[Accuracy alone is not enough: cognitive safety of artificial intelligence in medicine.]

Current evaluation of artificial intelligence (AI) in healthcare remains largely focused on model accuracy, clinical outcomes, efficiency, and the formal availability of human oversight. These dimensions are necessary but insufficient. A system may improve today's decision while, through repeated use, weakening the clinician's ability to recognise tomorrow's error. Cognitive safety is proposed here as a longitudinal property of the clinician-AI-organization sociotechnical system: its capacity to support or improve clinical performance without eroding independent hypothesis generation, uncertainty calibration, reasoned dissent, metacognitive control, and resilient performance when AI is wrong or unavailable. Automation and augmentation should not be treated as ideological alternatives, but as task-sensitive regimes selected according to ambiguity, reversibility, normative content, and the need to preserve skill formation. A Clinical Cognitive Impact Assessment is outlined to make this proposal empirically testable across pre-deployment evaluation and post-implementation monitoring. Keeping a physician formally in the loop is not enough: healthcare systems must preserve over time the cognitive capacities required to understand, challenge, and, when necessary, interrupt that loop.