JAMA: don't lock a doctor into the AI loop
A JAMA Perspective published Monday asks regulators not to hard-code a physician veto into medical AI — just as the evidence for that future is still mostly simulated.
Ezekiel Emanuel and three coauthors argue in JAMA that autonomous AI will soon outperform any doctor-plus-model team at medical reasoning, and that writing a human into the loop now would lock in worse care. The piece, dated August 17, is titled "Will Autonomous AI Exceed AI-Aided Physicians as the Best Medical Care?" Emanuel is a University of Pennsylvania bioethicist and an architect of the Affordable Care Act. His coauthors are colleague Abe Baker-Butler, Curai Health CEO Neal Khosla, and Vinod Khosla of Khosla Ventures — an OpenAI investor, and Neal's father. Two of the four names on the byline run or fund the industry the paper is asking medicine to get out of the way of.
The Decoder's recap of the Perspective walks through the evidence the authors treat as settled. In studies since 2024, they say, AI alone matches or beats doctors at the five core reasoning jobs of medicine: history, diagnosis, test choice, guideline-based treatment, and chronic-disease management. Google's AMIE scored higher than primary-care doctors in simulated visits. Across 377 complex cases, ChatGPT o3 named the correct diagnosis first 60 percent of the time, against 15.9 percent for 20 internists. Microsoft's diagnostic orchestrator found the right answer about four times as often as doctors, at lower cost. A 2024 JAMA Network Open study they lean on found GPT-4 alone scored 92 percent on diagnostic reasoning; doctors given the same model scored 76 percent. Contrary findings, the authors say, are outdated or methodologically weak.
The load-bearing claim is not that AI is good at quizzes. It is that once the machine is clearly ahead, the doctor who checks its work stops being a safety net and starts being a source of error. That is not just a forecast. A 2024 Nature Human Behaviour meta-analysis of 106 experiments, which the JAMA authors cite, found that human-AI teams on average did worse than the better of the two working alone. When the human was stronger, the combo helped. When the AI was stronger, people overrode it in the wrong places and made the result worse. The authors add a second, uglier mechanism: doctors who use AI lose their own skills. A 2025 Lancet study of colonoscopies, also in their reference list, is the exhibit.
The policy ask follows. Physician groups still say AI should augment doctors, never replace them. Robert Wachter has called AI-only care the "economy class" of medicine. The JAMA authors want regulators to stop writing that hierarchy into the rules. By 2030, they expect autonomous systems to be ready for some — maybe many — cognitive workflows. Liability, payment, training, and FDA-style sign-off all need a rethink now, they argue, because a mandated physician veto will be hard to unwind once it is in the statute book. They concede the caveats: almost all the evidence is simulated single tasks, not real patient care; the handoff between human and model is a known weak spot; surgery, childbirth, and other physical work stay with humans; and autonomous systems fail in ways doctors don't — hallucinations, outages, cyberattacks.
The bottom-up version of this argument is already in clinics. Families are bringing models to appointments that doctors cannot close — we covered that last weekend in Patients turn to AI to solve rare-disease mysteries. The JAMA piece is trying to turn that messy adoption into a regulatory principle. The honest read is that the Vaccaro finding is real and uncomfortable, and that a Perspective coauthored by the Khoslas is a very convenient place to convert it into "please don't require a doctor." Simulation superiority is not a license. It is a reason to run the trials the authors admit we do not have.
What to watch: whether FDA, CMS, or state medical boards treat this as a serious bid to loosen human-in-the-loop rules — or as a well-placed op-ed from the people who would collect the upside.
If the model is already more accurate than the doctor checking it, who should be allowed to overrule whom? Tell us in the comments.
Sources: JAMA · The Decoder · Vaccaro et al., Nature Human Behaviour · Google Research — AMIE
