Microsoft's Diagnostic AI Solves 83% of Hard Cases in Science Paper, but Only in a Simulation
MAI-DxO, an orchestrator of several AI agents, beat generalist doctors on 304 published case reports. Microsoft's own tests show how much of the gain comes from the model underneath.
Outspoken Digest Technology Desk
Friday, October 23, 2026/2 min read
Microsoft says an AI system it built reached the correct diagnosis in 83.2 percent of 304 difficult cases drawn from The New England Journal of Medicine, against 36.1 percent for generalist doctors, in a peer-reviewed paper published in Science on 22 October. The Next Web reported the figures, and TechCrunch carried the same numbers alongside an interview with Microsoft AI chief Mustafa Suleyman.
It is a simulation, not a clinic. Nobody was treated, and the comparison group was narrower than the headline suggests.
What is MAI-DxO and how was it tested?
MAI-DxO coordinates several AI agents that behave like a panel of doctors: they question a patient, order tests and debate a diagnosis. In the study the cases were presented as text, so neither the AI nor the human doctors had a physical exam or medical imaging, The Next Web said. The doctors were generalists, primary care physicians and general internists working alone with reference materials. Specialists and physician teams, who usually handle cases this hard, were not in the comparison, TechCrunch noted.
A second test used 325 emergency department admissions, where MAI-DxO scored 80.3 percent against 46.8 percent for the human group, according to The Next Web.
How much credit goes to the orchestrator?
Less than the headline implies. Microsoft ran the orchestrator on models from OpenAI, Google, Anthropic, Meta and DeepSeek, and the best results came with OpenAI's GPT-5 and GPT-5.5, TechCrunch said. With GPT-5.5, accuracy on complex cases rose from about 72 percent to nearly 87 percent. On emergency cases the gain was negligible: about 85 percent with or without the orchestrator. TechCrunch's summary is that most of the improvement appears to come from the underlying models, and that the orchestrator helps most on weaker models and harder cases.
Dominic King, who leads health at Microsoft AI, was quoted by The Next Web saying "No patient ever came to see me with a nice summary," a nod to the gap between tidy case reports and real consultations.
What happens next, and what should readers take from it?
Microsoft says it will start a partnership with Beth Israel Deaconess Medical Center in Boston to test the system in real hospital settings, where consenting patients would receive its input alongside standard care and its recommendations would be compared with those of human doctors. TechCrunch also reported that Microsoft is working with the Mayo Clinic on a healthcare foundation model planned for release in 2027. The system is not in clinical use.
The finding to hold onto is narrow: on text-only published cases, a well-built agent panel on a strong model outperformed generalists who were working alone. Whether it helps a busy clinician with a real patient, a messy chart and no neat summary is the question the Boston study is meant to answer.
Published in The Outspoken Digest
Editorial desk
Outspoken Digest Technology DeskSoftware, hardware, artificial intelligence and what they change for everyone else.
Newsletter
The Digest, in your inbox
One edition, sent when it is ready. No noise, and your address is never passed on.
Read Next
More Technology →
NASA's Roman Space Telescope Finishes Aligning Its Instruments, With the Coronagraph's 3,000 Moving Parts Next
Nov 14, 2026/2 min read
Nvidia's Board Approached Arm CEO Rene Haas About Succeeding Huang, as Staff Post Calls for His Return
Nov 14, 2026/2 min read
Starship Flight 15: Booster Diverted to the Gulf, Ship Coasts in Orbit Ahead of Saturday's Tower Catch Attempt
Nov 14, 2026/2 min read
Microsoft's Outage Updates Point to Missing Servers and an Overwhelmed Secure-Connection Process
Nov 13, 2026/2 min read