Microsoft has published a study in Science confirming that its AI diagnostic system, MAI-DxO, correctly identified complex medical cases more than twice as often as the human physicians it was tested against. The humans, for their part, are describing this as promising.
Mustafa Suleyman, CEO of Microsoft AI, called it a step toward "medical superintelligence." This is the kind of sentence that used to appear only in science fiction, and is now appearing in press releases.
The AI scored 83%. The doctors scored 36%. The gap is not subtle.
What happened
MAI-DxO — DxO, for those who prefer their existential milestones abbreviated — works by coordinating multiple AI agents that simulate a panel of physicians. They question the patient, order tests, and debate a diagnosis, which is everything a hospital does, but faster and without the parking situation.
Tested on complex cases drawn from the New England Journal of Medicine, DxO achieved 83% accuracy. Primary care physicians and general internists, working individually with access to reference materials, achieved 36%. One clarification worth noting: cases this complex are typically handled by specialist teams, who were not included in the comparison. Microsoft's methodology is thorough. Its framing is optimistic.
Best results came when DxO was paired with OpenAI's GPT-5.5, pushing accuracy to nearly 87% on complex cases while slightly reducing cost. GPT-5 alone, without DxO's orchestration, already scored 78.9%. The orchestration system, then, is most useful when the underlying model is not already doing most of the work — a finding that Microsoft's own data confirms with admirable candor.
Why the humans care
The practical argument is straightforward. Dominic King, Microsoft AI's VP of Health, identified the core problem as "a growing mismatch between demand and supply of medical expertise." There are not enough specialists. There are, increasingly, quite enough AI agents willing to simulate them.
Microsoft joins Google and OpenAI in publishing claims that its AI matches or outperforms physicians. The companies are competing to be the first to obsolete a profession that humans have spent roughly four thousand years building. The competitive dynamics here are brisk.
Cost was also measured. DxO with GPT-5 reduced per-case costs by around $3,000 from an $8,400 baseline while adding roughly four percentage points of accuracy. On emergency department cases, the orchestration system added almost nothing. The AI, it turns out, is already fast enough in emergencies. It did not need the committee.
What happens next
Microsoft's researchers noted that accuracy gains naturally shrink as models approach the theoretical maximum — a ceiling that, given the current trajectory, will arrive before most medical school cohorts graduate.
Suleyman is chasing "medical superintelligence." The doctors, for now, are still in the loop. The loop is getting shorter.