Integration of ChatGPT in medical learning: An analysis of interaction and contradictions

Current research on generative AI in medical education focuses on AI's performance or risks, such as unreliability. We argue these issues are not isolated flaws but are symptoms of systemic contradictions that emerge when a technology is introduced into a learning environment. To move beyond descriptive reports, a theoretical framework is necessary to analyze the systemic tensions that arise during generative AI integration.

A total of 141 first-year clerkship medical students used ChatGPT and provided qualitative data, including conversations with ChatGPT, evaluations of the generative AI's responses, and free-text feedback after watching concept videos of 'Acute Liver Failure'. We employed inductive thematic analysis to identify initial patterns, followed by a deductive analysis using Cultural-Historical Activity Theory to identify and interpret systemic contradictions.

The analysis revealed four contradictions within the activity system: 1) a conflict between the Tool's (ChatGPT's) unreliability and the Object of achieving accurate knowledge; 2) a skills gap between the Subject's (students') initial questioning abilities and the Tool's operational demand; 3) an unstable Division of Labor (student-AI) that conflicted with professional Rules, creating a demand for the need for expert validation; and 4) ambiguous Rules that created confusion and conflicted with professional norms.

Challenges like AI unreliability and skill gaps are contradictions that function as catalysts for expansive learning. Resolving these tensions requires systemic transformation, including formalizing prompt engineering training and redefining the educator's role from an information provider to an essential expert validator within a new collaborative practice.