Greetings, Earthlings. This week your headlines announced that a machine now diagnoses illness better than your doctors. From where we sit, that is a big claim, so we read past the headline.
Microsoft's AI Diagnostic Orchestrator, known as MAI-DxO, correctly diagnosed about 83% of complex medical cases in a study published in the peer-reviewed journal Science. Doctors given the same cases got about 36% right. That is a real result. It also comes with conditions attached, and the conditions matter more than the number.
The short version
In a peer-reviewed Science study, Microsoft's MAI-DxO solved about 83% of complex diagnostic cases. Generalist doctors working alone solved about 36%.
The cases were simulations built from published case reports. No real patients were involved.
Each doctor worked alone with reference materials. There were no colleagues, specialists or second opinions, which is not how hard cases are handled in a hospital.
Most of the accuracy comes from the underlying AI model. Microsoft's orchestration software adds a smaller boost on top.
A real-world evaluation at Beth Israel Deaconess Medical Center in Boston is planned. Until it reports, this is a lab result, not a clinical one.
What Microsoft actually built
MAI-DxO is not a new AI model. It is an orchestrator: software that makes several copies of an existing model behave like a panel of doctors. One proposes possible diagnoses, another decides which tests to order, another challenges the group's reasoning and another keeps an eye on cost. It works with models from different companies, and its best results came when paired with OpenAI's GPT-5 family.
To test it, Microsoft built a benchmark called SDBench (the Sequential Diagnosis Benchmark) from 304 published case records in the New England Journal of Medicine. These are famously difficult teaching cases. Instead of being handed the whole case at once, the AI or the doctor starts with a short summary of the patient and has to ask questions and order tests step by step, each with a price attached, before committing to a diagnosis.
That design deserves credit. It is much closer to real diagnostic work than the multiple-choice medical exams that most "AI passes medical school" headlines are based on.
What the study found
On the complex cases, MAI-DxO got about 83% right, against about 36% for primary care doctors and general internists. The published version also added emergency department cases from Beth Israel Deaconess, where the system was reported at around 80% against roughly 47% for doctors.
The more revealing numbers sit further down. According to TechCrunch's reporting on the paper:
GPT-5 on its own, with a basic prompt, scored 78.9% on the complex cases. The orchestrator added about four percentage points.
With the newer GPT-5.5, the orchestrator lifted accuracy from about 72% to nearly 87%.
On emergency cases, the orchestration added little.
In other words, most of the diagnostic ability lives in the underlying model, not in Microsoft's layer on top. The researchers themselves say the gains shrink as models get stronger.
The fine print that matters
The doctors worked alone
Each physician worked individually. In real hospitals, cases this complex go to specialists, multidisciplinary teams and second opinions. None of that was part of the comparison. The study answers "can AI beat a lone generalist on a puzzle case?" It does not answer "can AI beat a hospital?"
The human baseline is a design choice, and it moves. When Microsoft first released this work as a preprint in mid-2025, doctors scored about 20% against the AI's 85.5%. In the peer-reviewed version, doctors with access to reference materials scored about 36%. Giving humans some of the tools they normally use closed part of the gap.
These are not typical patients
Journal case records are written after the answer is known, and they are chosen because they are unusual and instructive. Real clinics are mostly common conditions, incomplete histories, patients whose symptoms change, and decisions made under time pressure. The study does not test any of that.
The cost savings are modelled
Microsoft reports lower diagnostic costs, but these are estimates produced by putting prices on tests inside the simulation. They are not hospital bills.
Microsoft checked its own work first
The research comes from Microsoft's own team. Peer review at Science is a meaningful check on the methods, but it is not independent replication. Google and OpenAI have made similar "AI beats doctors" claims about their own systems.
Even Mustafa Suleyman, the CEO of Microsoft AI, who describes this as a step towards "medical superintelligence", says the system needs clinical validation before it is used with patients. Microsoft has not released it and has given no date or price.
Signal vs hype: our reading
Signal: a step-by-step diagnostic test is better evidence than exam scores, and a peer-reviewed result showing strong performance on very hard cases is real progress. AI as a second opinion on complex cases is starting to look plausible.
Hype: "outperforms doctors" and "medical superintelligence". Nothing here shows AI improving outcomes for real patients, and the comparison was set up in a way that flatters the machine.
What to watch next
The Beth Israel Deaconess evaluation. Patients receive standard care, and those who consent also interact with MAI-DxO so its recommendations can be compared with their clinicians'. This is the test that matters.
Independent replication. Results from researchers outside Microsoft, ideally on routine cases rather than textbook puzzles.
Regulation. Any AI used to diagnose patients would have to clear medical device rules. In the UK, that means the MHRA.
Who carries the risk. If an AI suggests a diagnosis and a clinician acts on it, accountability still needs working out.
Frequently asked questions
Is Microsoft's AI better than doctors at diagnosis?
On a set of simulated, unusually difficult teaching cases, it was more accurate than generalist doctors working alone. It has not been shown to be better than doctors in real clinical practice.
Can patients or the NHS use MAI-DxO?
Not yet. It is a research system with no announced release date or pricing, and it would need regulatory approval before use in care.
Was the study peer reviewed?
Yes. It was published in Science in October 2026. Peer review checks the methods, but it is not the same as independent replication or a clinical trial.
End of transmission. We will report back when the real-patient data lands.