Microsoft's diagnostic AI outscores generalist doctors, and the job it absorbs is the first-pass judgement

Gizmodo reported on Friday that Microsoft's medical AI system, the Microsoft AI Diagnostic Orchestrator (MAI-DxO), diagnosed difficult cases more than twice as accurately as practising generalist physicians. The study, "Sequential Diagnosis with Language Models", first appeared last year as a preprint on arXiv. It is now published in the peer-reviewed journal Science. On complex cases published in The New England Journal of Medicine, DxO reached 83.2% diagnostic accuracy. Generalist physicians with access to reference materials reached 36.1%. Microsoft's AI health lead, Dr Dominic King, told The Verge that testing will begin at Beth Israel Deaconess Medical Center in Boston within the next couple of months.

What the headline leaves out

The design matters more than the score. DxO works sequentially. It starts with a chief complaint, then decides which follow-up questions to ask and which tests to order before it reaches a diagnosis. According to the paper, the system "simulates a panel of physicians". Separate AI agents propose diagnoses, select tests, challenge assumptions and weigh costs.

That is a team's reasoning process, split into roles and run by software. This is how AI absorbs work. A title stays intact while one of its tasks moves to the machine. Here, the task is the generalist's first pass at a hard case.

The source's own evidence sets the limits. The comparison was against generalists, and the researchers wrote that they "would expect the performance gap to narrow" against expert specialists. In a separate test on emergency department admissions, the margin was narrower: 80.3% to 46.8%. Microsoft says DxO reaches the right diagnosis "at half or even one third of the cost". Gizmodo notes that those savings are estimates drawn from simulated cases. Microsoft's AI chief Mustafa Suleyman told The Verge that real-world deployment would still require clinical testing.

Gizmodo also flags the question nobody can delegate. Widespread adoption would raise legal questions about accountability when an AI-assisted diagnosis goes wrong. The system can produce the answer. A named person still has to sign for it.

The number to watch is the one that fell

Two details in the report matter more to an operator than 83.2%.

The first: when physicians had no reference materials, their accuracy dropped to roughly 22%. Tools were already doing part of the work. The question was always who directs them.

The second: a 2025 study found that doctors who relied on AI to identify potential cancer risks during colonoscopies performed worse than before once the technology was taken away. The tool improved results while it was running and weakened the skill underneath.

That is the risk for any capable professional who adopts these systems carelessly. AI does the routine work. You do the thinking. Hand over the thinking as well and you lose the capability that makes your sign-off worth anything.

Your Next Move

1. Find your sequential diagnosis. Pick the parts of your work that follow the same loop: gather information, form a hypothesis, test it, revise it. Orchestrated agents are built for those tasks. To work out what that means for your own role, use the step-by-step method in how to find out whether AI will take your job.

2. Run your own panel, and keep the first draft. Copy DxO's structure for your next important decision. Give one AI prompt the job of proposing, one of challenging your assumptions and one of weighing cost. Write your own answer before you read any of theirs. That habit guards against the decay the colonoscopy study recorded.

3. Move toward the signature. Accountability is the part Gizmodo identifies as unresolved, and it is the part that stays with a person. Take on work where your name goes on the decision: review, approval and the specialist depth the researchers expect to narrow the gap.