Psychometric network analysis of AI constructs related to students' academic motivation and basic psychological needs in Higher Education

Abstract The accelerating integration of generative AI into higher education renews questions about how students’ motives, needs, meaning, and redundancy beliefs interrelate. This study modeled the conditional architecture of these constructs using a regularized partial-correlation network in a multi-institution undergraduate sample (N = 904). Nodes were AIM–N composites spanning redundancy (Task-, Motivational-), general motivation (Intrinsic, Identified, Controlled, Amotivation), AI-specific motivation (AI-Intrinsic, AI-Identified, AI-Controlled, AI-Amotivation), academic meaning (Presence, Search), and basic psychological needs (Satisfaction, Frustration). A graphical LASSO network was estimated and evaluated via nonparametric bootstrapping (edge accuracy, centrality stability), community detection (modularity optimization), and permutation-based network comparison tests (NCT) across STEM vs. non-STEM and High- vs. Low-AI use subgroups.The whole-sample network resolved into three coherent communities: (i) a self-determined/meaning block (Intrinsic, Identified, Meaning-Presence, Need-Satisfaction, with Meaning-Search nearby); (ii) a frustration/amotivation block (Need-Frustration, Amotivation, AI-Controlled, AI-Amotivation); and (iii) a compact redundancy block (Task- and Motivational-Redundancy). Intrinsic exhibited the highest strength centrality, followed by Need-Frustration and AI-Amotivation. Among the strongest partial edges were Task–Motivational Redundancy, Intrinsic–Need-Satisfaction, Need-Frustration–Amotivation, and AI-Controlled–AI-Amotivation; bridges were sparse and theory-consistent (e.g., links from Need-Frustration and AI-amotivation/controlled toward the autonomous cluster, and selective ties from redundancy to the amotivation/controlled cluster). Bootstrap diagnostics supported stable core edges and interpretable centrality ranks.Across subgroups, global strength was comparable. For STEM vs. non-STEM, NCT indicated no difference in global strength and a modest structural difference; for High- vs. Low-AI use, no reliable differences were detected in either global strength or overall structure. Taken together, the network view highlights actionable targets: reinforcing autonomous motivation and need satisfaction while reducing redundancy beliefs and need frustration may yield the largest downstream gains—dampening controlled motives and amotivation and preserving meaning in AI-rich learning contexts.