Medicine Should Seriously Consider Autonomous AI Doctors

Summary

This opinion article by Ezekiel J. Emanuel and Vinod Khosla argues that medicine should test whether specialized AI systems can autonomously control some clinical decisions, rather than reserving AI for a supporting role. The authors say much routine care is algorithmic, while physician error contributes to substantial misdiagnosis and preventable deaths; a medical model with domain training and guardrails could check drug interactions and follow up with patients more consistently. They cite a study in which Google’s Articulate Medical Intelligence Explorer was judged better than physicians at eliciting information from patient actors, and a randomized type 2 diabetes study in which autonomous AI reached stable insulin doses faster and was associated with more reliable medication use and less distress. Evidence for physician-AI hybrids is less decisive: a 2025 review of 52 studies found that hybrids generally did not outperform medical AI alone or the best clinicians, although some teams did better when AI controlled the final decision. The American College of Physicians accepts that fully autonomous AI is possible but favors rare use in low-risk cases with clinician access, while the AMA remains opposed to AI taking the final role. The authors acknowledge that some reviewed studies favored humans, relied on simulations or older models, and lacked long-term real-world outcomes. They also cite surveys and reviews suggesting patients can perceive AI as more empathetic than doctors. They conclude that researchers, regulators, and medical educators should investigate autonomous care rather than dismiss it, while recognizing that AI will remain imperfect and require safeguards.