Microsoft’s AI-based diagnostic system, the Microsoft AI Diagnostic Orchestrator (MAI-DxO), achieved substantially higher diagnostic accuracy than general physicians in simulation studies reported in the journal Science. Mustafa Suleyman, Microsoft’s CEO for AI, described the results as a step toward so-called medical superintelligence.
Study setup and key results
Researchers evaluated MAI-DxO on two different case sets. The first dataset comprised 304 complex, particularly difficult cases that often involved rare or complicated diseases or otherwise puzzling symptom constellations. In this set MAI-DxO reached about 83 percent diagnostic accuracy. By comparison, general physicians who were allowed to use reference books and internet sources achieved roughly 36 percent accuracy; without external aids their accuracy was about 22 percent.
In the second test, which focused on scenarios seen in emergency departments, MAI-DxO again outperformed general physicians, though by a narrower margin: the system achieved approximately 80 percent accuracy versus about 47 percent for the physicians. The study authors note that the performance gap would likely narrow if the system were compared to experienced specialists in the relevant fields.
How the system works
DxO operates sequentially: patients present an initial complaint rather than uploading a complete list of history and test results, and the system determines what further questions to ask and which tests to order before issuing a diagnosis. This workflow more closely resembles how clinicians typically reason in practice than single-step models do.
The platform coordinates multiple AI-driven agents assigned to different roles: some suggest diagnoses, others select tests, challenge assumptions, or weigh costs. The researchers argue that this committee-like structure can help mitigate certain cognitive biases affecting individual clinicians, such as anchoring on an initial diagnosis or overlooking alternative explanations.
Cost and model testing
Microsoft reported that in some comparisons the system operated at roughly half to a third of the traditional costs while maintaining high diagnostic accuracy. These cost figures are based on simulated cases and estimated test costs, not on a full economic evaluation in a real hospital setting.
MAI-DxO was tested with models from several companies, including OpenAI, Google, Anthropic, Meta, and DeepSeek, and produced its best results with OpenAI GPT-5 and GPT-5.5 models.
Plans for clinical trials and expert caveats
Microsoft plans to evaluate the tool in real clinical settings: the company will test the system in collaboration with Beth Israel Deaconess Medical Center in Boston to examine how it performs within emergency department workflows. Microsoft is also working with the Mayo Clinic to explore the system’s usefulness beyond diagnostic recognition.
Medical experts emphasize that strong performance on diagnostic tests does not mean AI systems can immediately replace physicians. The prevailing view is to position MAI-DxO as an assistant that provides second opinions, suggests alternative diagnoses, highlights subtle but important findings, and helps determine which tests are likely to yield the most new information. Widespread deployment would also raise legal and accountability questions, for example about responsibility if an AI-assisted diagnosis leads to harm.
Related research on tumor cell analysis
The article also notes that researchers at the Szegedi Biológiai Kutatóközpont developed an AI-based method to analyze single tumor cells that incorporates both visual features and molecular properties, potentially enabling new types of tumor diagnostics and research.
Conclusion
MAI-DxO’s simulated results are promising: its sequential, multi-agent design and higher diagnostic accuracy—especially in very difficult cases—stand out. The next step is real-world clinical trials to determine practical applicability, cost implications, legal considerations, and how clinicians accept and integrate such tools into care.
