Calibrating Generative AI Use for Sustainable Higher Education: Cross-Sectional Evidence on Perceived Capability and Dependence

Generative artificial intelligence (GenAI) can support educational work while changing how users engage in independent generation and evaluation. This study examines whether the GenAI task use frequency has different associations with perceived role-specific capability and AI dependence. A cross-sectional survey of 1105 respondents (659 students and 446 teachers) from five Beijing universities measured all focal variables by self-report. Hierarchical regressions with HC3 standard errors were complemented by formal inverted U tests, bootstrap intervals, measurement-model sensitivity analyses, and nested out-of-fold prediction. Both capability outcomes showed positive lower-range and negative upper-range slopes, with the maxima located within the observed use range and supported by observations on both sides within each literacy stratum. Between the 25th and 75th percentiles of AI literacy, the model-implied maximum shifted by 0.715 use scale units for students (95% bootstrap CI [0.525, 0.973]) and 0.643 units for teachers (95% CI [0.450, 0.883]). The primary literacy measure was a 12-item mean; four-dimensional and second-order CFA specifications supported its interpretation as a broad composite. Separate exploratory dimension models yielded positive literacy-by-use interactions, with the largest point estimate for critical evaluation. Dependence was positively associated with use, habit, and trust and negatively associated with literacy. The combined squared use and interaction increments for dependence were 0.0024 and 0.0004, with 90% interval upper bounds of 0.0058 and 0.0031, below an exploratory 0.01 R2 benchmark. Belief-centred and trust omission specifications supported the broader reliance pattern, although trust and dependence remained conceptually close. Nonlinear ridge recovered most of the capability prediction improvement over linear ridge. The findings describe between-person associations, not individual optimal use levels or intervention effects. They motivate the evaluation of literacy-oriented task design and complementary approaches to trust calibration and habitual reliance.