnosis, and Microsoft measured cost per case alongside accuracy, which is the right instinct for anything that might touch real budgets.

Now the fine print, because it matters. The comparison was against generalists, not the specialists or physician teams who usually handle cases this complex. It was a simulation, not a clinic. And the study's own data suggests much of the improvement comes from stronger underlying models, mostly OpenAI's GPT-5 and GPT-5.5, rather than from Microsoft's orchestration layer. On complex cases, DxO lifted GPT-5.5 from 72 percent to nearly 87 percent. On emergency department cases, it added little to either accuracy or cost. The orchestration helps most with weaker models and hard cases, and the gap between weak and strong models was bigger than the boost DxO gave.

That tells us the model race, not the wrapper, is doing the heavy lifting. It also tells us headline percentages shouldn't travel without their context.

The next step is the right one. Microsoft is partnering with Beth Israel Deaconess Medical Center to run DxO in parallel with standard care for consenting patients and compare its recommendations against doctors'. Google ran a similar real-world study with Beth Israel, published in The Lancet this month. Parallel testing, with humans still making decisions, is how this should go.

Where this leads if it scales is the question I'm watching. Suleyman expects health models to get an order of magnitude better within 24 to 36 months and says Microsoft is training a healthcare foundation model with the Mayo Clinic for release next year. If diagnostic AI becomes a product Microsoft must lead commercially, the pressure to ship will be enormous. Liability, bias across patient populations, and what happens when clinicians defer to a confident machine all need answers before deployment, not after.

Microsoft, Google and OpenAI have all claimed their AI can match or beat doctors on some tasks. Microsoft already sells dictation and radiology tools to clinicians and sees DxO as an extension of that business. Suleyman, a DeepMind co-founder, has led Microsoft's AI unit since 2024.