LLMs in Problem-Based Learning: The Grounding Issue

Large language models (LLMs) are increasingly used in medical education, including in problem-based learning (PBL). Their ability to summarize cases, generate differential diagnoses, and structure discussion raises the question of whether they might eventually replace PBL rather than merely support it. This paper offers a conceptual analysis of that question, using a renal-tumor PBL case derived from recent ChatGPT-assisted teaching research as an illustrative use case. It distinguishes among semantic support, representational uncertainty, and situated uncertainty. LLMs can contribute substantially at the first two levels: they can reformulate cases, identify missing information, keep multiple diagnostic possibilities in play, and suggest coherent next steps. What they do not reproduce is the situated uncertainty through which PBL forms judgment in interaction with peers and tutors. More tools, retrieval, modalities, or sensor input may enrich representation, but they do not by themselves close this gap. The paper therefore argues that LLMs can augment PBL in important ways but do not replace its role in the formation of clinical judgment.