On this page

Synthesis: Su, Zhang, Chen, Wang, He and Zhang (2026) surveyed 1,461 Chinese adolescents in Grades 7, 8, 10 and 11 and modeled generative-AI dependence against school examination performance using structural equation modeling. Dependence correlated negatively with attentional control (r = −0.306), physical fitness (r = −0.124) and academic performance (r = −0.193), and the model estimated a direct dependence-to-performance path of −0.068 alongside three negative indirect paths: through attentional control, through fitness, and through attentional control followed by fitness. Attentional control carried most of the indirect effect (76.19% of the total). Grade level moderated both mediational chains, with the positive attentional-control and fitness paths to performance significant at lower grades and non-significant at higher grades, so the negative association between dependence and performance was weaker for older students.

Key Findings

  • Dependence was negatively associated with performance. The direct path from generative-AI dependence to academic performance was small and negative, and the bivariate correlation was moderate.
  • Attentional control was the main channel. The indirect effect through attentional control accounted for 76.19% of the total indirect effect, far more than the other two paths combined.
  • Physical fitness added a second, smaller channel. The indirect effect through fitness accounted for 15.87%, and the serial path through attentional control then fitness for 7.94%.
  • Dependence predicted lower attentional control most strongly. The dependence-to-attentional-control path was the largest in the model, larger than its paths to fitness or performance.
  • Grade level moderated the mediational chains. Both interactions were negative and significant, and at one standard deviation above the grade mean the indirect paths were no longer significant.
  • The model fit the data adequately. Confirmatory fit indices met conventional thresholds, and the correlations were consistent with the hypothesized ordering.

What the model contains

Dependence was measured with a 17-item student instrument covering cognitive-behavioral, functional and emotional dependence (α = 0.898); attentional control with a 21-item Chinese revision of the Attentional Control Scale; fitness with the National Student Physical Health Standard, which weights body morphology at 15%, physiological function at 15% and physical performance at 70%; and academic performance with standardized scores in Chinese, mathematics and English from school-administered examinations. Paths were estimated in AMOS with maximum likelihood, indirect effects were bootstrapped with 5,000 resamples, and moderation was tested with product terms formed from centered mediators and centered grade level.

The theoretical case is that cognitive offloading erodes the attentional capacity that deep processing depends on, while fitness supplies a behavioral and physiological route that supports the same capacity. The two mediators are treated as serially ordered, with attentional control preceding fitness, because the authors argue that cognition shapes behavior.

Reading the grade-level moderation carefully

The moderation is the part most likely to be misread. Lower-grade students showed a positive attentional-control to performance path of 0.429 and a fitness to performance path of 0.367, both significant; at one standard deviation above the grade mean neither path was significant, and the corresponding indirect effects lost their confidence intervals' distance from zero. The authors explain this developmentally: in the lower grades, where learning is still largely receptive, stable attention and physical condition carry more of the achievement signal, whereas senior secondary study rewards active learning, intrinsic motivation and autonomy, which the model does not measure.

That is a plausible reading, but it is a cross-sectional pattern in which "grade" is a proxy for age, curriculum and examination regime at once. What the data support is the descriptive claim that the dependence–performance association was weaker at higher grades, not a claim that maturity protects students.

What this means for practice

  • Instructors. Watch attentional control rather than AI-use time; the model puts the largest dependence-related path through a student's capacity to resist distraction and stay on task.
  • Instructors. Keep unaided reasoning tasks in the weekly routine so that the attentional work AI can shortcut is still practiced.
  • Administrators. Target support at the lower secondary grades, where the attentional-control and fitness paths to achievement were strongest and the dependence-related indirect effects were largest.
  • Researchers. Measure attentional control directly rather than inferring it from grade, and test the serial order with longitudinal data before treating it as a mechanism.

Limitations

  • Cross-sectional design. The questionnaire cannot establish causal ordering, and the authors note that poor performance may itself intensify dependence.
  • Few covariates. Only sex, class and age were controlled; socioeconomic status and parental education were omitted because the sampled schools were all urban.
  • Urban sample only. All schools were in urban China, so rural and cross-regional generalization is untested.
  • Self-reported dependence and attention. Two of the four constructs are questionnaire measures, with the usual common-method and self-report limits.
  • Grades 9 and 12 excluded. Graduating cohorts were deliberately left out, which removes the highest-stakes examination years from the sample.

Citation

Su, Y., Zhang, R., Chen, P., Wang, X., He, Y., & Zhang, X. (2026). Generative artificial intelligence dependence and academic performance: serial mediation through attentional control and physical fitness and moderation by grade level. Frontiers in Psychology, 17, 1993658.

Embed this page

Copy the code below to embed this page in a learning management system or other website. The embedded version shows only the page content — it hides the site header, navigation, and footer.