ChatGPT as cognitive scaffold: computational thinking and metacognition in AI-embedded instruction

This is the authors' abstract. We don't add key points for this paper.

Drawing on a computational thinking model and cognitive offloading theory, this exploratory mixed-methods study examined whether structured AI scaffolding is associated with differential computational thinking (CT) outcomes and explored the role of metacognitive self-regulation. Two naturally occurring sections (n = 47) of the same liberal arts course, taught by the same instructor in consecutive Fall semesters, one embedding ChatGPT and one not, were compared using independent-samples t-tests, multiple regression, and inductive content analysis. Students in the ChatGPT-embedded section reported significantly higher problem-solving scores on a validated CT instrument. Digital literacy scores did not differ, possibly reflecting measurement limitations or the intervention's focus on reasoning over general digital fluency. Problem solving correlated significantly with metacognitive strategy use and contributed to a significant overall regression model, but did not remain a unique predictor after adjusting for digital literacy and class membership. Qualitative analysis identified three themes: ChatGPT as a cognitive scaffold, ChatGPT as a cognitive risk, and student-initiated boundary-setting. Most students expressed all three simultaneously, suggesting structured AI-embedded environments may foster metacognitive vigilance alongside CT gains. Findings support deliberate instructional design and explicit metacognitive supports when integrating ChatGPT in higher education.