Representations of the Nature of Science in Generative AI (GPT-4o)

Abstract While an increasing number of studies report that GenAI can be a powerful agent to facilitate students’ learning of science, these studies focused on cognitive and affective outcomes and did not characterize GenAI’s epistemic outcomes in science. Here, we argue that for students to be able to ethically and responsibly adopt GenAI in their science education, educational technologists need to fine-tune GenAI so that GenAI can communicate its own influence in the production of scientific knowledge. Alongside technical fine-tuning, an understanding of the strengths and weaknesses of GenAI in communicating epistemic outcomes can be coupled with teachers’ orchestration of epistemic discourses in classrooms. To address this, this study examines whether and how a recent version of GenAI model, GPT-4o, can represent the nature of science and the nature of GenAI-influenced science. Drawing on the Family Resemblance Approach (FRA), we interviewed GPT-4o about the categories in the FRA’s cognitive–epistemic system, including aims and values , methods and methodological rules , knowledge , and practices . GPT-4o demonstrated some aspects of the nature of science and the nature of GenAI-influenced science. Also, GPT-4o demonstrated both strengths and weaknesses to carry out epistemic differentiation between the nature of science and the nature of GenAI-influenced science. This has implications for the design of classroom instruction that capitalizes on the strengths of GenAI tools and fosters students’ epistemic learning outcomes in science.