An artificial intelligence of Microsoft obtained better diagnoses than general practitioners in a study published on October 22 in the journal Science. In complex medical cases, the MAI-DxO system reached 83% accuracy, compared to 36% for physicians authorized to consult books and the Internet.

The result is impressive, but it does not mean that a digital doctor is ready to receive patients. MAI-DxO worked on historical files transformed into structured simulations. His next trial will be a prospective study in a real hospital.

A virtual team that moves step by step

Conventional medical assessments often give AI all symptoms, examinations and results at the same time. In real life, a doctor starts with incomplete information. It must question the patient, examine different assumptions and select the tests that will confirm or eliminate them.

The [Microsoft AI Diagnostic Orchestrator] (https://microsoft.ai/news/building-medical-superintelligence/) reproduces this process. Several software agents play the roles of a group of clinicians: they propose possible diagnoses, discuss missing information and choose the next questions or analyses. An orchestrator coordinates their exchanges, then updates the conclusion as new data appear.

The system is not limited to a single model. Microsoft claims to have tried its architecture with models such as OpenAI, Google, Anthropic, Meta and DeepSeek. The best published results were obtained with GPT-5 and GPT-5.5 models.

Much higher results in simulation

The study, entitled Sequencial diagnosis with language models, is based first on 304 complex cases from the clinical pathological lectures of the New England Journal of Medicine*. Each participant received a short initial presentation and had to request the necessary information before proposing a diagnosis.

MAI-DxO obtained 83% accuracy. General practitioners reached 22% without external resources and 36% with access to references. In a second assessment based on MMIC-IV-ED emergency admissions, the AI obtained 80%, compared with 47% for physicians.

The researchers also allocated an estimated cost to the examinations requested. Some MAI-DxO configurations maintained high accuracy by choosing cheaper investigations. According to Microsoft, no other human or automated approach evaluated was simultaneously more accurate and cheaper.

These costs remain theoretical: they are based on integrated pricing on the test bench and not on the actual costs of a Canadian hospital.

A better test, but not the real clinic yet

The SDBench test bench is an improvement over the questionnaires where all information is provided in advance. It obliges the AI to decide what to ask for, avoiding that it automatically benefits from a full file.

Nevertheless, there is still a simulation. A program called "Gatekeeper" reveals the data when the AI or the doctor requests it. A true patient can describe his symptoms in an imprecise way, forget a drug, present several problems at the same time, or show signs that a clinician observes without them being included in the record.

New England Journal of Medicine cases are also selected because they are difficult or unusual. The comparison group consisted of generalists, not teams of specialists already familiar with the diseases concerned. The percentages cannot therefore be transposed directly to all consultations.

Availability, prices and Canada

MAI-DxO is a research system, not a service that patients, doctors or hospitals can buy today. Microsoft did not report any prices, marketing dates, or Canadian availability. Nothing in the advertisement indicates a Health Canada authorization for clinical use.

Microsoft plans a prospective study with the Beth Israel Deaconess Medical Center in Boston. Patients should interact directly with the system, while its findings will be compared to independent physician assessments. This step will measure his behaviour with real and imperfect information.

Limits or Considerations

The authors of the study work mostly for Microsoft, which develops MAI-DxO. Publication in Science enhances methodological credibility, but does not replace an independent reproduction or regulatory evaluation.

High average accuracy can also mask serious errors in certain groups or diseases. Prior to deployment, false diagnoses, biases, confidentiality of information, responsibility for error, and the tendency of professionals to overtrust a convincing response will need to be measured.

Conclusion

MAI-DxO shows that a team of agents who question, debate and revise their assumptions can do better than a model that is responsible for responding immediately. The results clearly exceed those of the generalists in the simulations studied.

However, technology remains a research tool. The next decisive step will not be a better percentage on a historical record, but the demonstration that it improves real care under the supervision of professionals, without creating new risks for patients.