Teacher-regulated generative AI in structured group tasks: independent performance and metacognitive judgements in higher education

Introduction As generative AI becomes increasingly embedded in higher education classrooms, an important question is whether it shapes not only students' task performance but also how they judge their own learning. This study examined whether participation in AI-supported structured group tasks was associated with differences in university students' independent post-task performance, performance predictions, confidence judgements, and cognitive calibration compared with non-AI-supported structured group tasks. Methods A group-randomized, post-test-only classroom design was implemented in two undergraduate classes using Rain Classroom, a digital instructional platform. Students worked in pre-existing groups, and instructional condition was randomly assigned at the group level. In the AI-supported condition, students were permitted to use the platform's built-in AI learning companion during structured group discussion; in the non-AI-supported condition, they were instructed not to use it. After previewing the post-task assessment items, students completed a task-specific judgement form before responding to the assessment itself. The final analytic sample consisted of 126 students. Results Linear mixed models accounting for class and group-level clustering indicated significantly higher independent post-task performance in the AI-supported condition, B = 0.506, SE = 0.188, p = 0.013, 95% CI [0.117, 0.894], with a small descriptive standardized mean difference ( d = 0.196). No statistically significant between-condition differences were observed in predicted performance, confidence judgements, calibration bias, or absolute calibration error. Discussion These findings indicate a modest between-condition difference in post-task performance without parallel statistically detectable differences in metacognitive monitoring.