Beyond access: guided versus free ChatGPT use and learning outcomes in higher education

Research on artificial intelligence in higher education has largely examined usage patterns rather than the cognitive dimensions of student-AI interaction. This mixed-methods experimental study investigated how undergraduates’ epistemic engagement with ChatGPT during academic reasoning tasks relates to learning outcomes. Grounded in constructivist learning theory, the study randomly assigned 120 undergraduates to three conditions: No-AI control, Free-AI (unstructured access), and Guided-AI (process-focused scaffolding). Students completed epistemic reasoning tasks requiring conceptual understanding and judgment under ambiguity. Analysis identified four interaction profiles: Delegative (AI-led), Instrumental (procedural), Co-Constructive (shared control), and Dialogic-Critical (student-led critique). Task quality varied by interaction pattern rather than AI access alone; Delegative users showed no measurable advantage over No-AI controls despite having AI access. A minimal intervention, a single-page handout emphasizing questioning strategies, substantially increased dialogic engagement (d = 2.29) and critical evaluation (d = 2.19). Because profiles were observed rather than assigned, profile-outcome relationships are reported as associations. Findings challenge binary debates over AI prohibition versus permission, indicating that educational value is associated with students’ epistemic stance during interaction. Pedagogical scaffolding, rather than surveillance or restriction, may be the most productive institutional response to generative AI.