Scaffolding Agency in STEM: The Effects of AI-Supported Self-Regulated Assessment on Performance and Self-Efficacy Wai Kit, Mak

Abstract

This study investigates the impact of varying levels of selection autonomy on STEM performance and self-efficacy among 126 junior secondary students. Utilizing a three-phase repeated measures design, students transitioned from teacher-led assessment to AI-assisted choice via a role-play mentor, and finally to unguided self-selection. Results from MANOVA and paired t-tests reveal a significant stepwise decline in academic performance as autonomy increased, uncovering an "Autonomy Paradox" where high-achievers under structured conditions struggled most in autonomous settings. However, STEM self-efficacy remained stable across all phases, suggesting that the agency provided by AI and self-selection acts as a psychological buffer. These findings highlight the critical role of AI as a cognitive scaffold in the Zone of Proximal Development, facilitating a shared-regulation model that balances learner agency with academic achievement.