From “usefulness” to “dependence”: the dynamic evolution mechanism of generative AI use behaviors

Background As generative artificial intelligence (GenAI) rapidly permeates Chinese higher education, university students’ AI use behaviors have come to be characterized by “high frequency, wide penetration, and rapid iteration.” Yet, a systematic understanding of the dynamic evolution mechanism from initial cognition to deep dependence remains lacking. Methods Employing a four-wave longitudinal design spanning 18 months (the first week of June 2024 to the first week of December 2025, with one wave every 6 months) and drawing on 524 valid responses to the 600 distributed questionnaires, retained in full across all four waves, this study systematically investigates the dynamic trajectories and mediating mechanisms of 12 core constructs, including perceived usefulness, trust, technology concerns, ethical concerns, use behavior, technology integration, usage habits, academic dependence, and moral disengagement. Results Repeated-measures ANOVA and mediation model testing reveal the following: (1) Perceived usefulness exhibits an asymmetric U-shaped trajectory of “sharp initial decline, stabilization, and rebound,” which does not support the hypothesized inverted U-shaped pattern. (2) Trust shows a trajectory of “adjusted decline, low-level stabilization, and internalized overshooting,” with an “overshooting” phenomenon in the later stage that surpasses the initial level. (3) Technology concerns decrease in a strictly monotonic fashion, whereas ethical concerns follow an asymmetric U-shaped trajectory of “precipitous decline, stabilization, and sharp rebound,” revealing a differentiated pattern of technological desensitization and ethical re-sensitization. (4) Technology integration partially mediates the transformation from actual use behavior to usage habits (indirect effect proportion = 23.39%), with the cognitive pathway accounting for only about one-quarter of the total association. (5) Early trust significantly and positively predicts continuance intention 18 months later (β = 0.400), and this effect is partially mediated by user satisfaction. (6) Exploratory analyses show that perceived usefulness negatively predicts academic dependence and moral disengagement through the sequential mediation of “actual use behavior → technology integration → usage habits,” suggesting that “high usefulness” may lead to benign autonomous use rather than problematic dependence. Conclusion This study is among the first to adopt a longitudinal design to reveal the nonlinear dynamic evolution patterns of GenAI use behaviors among Chinese university students. It provides crucial empirical evidence for extending technology acceptance theories along the temporal dimension, for understanding the stratified evolution of risk perception, and for informing AI education interventions. The sample was recruited by convenience sampling and consists almost entirely of female students in the humanities and social sciences; the findings describe this sample, and no claim of generalization to all Chinese university students is made.