The Microsoft AI diagnostic study reports that an experimental system correctly identified 83% of complex medical cases, compared with 36% for generalist physicians. The findings, published Oct. 22 in Science, highlight advances in AI-assisted diagnosis but do not establish superiority in real-world patient care.
Microsoft’s system, called MAI-DxO, coordinates multiple AI agents that simulate a medical team. They review symptoms, request tests, and evaluate possible diagnoses before concluding.
Microsoft AI CEO Mustafa Suleyman described the results as progress toward medical superintelligence. The company’s broader ambition is to develop AI systems that support clinical decisions throughout a patient’s care.
Microsoft AI Diagnostic Study Reveals Performance Differences
Researchers evaluated MAI-DxO using difficult cases from The New England Journal of Medicine. The system achieved approximately 83% accuracy, while individual primary care physicians and general internists scored 36% with access to reference materials.
However, these cases typically require specialist expertise or multidisciplinary teams. The study did not compare the AI against those teams, limiting conclusions about its performance in hospitals.
Microsoft tested the system using models from OpenAI, Google, Anthropic, Meta and DeepSeek. OpenAI’s GPT-5 and GPT-5.5 produced the strongest results.
On complex cases, GPT-5 achieved 78.9% accuracy without orchestration. MAI-DxO improved that result by roughly four percentage points while reducing estimated diagnostic testing costs.
GPT-5.5 showed a larger improvement, rising from approximately 72% accuracy to nearly 87% with DxO. However, the added system delivered little measurable benefit on emergency department cases, where both configurations scored around 85%.
The results suggest that stronger underlying models account for much of the progress. Microsoft’s orchestration system can provide additional gains, particularly on challenging cases and with less capable models.
Hospital Testing Will Examine Real-World Performance
Microsoft plans to evaluate MAI-DxO with Beth Israel Deaconess Medical Center in Boston. The research will compare its diagnostic recommendations with those of practicing clinicians.
Consenting patients will receive standard medical care while also interacting with the AI system. Researchers will compare its assessments with physicians’ decisions across emergency and urgent care settings.
The clinical evaluation follows broader research into agentic AI in healthcare. These systems can coordinate multiple diagnostic tasks but still require evidence of clinical reliability and safety.
Google has also advanced its medical AI research. A June study of AMIE reported strong performance on simulated disease-management tasks, including several measures where the system outperformed physicians.
In October, Google reported results from a separate clinical study involving 98 patients at Beth Israel. Doctors supervised the AI conversations, and clinicians found its summaries helpful in preparing for appointments in 75% of cases.
OpenAI has also expanded its involvement in healthcare, as NPR’s coverage of AI doctors illustrates. However, successful diagnostic simulations do not automatically demonstrate safe, independent patient care.
Microsoft Targets Medical Superintelligence
Suleyman is also working with Mayo Clinic to develop a specialized healthcare foundation model. Microsoft plans to introduce the model in 2027, though its clinical capabilities are still under development.
The company already provides medical documentation and radiology software, including Dragon Copilot. Its expansion into diagnostic reasoning builds on broader efforts to move AI beyond clinical note-taking.
Microsoft expects medical AI capabilities to improve substantially over the next several years. Suleyman believes AI could eventually help clinicians anticipate changes in patient records and identify medical problems earlier.
For now, MAI-DxO remains an experimental diagnostic system. Hospital studies must establish whether benchmark improvements translate into reliable clinical decisions, lower costs, and better patient outcomes.
