A research paper on Microsoft AI's diagnostic system MAI-DxO has been published in Science, a major step toward what the company calls "medical superintelligence."
1. 83% vs 36% on hard cases. The study used 304 complex cases from the New England Journal of Medicine, simulating real clinical workflow: the AI could not just answer exam questions — it had to question the patient, order tests, and adjust its judgment like a doctor. A "consultation panel" of multiple AI agents correctly diagnosed 83% of the hard cases, while primary care physicians and general internists working individually with reference materials managed only 36%. On emergency-department cases it was 80% vs 47%.
2. A multi-agent "consultation panel." MAI-DxO's core is not a single model answering questions, but several AI agents each acting as a consulting doctor — proposing diagnoses, picking tests, debating over multiple rounds, and converging on a final diagnosis. Microsoft validated the framework across models from OpenAI, Google, Anthropic, Meta, and DeepSeek, with the best results on GPT-5 / GPT-5.5, and open-sourced the SDBench evaluation benchmark for the whole industry to check.
3. The caveats are clear. The paper states that such hard cases are normally handled by specialists or physician teams, who were not in the comparison — and that most of the accuracy gain comes from stronger base models, with the orchestration adding gains on top. Microsoft AI CEO Mustafa Suleyman says the next step is real-world clinical trials with partners such as Mayo Clinic.
Source: Read the original article
Comments