From Code Generation to Code Auditing: Constrained AI Scaffolding, Transfer, and Cognitive Efficiency in Debugging

Abstract

Generative artificial intelligence is reshaping programming education, yet its effects on skill development depend partly on how learners interact with artificial intelligence-supported systems. This study introduces the Artificial Intelligence-Scaffolding Interaction Framework, which conceptualizes constrained, question-driven artificial intelligence as a cognitive scaffold rather than a direct solution provider. A quasi-experimental study with 54 novice programmers compared Socratic artificial intelligence scaffolding with traditional Web-based information retrieval during debugging activities. Students in the Socratic condition obtained higher immediate debugging-transfer scores, corresponding to a moderate effect size (Hedges’ $g=0.45$ ), although the primary between-group difference did not reach conventional statistical significance ( $p=.100$ ). The largest item-level difference was observed for mental code tracing. Perceived-workload profiles also showed lower Effort in the Socratic condition; however, this difference did not remain statistically significant after correction for multiple comparisons. Exploratory moderation analyses detected no significant interaction with prior academic achievement or artificial intelligence familiarity, a pattern consistent with, but not sufficient to establish, the proposed Equalizer Hypothesis. Post-task reflections in the Socratic condition were longer and contained more reasoning-oriented language, whereas trial-and-error references were more common in the search condition. Together, the findings suggest constrained Socratic scaffolding as a promising instructional design for debugging education while highlighting the importance of distinguishing observed learning outcomes from the cognitive mechanisms proposed to explain them.