When Generative <scp>AI</scp> Meets Socratic Method: Investigating Programming Learning Dynamics Through Behaviours, Interaction Qualities and Perceptions

ABSTRACT Background The integration of generative artificial intelligence (GAI) tools like GPT into programming education offers transformative potential through personalised guidance and instant feedback, yet risks fostering overreliance and superficial learning due to their tendency to deliver direct, context‐free answers. Objectives This quasi‐experimental study addresses this gap by proposing a Socratic questioning framework to optimise GAI‐facilitated programming instruction, emphasising critical thinking over passive solution retrieval. Methods We compared two pedagogical approaches: GAI‐Scaffolded Learning (GSL), where GPT employs structured Socratic dialogue to guide problem‐solving and GAI‐Direct Learning (GDL), which provides immediate answers without guided inquiry. This research collected learners' programming behaviours, interactions data with GPT from screen recordings and platform log data and perceptions data. This research further utilised multiple learning analytics approaches (i.e., click stream analysis, lag‐sequential analysis, epistemic network analysis [ENA] and statistics) to compare learners' programming behaviours, interaction patterns and perceptions under two approaches. Results and Conclusions Through an analysis of 80 college students' programming behaviours, interaction qualities and perceptions, we found some intriguing results. First, GSL engaged in cyclical, reflective practices (debugging, Socratic questioning, console use), while GDL prioritised rapid fixes via trial‐and‐error with GPT code, risking superficial mimicry and over‐reliance on external resources. Second, ENA highlighted GSL's deeper engagement through interconnected feedback, emotional support and iterative inquiry, reducing frustration and sustaining persistence and GDL interactions focused on surface‐level queries, lacking scaffolding for emotional/heuristic integration. Third, GSL maintained positive attitudes due to structured prompts aligning expectations and easing cognitive load. GDL attitudes declined from mismatched expectations and frustration. Implications Based on these findings, the study proposes pedagogical and developmental implications for future design and development of AI‐augmented curricula, providing actionable insights for educators seeking to harness GAI's potential while nurturing critical thinking in programming education.