In JAMA, two authors ask regulators not to impose a doctor in the loop

Ezekiel Emanuel and Neal Khosla argue that autonomous AI will eventually surpass any doctor-machine tandem, and that regulations should not freeze the requirement for human oversight.

Published on August 18 in JAMA, the forum signed by bioethicist Ezekiel Emanuel (University of Pennsylvania) and Neal Khosla, CEO of Curai Health, presents an unusual thesis: from a certain level of performance, a diagnostic AI delivered to itself would do better than a combination of a doctor and the same AI. The authors draw a regulatory recommendation — not to include in the texts the obligation of a human in the loop.

Their demonstration is based on a series of recent results: Google's AMIE conversational system is ahead of general practitioners on almost all the criteria assessed; of 377 complex cases, ChatGPT o3 places the right diagnosis at the top in 60% of cases, compared with 15.9% for 20 internists; A Microsoft diagnostic tool finds the right answer about four times more often than practitioners, at a lower cost. One figure often returns in the debate: GPT-4 alone gets 92% on a diagnostic reasoning test, compared to 76% for doctors using the same model.

The authors themselves recognize the limitations of their records: these results come from simulations, not from actual patient care, technical gestures remain out of reach, and autonomous systems are exposed to failures — hallucinations, breakdowns, cyberattacks — to which a doctor is not subjected. Interests are also at stake: Neal Khosla runs a digital health company, and his father Vinod Khosla has invested in OpenAI as well as Curai Health. The American Medical Association and the American College of Physicians defend an AI that assists the doctor without replacing him.