A JAMA opinion piece challenges the regulatory assumption that human physicians must retain final decision-making authority over AI systems in medicine. The authors argue that autonomous AI will soon surpass any doctor-AI collaboration at medical reasoning tasks, making mandatory human oversight counterproductive.
The piece questions a common regulatory instinct to preserve physician veto power as a safety mechanism. Instead, the authors contend that once AI systems demonstrate superior diagnostic and treatment recommendations, requiring doctors to review and approve every decision introduces friction without improving outcomes. They suggest regulators should evaluate AI systems on performance metrics, not on whether humans participate in the loop.
The argument rests on a specific claim: AI performance in medical decision-making will reach a point where adding human judgment degrades overall quality. This reflects emerging patterns in other domains. In radiology, some AI systems already match or exceed radiologist accuracy on specific tasks. Similar performance gaps appear in pathology and diagnostic screening.
However, the authors acknowledge a critical limitation. Most evidence supporting their position comes from simulation studies and controlled datasets, not real-world patient care. No study yet demonstrates that fully autonomous AI makes better decisions than human-AI teams in actual clinical practice with all its complexity, variability, and edge cases.
The piece raises a genuine tension in medical AI regulation. Mandating human involvement slows decision-making and may waste resources if AI truly outperforms physicians. But removing humans entirely from high-stakes medical decisions carries reputational and accountability risks that extend beyond pure performance metrics.
The practical path likely involves context-specific approaches rather than blanket rules. Emergency triage might tolerate more autonomy than rare disease diagnosis. Routine screening differs from complex treatment planning. Regulators must also address liability questions: who bears responsibility when autonomous AI errs, and how does that differ from shared decision-making?
The JAMA piece usefully challenges automatic assumptions about human oversight. But the gap between