<![CDATA[The integration of generative artificial intelligence (AI) into undergraduate mathematics classrooms has begun to shift the dynamics of the learning ecosystem — reconfiguring how reflective interaction, metacognitive regulation, and cognitive engagement relate to one another within the Community of Inquiry (CoI) framework. Latent-variable structural models have documented that this reconfiguration exists, but they cannot pinpoint which specific learner behaviors carry the shift. This study applies psychometric network analysis to produce a behavior-level map of the reconfigured ecosystem. Using cross-sectional data from 200 undergraduate mathematics education students at Universitas Negeri Makassar who used conversational AI during a differential calculus course, we estimated a regularized partial-correlation network over 19 items from an AI-Integrated CoI survey using the Graphical Lasso with cross-validated penalty selection. Node strength and bridge strength were computed to identify hub and boundary-spanning behaviors, and bootstrap resampling (300 iterations) was used to assess stability. The network retained 87 edges (density = 0.51). The four items with the highest bridge strength were RMP2 (using AI to check one's own solution steps — 0.45), LP3 (monitoring understanding of AI explanations — 0.45), RMP5 (asking follow-up questions to deepen understanding — 0.40), and LP6 (reflecting on whether AI is helping — 0.39). Three of these four top-ranked items describe monitoring and verification behaviors that occur during the AI interaction rather than the post-hoc evaluation of AI outputs most often emphasized in AI-literacy discourse. Bootstrap Spearman correlation between bootstrapped and reference bridge rankings averaged 0.80, indicating moderate stability. The findings reframe the AI-induced shift as a shift toward in-interaction epistemic monitoring rather than a shift toward output-checking vigilance.]]>
