Taylor and colleagues' 1 observational study of preclinical case based learning shows that ChatGPT is already embedded in students' workflow, often when groups hit an impasse. The key risk is not use but unexamined incorporation. In group settings, fluent output can be mistaken for authority; errors can be copied into shared notes and reinforced through repetition. This epistemic drift is accelerated by time pressure, diffusion of responsibility and cohesion: If no one can confidently adjudicate a claim, acceptance becomes the fastest route forward. Prohibition is therefore both unrealistic and inequitable, pushing use underground and advantaging students with better access and digital fluency. Curricula should instead be engineered for epistemic safety. First, require claim provenance in workbooks: Any AI informed statement should link to a primary source (teaching material, guideline, or paper) and include a brief uncertainty note (‘what evidence would overturn this?’). Second, introduce a rotating red team role that challenges at least one AI supported claim per case and records the verification step. Third, train tutors to assess process by routinely asking, ‘How did you verify that?’, and by modelling quick triangulation strategies. Fourth, run low stakes audits of group notes to surface recurring error types and effective corrections, with feedback focused on learning rather than penalty. Finally, align assessment with these behaviours by rewarding audit trails, justified revisions and transparent uncertainty alongside clinical reasoning. Challenges include added workload, variable tutor confidence, unequal model access and performative tagging; lightweight templates, institutional licences and formative feedback can mitigate these. Waseem Jerjes: conceptualization, investigation, writing – original draft, data curation, resources, methodology, visualization, writing – review and editing, validation. The author has nothing to report. The author declares no conflicts of interest. Data sharing is not applicable to this article as no datasets were generated or analysed during the current study.
Waseem Jerjes (Sat,) studied this question.