The rise of AI in healthcare offers major potential for diagnosis and learning but also exposes to a worrying risk of "deskilling." Recent data from colonoscopy shows that prolonged exposure to assistance systems reduces the performance of endoscopists illustrating a phenomenon of "deskilling" that is potentially harmful to patient safety. Clinical reasoning, based on the integration of data, uncertainty, context, and patient values, cannot be reduced to a simple exercise in prediction. Structured and supervised educational integration of AI (simulation, feedback, formative assessment, prompting learning) would preserve the cognitive autonomy and critical thinking of future clinicians by making AI an ally rather than a substitute.