Promise, Mimicry, and Surveillance: Responsibly Integrating Artificial Intelligence With Socratic Inquiry in Medical Education.

Large language models are increasingly proposed as a way to deliver Socratic dialogue at scale in medical education, offering personalized, always-available, and lower-stakes inquiry that human educators cannot logistically provide. In this viewpoint, I argue that this promise is plausible but unproven and that 2 failure modes deserve more attention than they currently receive. The first is the mimicry trap: a system that fluently generates Socratic-sounding questions can appear to cultivate reasoning while functioning as interactive content delivery. I position this as an educational instance of the ELIZA effect and of the proxy-outcome problem; separating conversational mimicry from genuine metacognitive gain remains a central unresolved evaluation problem. The second is what I term the Panopticon Paradox: the data collection that makes AI tutoring effective may erode the psychological safety that Socratic inquiry requires, pushing learners toward performative rather than authentic engagement, in a manner analogous to gaming behaviors documented in intelligent tutoring systems. I present this second construct as a testable causal model with specified mediators, moderators, and falsifiable predictions rather than as an established finding. Because the evidence base is dominated by proof-of-concept tools, cross-sectional surveys, and short-term evaluations, I argue that AI is complementary rather than a replacement technology, and I propose a 3-pillar framework (governance, curriculum, and faculty development) tied explicitly to the 2 failure modes. I distinguish 4 modalities of delegation and argue that ethical deliberation, emotionally complex communication, and ambiguous clinical judgment cannot presently be recommended for autonomous or primary AI delivery.