GenAI dependence and graduate students’ research creativity: the roles of critical thinking and human–AI collaboration quality

Synthesis: Yin, Wu, Liu and Li collected questionnaire data from 1,157 graduate students in China through online convenience sampling and used structural equation modeling plus PROCESS Model 6 with 5,000 bootstrap resamples to relate GenAI dependence to self-reported research Creativity. They split dependence into functional (task-oriented, instrumental) and existential (deeper psychological reliance) forms, and tested critical thinking and human–AI collaboration quality as serial mediators. Both forms were positively associated with research creativity, but the total association was far stronger for functional dependence (β = 0.457) than for existential dependence (β = 0.147), and the pathway through critical thinking carried approximately 85.0% of the functional form's total indirect effect. The design is cross-sectional, so the authors report every path as a statistical association rather than a causal effect.

Key Findings

• Functional dependence was the stronger correlate of research creativity. Its total effect was β = 0.457 (p < 0.001) against β = 0.147 (p < 0.001) for existential dependence, and both direct effects stayed significant after the mediators entered the model (β = 0.052 and β = 0.045, both p < 0.05), indicating partial mediation in each case.
• Critical thinking was the dominant indirect pathway for functional dependence. The total indirect effect was 0.4055, 95% CI [0.3470, 0.4608], of which the critical-thinking path (Ind1) was 0.3449, 95% CI [0.2925, 0.3967], approximately 85.0% of the total.
• All three indirect pathways were significant for both forms of dependence. Functional: Ind1 through critical thinking 0.3449 [0.2925, 0.3967]; Ind2 through human–AI collaboration quality 0.0431 [0.0182, 0.0703]; Ind3 through both in sequence 0.0175 [0.0072, 0.0301]. Existential: total indirect 0.1025 [0.0395, 0.1645]; Ind1 0.0734 [0.0163, 0.1301]; Ind2 0.0229 [0.0117, 0.0364]; Ind3 0.0062 [0.0012, 0.0128].
• The serial pathway was larger for functional dependence (Ind3 = 0.0175) than for existential dependence (Ind3 = 0.0062), and existential dependence showed substantially smaller effect sizes across every specific indirect pathway.
• In the overall SEM, GenAI dependence was positively associated with critical thinking (β = 0.492, p < 0.001), human–AI collaboration quality (β = 0.538, p < 0.001) and research creativity (β = 0.064, p < 0.05); critical thinking was associated with human–AI collaboration quality (β = 0.348, p < 0.001) and research creativity (β = 0.731, p < 0.001); human–AI collaboration quality was associated with research creativity (β = 0.101, p < 0.001).
• The model fit the data acceptably (χ2/df = 7.541, CFI = 0.988, GFI = 0.981, SRMR = 0.017), though the authors note the χ2/df ratio was relatively high.
• In the separate mediation models, the path coefficients diverged by form of dependence. For functional dependence: dependence → critical thinking 0.464, dependence → collaboration quality 0.443, critical thinking → collaboration quality 0.388, critical thinking → research creativity 0.744, collaboration quality → research creativity 0.097 (all p < 0.001). For existential dependence: 0.097 (p < 0.01), 0.205, 0.571, 0.755 and 0.111 (all p < 0.001).

Study Design and Measures

This is a cross-sectional survey of graduate students in China, recruited through the Wenjuanxing platform and distributed via university-based WeChat groups and social media; participation was voluntary. Of 1,277 questionnaires collected, 1,157 valid cases remained after excluding patterned answering, implausible completion times and failed attention checks, a valid-response rate of 90.60%. The sample was 47.796% male (553) and 52.204% female (604).

Dependence was measured with the 18-item AI Dependence Questionnaire (Cronbach's α = 0.913), covering functional and existential dimensions on a 5-point scale; critical thinking with a 5-item scale (α = 0.863); human–AI collaboration quality with an 11-item scale covering outcome quality, comfort and efficiency (α = 0.935); and research creativity with a 6-item scale (α = 0.885). Gender, degree type, grade level and academic discipline were controls in all mediation analyses. Harman's one-factor test extracted five factors with eigenvalues above one, the first accounting for 34.119% of variance, below the 40% threshold; a single-factor CFA fit poorly (χ2/df = 9.812, GFI = 0.633, AGFI = 0.590, CFI = 0.782, RMR = 0.143, NFI = 0.763).

Variable means on the 5-point scale ranged from 3.11 to 3.77, and all study variables correlated positively: functional dependence with existential dependence r = 0.339, critical thinking r = 0.457, human–AI collaboration quality r = 0.620 and research creativity r = 0.456; existential dependence with critical thinking r = 0.094, collaboration quality r = 0.258 and research creativity r = 0.148 (all p < 0.01).

What this means for practice

• Instructors and supervisors. Treat dependence as differentiated, not as a single dial to turn down: the paper's practical reading is to promote purposeful, task-oriented use while watching for the psychological form. Functional dependence — using GenAI for procedural tasks while staying engaged in problem framing and reasoning — carried the stronger positive association.
• AI literacy programs. Build critical thinking into training rather than stopping at prompt technique. Because critical thinking carried most of the indirect association for functional dependence, verification of AI-generated claims, comparison of alternatives, source checking and justification of accept-or-reject decisions are the actionable practices.
• Supervisors. Encourage reflective collaboration and calibrated trust: the paper recommends students document how AI-generated content was evaluated and modified, periodically complete important tasks without GenAI, and use human review for high-stakes decisions.
• Institutions. Avoid framing GenAI dependence as uniformly desirable or undesirable; the authors argue the goal is neither unconditional adoption nor categorical avoidance, but human-directed, transparent AI-assisted research that preserves cognitive agency.

Limitations

• The cross-sectional design "prevents strong causal inference and does not establish the temporal ordering" among GenAI dependence, critical thinking, human–AI collaboration quality and research creativity; the authors state the indirect effects should be read as statistical associations, not causal mediation, and note reverse causality is possible — students with stronger critical thinking or higher research creativity may be more likely to use GenAI strategically.
• Online convenience sampling via WeChat groups and social media may introduce self-selection bias and limits representativeness across institutions, disciplines, regions and cultures. Unmeasured variables such as prior AI experience, research Self-Efficacy, academic ability, supervisory support and previous research productivity may contribute to the observed associations.
• All constructs were self-reported, which may inflate shared-method variance; research creativity in particular reflects participants' perceptions of their own creativity rather than independently evaluated output.
• The authors caution that the positive associations should not be read as "greater dependence is uniformly beneficial"; existential or excessive dependence may carry risks including cognitive offloading, reduced originality, algorithmic conformity, weakened independent reasoning and diminished confidence without AI assistance.

Citation

Yin, S., Wu, R., Liu, X., & Li, R. (2026). GenAI dependence and graduate students' research creativity: the roles of critical thinking and human-AI collaboration quality. Frontiers in Psychology, 17, 1896963.