Designing Epistemic Partnerships with Generative AI: A Four-Stage Model of Student Engagement in Scientific Argumentation

Abstract

Abstract This study examined how a design-based learning environment supported first-year undergraduates’ ( N = 31) epistemic engagement with generative AI in a science argumentation course. A two-year design-based research (DBR) study was conducted across two cycles (2024: n = 11; 2025: n = 20). Within a Toulmin-structured argumentation framework, learners interacted with ChatGPT, and their prompts were analysed as epistemic indicators across four stages of the Four-Stage Dialogic Partnership Model (4S-DPM): Epistemic Retrieval, Epistemic Scaffolding, Epistemic Critique, and Epistemic Agency. Comparative analysis revealed a significant shift toward higher-order epistemic engagement (Fisher’s exact test, p < .001): Stage 3–4 prompts increased from 12.5% to 47.0%, with Epistemic Agency prompts rising 3.6-fold (8.3% to 30.0%). Qualitative analysis identified dialogic restructuring, an iterative process in which learners examine and reconstruct their arguments through AI dialogue, as characteristic of Stage 4. Three design elements supported this trajectory: externalisation of argumentative structure, peer dialogue, and understanding of AI’s epistemic characteristics. These findings suggest that generative AI does not automatically function as an epistemic partner; rather, its role depends on the design of the learning environment. The 4S-DPM provides an analytical and pedagogical framework for designing science learning environments in which AI supports epistemic engagement.