Generative Artificial Intelligence in Clinical Reasoning Education: A Focused Review and Framework

Clinical reasoning education requires learners to generate and justify diagnostic plans under uncertainty, yet supervised practice and feedback are uneven.Generative artificial intelligence (AI), large language models, and virtual patients may expand practice, but improved assisted performance may not represent durable unaided reasoning.We conducted a focused narrative review of literature identified through PubMed, Scopus, Web of Science, and Google Scholar, supplemented by reference-list review and citation searching.The synthesis covered AI-powered virtual patients, large language model-based diagnostic coaching, and automated assessment or feedback, with emphasis on evidence involving learners and on indirect evidence that clarifies educational mechanisms or risks.Most studies report associations or short-term improvement rather than AI-specific educational effects, and several rely on satisfaction outcomes, pre-post designs, or historical controls.Current evidence therefore supports feasibility and selected short-term gains in practice, diagnostic performance, and feedback but does not establish durable learning.However, model errors, unreliable references, weaker performance on image-based tasks, and limited evidence for retention or transfer constrain high-stakes use.Recent reviews describe broad applications of AI in medical education, but less attention has been paid to when assistance should enter a reasoning task and how to preserve foundational competence.We propose AI-augmented cognitive apprenticeship, a framework organized around four principles: struggle first, AI second; AI for formative practice, humans for summative judgment; transparent AI literacy; and competence-calibrated AI assistance.The framework treats never-skilling, defined as failure to acquire a foundational competence, as a testable educational risk; operationally, neverskilling is failure to reach a predefined unaided competency threshold after AI-supported training.AI should therefore be introduced after an independent attempt, calibrated to demonstrated competence and task risk, and paired with periodic unaided assessment.Prospective longitudinal studies are needed to determine whether AI-supported practice transfers to durable clinical reasoning.