Epistemic framing in AI-mediated university physics problem solving: a qualitative multiple-case study

Research on large language model (LLM)-based generative artificial intelligence (GenAI) in physics education has examined model performance, students’ evaluations of generated responses, and learning outcomes. Less is known about how generated explanations and calculations enter students’ observable physics problem-solving activity. This qualitative multiple-case study examines epistemic-frame sequences and their relation to solution quality, alongside students’ interactions with ChatGPT and their subsequent use of generated contributions. It analyses 30 submissions from 73 students across five university courses: 22 using ChatGPT and eight without it. ChatGPT was selected because its free version was students’ most frequently reported GenAI tool for independent study. Written artefacts and interaction records were analysed using Nguyen et al.’s problem framing framework, with solution quality assessed independently. Activity recurrently moved between conceptual and algorithmic physics, but frame-sequence length did not distinguish solution quality. Challenges and reformulations made students’ evaluation of ChatGPT responses visible, whereas incorporation without documented evaluation sometimes carried generated errors into final submissions. The heterogeneous dataset confines interpretation to qualitative patterns and precludes causal comparison. The findings identify students’ observable evaluation and use of generated responses as a key link between GenAI output and documented physics reasoning.