Which Human-in-the-Loop? Why Context, Culture, and Health Systems Matter

AbstractArtificial intelligence-based clinical decision-support systems are increasingly integrated into routine patient care, yet broad consensus holds that they should not operate autonomously. The "human-in-the-loop" principle is widely endorsed but poorly defined. This article argues that human oversight in medical AI comprises three distinct and interdependent domains: the clinical loop at the point of care; the governance loop governing approval and deployment; and the learning loop overseeing monitoring and model updates after deployment. Oversight failures in any one of these domains undermine safety, accountability, and alignment with societal values. Drawing on international experience, we show that human-in-the-loop practices are shaped by health system structures, reimbursement models, regulatory frameworks, and cultural priorities, and, therefore, cannot be universally standardized. AI systems developed in the United States may embed assumptions about resources, incentives, and acceptable interventions that do not translate to other settings. Effective oversight requires treating human-in-the-loop as a design specification that can be deliberately adapted to local contexts.