The causal effects of artificial intelligence use on metacognition, engagement, and knowledge transfer in educational...

Artificial intelligence applications in learning environments are becoming more common, but there is no clear causal evidence regarding the impact of AI-assisted tutoring on metacognition, engagement, and knowledge transfer.

Using EdNet log data, AI use was operationalized as exposure to tutoring-system assistance events during learning episodes. Fixed effects models with temporal controls within learner and clustered inference were estimated. Monitoring proxies and control proxies were used to assess metacognition, time on task, activity intensity, persistence after errors, return probability, and spacing. Engagement, near transfer, topic-shift transfer, and delayed performance were assessed.

AI exposure increased monitoring behaviors (β = 0.028-0.047) and control behaviors (β = 0.036-0.061). Longer time on task (β = 0.083), higher activity intensity (β = 0.057), higher persistence after errors (β = 0.064), and higher return probability on the next day (β = 0.031) increased engagement, while shorter interval between activities (β = -0.018) decreased it. The strongest transfer gains were for near transfer (β = 0.052), topic-shift transfer (β = 0.021), and delayed performance (7 days) (β = 0.017).

AI-supported tutoring appears to strengthen metacognitive regulation and engagement, while broader transfer remains conditional on scaffolds that promote verification, retrieval, and abstraction.