The dark side of AI in education: AI dependency as a mediator linking academic self-efficacy and teacher support to learning burnout among university students

Synthesis: In a cross-sectional survey of 276 Chinese undergraduates, both academic self-efficacy and teacher support were negatively associated with AI dependency, while AI dependency was positively associated with learning burnout. AI dependency fully mediated both relationships, meaning the effect of each resource on burnout ran almost entirely through how dependent students had become on AI. The study frames excessive AI reliance as a resource-depletion process that links individual and contextual resources to students' emotional exhaustion and disengagement.

Key Findings

• Academic self-efficacy lowers AI dependency. Students who felt more confident about completing learning tasks on their own were markedly less likely to depend on AI. The authors read this through Cognitive Offloading Theory: low-confidence students delegate demanding tasks to AI, which habitual delegation then reinforces.

• Teacher support also reduces AI dependency, but only weakly. Students who received clearer guidance, involvement, and autonomy support from teachers showed less AI dependency, yet teacher support explained only a small share of the variance in AI dependency. The authors take this as a sign that their model captures only part of what drives dependency — other individual and contextual factors matter too.

• AI dependency predicts learning burnout. The more students relied on AI, the more they reported emotional exhaustion, cynicism, and diminished academic efficacy — the hallmarks of learning burnout. AI dependency accounted for a substantial share of the variance in burnout.

• AI dependency fully mediates both paths. Once AI dependency was included in the model, the direct effects of academic self-efficacy and teacher support on burnout became non-significant; the indirect effects ran through dependency and accounted for the bulk of each total effect. This positions AI dependency as the transmission channel linking individual and contextual resources to burnout.

• Reverse models were weaker. Alternative models swapping the direction of mediation produced significant but substantially smaller indirect effects, which the authors treat as supporting their proposed direction.

• Dependency varied with usage, not gender. AI dependency did not differ by gender but rose with frequency of AI use and, to a small degree, across academic year (seniors more dependent than juniors). These are group differences rather than formal moderation effects.

Study Design & Method

A cross-sectional survey design was used. An online structured questionnaire distributed via Questionnaire Star reached undergraduates in selected courses at several comprehensive universities in Chengdu, China, recruited through convenience sampling. Of 311 questionnaires returned, 276 were retained as valid (an 88.75% effective rate). The sample was 55.8% male, juniors made up about two-thirds, and most students used AI tools frequently — over half reported using them once a day or more.

Four instruments were used: an eight-item Academic Self-Efficacy Scale adapted from Owen and Froman, a ten-item Teacher Support Scale from Skinner and Belmont covering Structure, Involvement, and Autonomy Support, a seven-item AI Dependency Scale adapted from Young's Internet Addiction Test with items reworded for academic AI use, and a nine-item shortened Learning Burnout Scale derived from the MBI-SS. All items used five-point Likert scales, and several self-efficacy items were rewritten to reference learning without AI assistance. The measurement structure was validated with exploratory and confirmatory factor analysis and showed satisfactory reliability and discriminant validity; a common method bias check found no substantial single-factor or collinearity concerns. Regression and mediation were run in SPSS with PROCESS, and alternative reverse-mediation models and models controlling for demographics were estimated as robustness checks, which did not materially change the findings.

Implications

The mediation results locate AI dependency as the channel through which both an individual resource (academic self-efficacy) and a contextual resource (teacher support) relate to burnout — reinforcing a Conservation of Resources reading that depleted internal resources push students toward AI as a compensatory external resource, which in turn erodes the cognitive resources needed to sustain engagement. Because the direct effects were non-significant once AI dependency entered the model, interventions aimed only at boosting self-efficacy or teacher support may have limited traction on burnout unless they also address how students actually use AI.

The authors' practical recommendations are threefold: educators should supply timely feedback, personalized guidance, and positive teacher–student interaction, and integrate AI meaningfully as a learning aid rather than banning it; institutions should embed AI literacy into general and professional curricula and cultivate self-regulated learning; and policymakers should issue clear guidelines on responsible AI use, academic integrity, and the prevention of over-dependency. Notably, the weak explanatory power of teacher support for AI dependency suggests single-factor interventions should be complemented by broader approaches.

Limitations

• The cross-sectional design cannot establish causality or capture change over time; reciprocal relationships are plausible (burnout may itself drive dependency), and longitudinal or experimental designs are needed.
• All measures were self-report, inviting social desirability and subjective bias; the adapted AI Dependency Scale showed satisfactory reliability and construct validity but its content validity was not assessed via expert review or pilot testing.
• The sample was limited to undergraduates at a few comprehensive universities in Chengdu, China, using convenience sampling — limiting external validity, since AI dependency may vary with curriculum, assessment practices, technology access, institutional AI policy, and teacher–student relationships across contexts.
• PROCESS uses observed composite scores rather than latent variables and does not model measurement error; latent-variable structural equation modeling with bootstrapped indirect effects is recommended.
• Other relevant antecedents (self-regulated learning, intrinsic motivation, critical thinking, academic engagement) and boundary conditions (AI literacy, digital competence, usage frequency, discipline, personality, achievement) were not modeled; the observed group differences are not formal moderation effects.

Connected Concepts

• AI Misuse and Learning Harm — AI dependency is framed as maladaptive reliance that shifts AI from learning aid to external regulator of cognition, with burnout as the harm
• Cognitive Offloading — Cognitive Offloading Theory explains why low-confidence students delegate demanding tasks to AI, and why habitual delegation depletes the resources that protect against burnout
• Self-Efficacy — academic self-efficacy is the key individual resource that reduces AI dependency, extending self-efficacy from achievement to AI-related behavior
• Teaching — teacher support (Structure, Involvement, Autonomy Support) acts as a contextual resource that guides appropriate AI use and buffers dependency
• Well-Being — learning burnout comprises emotional exhaustion, cynicism and diminished academic efficacy, the psychological cost of dependency
• Higher Education — the study targets university students in AI-assisted learning environments
• Student Engagement — dependency is linked to passive learning, weakened participation and reduced engagement

Connected Articles

• Are Students Dependent on AI in Writing Courses? Analyzing Factors Influencing Dependence on Generative AI Through the I-PACE Model — companion study of AI dependence using the I-PACE model; same construct and population, different theoretical frame (writing courses)
• Understanding Student Dependency on AI: The Role of AI Literacy, Academic Self-Efficacy, and Resource Management Strategies — directly overlaps on AI dependency, academic self-efficacy and higher-education students
• A Critical Narrative Synthesis of Psychological Correlates, Measurement, and Reported Findings on Conversational AI Engagement and Dependence-Related Constructs — review of dependence-related constructs and their psychological correlates, including measurement concerns relevant to the adapted IAT scale
• Epistemic Dependence in AI-Mediated Learning — theoretical treatment of dependence in AI-mediated learning and why it erodes autonomy
• Meta-Cognitive Insights into Cognitive Offloading: Mechanisms, Interventions, and Educational Implications — mechanisms and interventions for cognitive offloading, the process underlying the dependency path tested here

Citation

Huang, Q., Tu, S., Lin, J., Lu, L., & Lv, C. (2026). The dark side of AI in education: AI dependency as a mediator linking academic self-efficacy and teacher support to learning burnout among university students. Frontiers in Psychology, 17, 1889053.