Abstract
<title>Abstract</title> <p>Artificial Intelligence (AI) tutoring in higher education has become an important part of the recent educational practices. This study introduces the Cognitive Autonomy Index (CAI), a theory-driven metric for characterizing student-AI interactions. The CAI captures the balance between behaviors that promote autonomy, such as asking conceptual questions, and those that undermine autonomy, such as requesting full solutions by AI tools. We analyzed longitudinal dialogue data from 110 undergraduates in an AI course to examine whether the CAI is associated with students’ performance on examinations completed without AI assistance and whether these relationships vary across different phases of the course. Analyses reveal that aggregating autonomy-related behaviors into a single score obscures critical phase-specific dynamics. Overall AI usage volume showed only weak associations with examination performance. However, Principal Component Analysis (PCA)-derived behavioral components demonstrated stronger associations with performance, indicating that how students use AI tools matters more than how much they use them. Early-phase conceptual questioning was positively associated with examination performance, while late-phase context-dumping (pasting assignment material) was negatively associated with performance. Writing requests, classified as autonomy-undermining in CAI, was positively associated with early performance, revealing that the same behavior has different meanings at different times.</p>