Overtrust in Generative AI and Students’ Critical Thinking: An Integrated PLS-SEM and Machine Learning Analysis in Vietnamese Higher Education

The rapid adoption of generative AI (GenAI) among university students has raised concerns that excessive trust in AI outputs may erode critical thinking. This study examines how overtrust relates to critical thinking through cognitive offloading and AI dependency, with AI literacy as a protective moderator, integrating partial least squares structural equation modelling (PLS-SEM) with machine learning. Survey data from 580 Vietnamese undergraduates were analysed through a five-step pipeline: exploratory structure learning, confirmatory path estimation, predictive benchmarking, explainable AI, and heterogeneous-effect estimation. Overtrust was associated with higher offloading (β = 0.390) and dependency (β = 0.402), both of which were associated with lower critical thinking (β = -0.251 and -0.348), while AI literacy buffered the association with dependency (interaction β = 0.154). Crucially, the direct effect of overtrust was non-significant and the two mediators carried 89% of its total effect, indicating indirect-only mediation. A parsimonious six-variable classifier separated at-risk students with AUC ≈ 0.77, whereas demographics alone performed at chance; SHAP analysis identified AI dependency as the dominant feature in the classification decision, and a causal forest confirmed the buffering pattern across the low-to-mid range while exposing a non-monotonic reversal under sparse support at the upper tail. Interventions should therefore target AI-use behaviour rather than trust itself. These findings are associational rather than causal: the design is cross-sectional with all measures self-reported at a single point, so the classifier performs concurrent screening rather than temporal prediction.