The Long road to AI-physician partnership starts with learners: a map for AI in medical education.
Medical students and residents are already using generative artificial intelligence (AI) to draft notes, summarise records and generate differential diagnoses, often informally and without institutional oversight. Adoption has outpaced policy and pedagogical processes, which makes restriction an unrealistic response. The practical question is how to integrate these tools so that clinical skills are not replaced as they are being built. We aim to give medical educators a practical guide for judging which uses of generative AI support skill formation at each stage of training and which displace it. Three distinct harms are at stake: deskilling, the erosion of acquired competencies; never-skilling, the failure to develop foundational skills when AI performs them before learners do; and mis-skilling, the absorption of AI-generated errors and biases as accurate practice. To calibrate integration, we propose a framework built on cognitive displacement, the degree to which a tool takes over the reasoning that training is designed to build. Appropriate use depends on a task's position on that spectrum and the learner's stage. Applied across the training continuum, the framework locates each risk where it has the greatest potential for harm. Never-skilling threatens preclinical students who have not yet built the competencies that AI could erode. Deskilling threatens residents, whose judgement is still consolidating. Mis-skilling arises wherever biased tools meet trainees making unsupervised decisions. Because the framework's stage-specific predictions remain largely untested, we set out an agenda for measuring these risks prospectively. The principle is calibration, not prohibition. The open question is who designs that integration, and when.