To the Editor: Making Clinical Reasoning Visible During Resident Use of Artificial Intelligence

In his recent AI Teaching Rounds article, Preiksaitis identifies a problem likely to define AI-enabled residency training: clinical work may appear more polished while the learner’s underlying reasoning becomes less visible.¹ This marks an educational shift. For generations, supervisors have inferred thinking from oral case presentations, written notes, differential diagnoses, and management plans. In AI-enabled settings, however, these products may no longer reliably reflect the learner’s own cognitive work. We therefore agree with the author’s emphasis on protecting “human-first” cognitive tasks such as problem representation, first-pass differential diagnosis, and the initial assessment and plan.¹ Frameworks such as DEFT-AI (Diagnosis, Evidence, Feedback, Teaching, and Recommendation for AI Use)² are valuable because they redirect faculty attention from the apparent quality of AI-shaped output to the learner’s underlying thinking.

However, an important difficulty remains. Supervision can detect that AI was used and probe what the trainee did in the moment, but it may still fail to capture how AI altered reasoning, narrowed inquiry, promoted premature closure, or displaced the effort through which clinical judgment develops.² The educational risk is therefore not limited to inappropriate outsourcing of tasks; it also includes erosion of metacognitive discipline. A trainee may appear efficient and accurate while becoming less practiced in generating, interrogating, and revising independent clinical interpretations. If this recurs, AI may not simply assist reasoning but begin to shape it in ways that remain insufficiently visible to faculty. This concern is increasingly relevant as medical education moves from debating whether AI should be present to determining how its use will be supervised, entrusted, and assessed.³

This is where an additional educational layer is needed. If programs wish to preserve reasoning rather than merely regulate tools, residents need a structured way to examine their cognition after AI interaction: what they thought before using AI, what the system contributed, what they accepted or rejected, and why. Without such reflection, disclosure risks becoming procedural while cognitive offloading becomes normalized. This issue is especially relevant in competency-based training, where the goal is progressive development of transferable judgment under conditions of uncertainty.⁴

One possible response is a reflective scaffold. We propose CLARITY (Clinical context, Logical clinical reasoning, AI integration, Reflective appraisal, Integrity safeguards, Transformative insight, and Yield to practice), a framework we are developing to make AI-mediated reasoning visible after the clinical encounter. Its value is not that it introduces another framework, but that it generates an assessable record of pre-AI reasoning, AI input, critical appraisal, and reflective synthesis. Such a scaffold could help faculty determine whether AI supported learning, merely improved presentation, or inadvertently replaced the cognitive struggle through which expertise is built. Used in this way, CLARITY would not compete with DEFT-AI; it would extend it. Structured reflection may also restore alignment between supervision and educational theory.

If AI increasingly enters the space between encounter and interpretation, reflective appraisal of AI use should become part of that developmental process rather than remain an optional add-on.⁵ Preiksaitis has advanced the conversation by asking what should remain human work. The next step may be to ask what form of reflection is needed to show that such work has, in fact, occurred.

Author Notes

Disclosure: AI was not used in writing this article.