ChatGPT-induced technostress among UAE university students: a PLS-SEM and IPMA investigation of AI anxiety, cognitive overload, and perceived authenticity

Abstract

University students in the United Arab Emirates are among the most intensive users of ChatGPT worldwide. Little is known, however, about the psychological costs associated with this adoption. This study aims to examine how three AI-related stressors – perceived ChatGPT authenticity, AI anxiety and cognitive overload – shape students’ trusting intention toward ChatGPT and, through that, their experience of technostress. Drawing on technostress theory, trust theory and cognitive load theory, a dual-pathway structural model was developed and tested using partial least squares structural equation modeling with bootstrapping (5,000 subsamples). Data were collected from 288 university students across multiple UAE higher education institutions. Robustness was assessed through collinearity diagnostics, full-collinearity common-method-bias testing, specific indirect effects and out-of-sample predictive assessment. Importance–performance map analysis (IPMA) targeting technostress was conducted to identify priority intervention targets. All six hypotheses were supported. AI anxiety was the dominant direct predictor of technostress (β = 0.493), followed by cognitive overload (β = 0.225) and trusting intention (β = 0.199). Cognitive overload was the strongest predictor of trusting intention (β = 0.403). Most strikingly, higher trust in ChatGPT was associated with greater, not lower, technostress – a double-edged dynamic in which deeper reliance accompanies as much strain as it relieves. All three indirect effects through trusting intention were significant, indicating partial complementary mediation. IPMA targeting technostress identified AI anxiety as the highest-priority intervention target (importance = 0.522), ahead of cognitive overload (0.305). The cross-sectional, single-country design precludes causal inference; all relationships should be read as associations. Future research should adopt longitudinal designs and replicate the model across other national and cultural contexts. The model’s explanatory power (R² = 0.670 for technostress) and out-of-sample predictive relevance support the framework’s value within the UAE higher education setting studied here; extension to other contexts remains an empirical question. Institutions should prioritize anxiety-reduction measures, which carry the largest total effect on technostress, alongside scaffolded AI literacy programs and structured task guidance that reduce cognitive overload. Instructors should normalize uncertainty about ChatGPT to mitigate anxiety-driven technostress. Policymakers should accompany rapid AI adoption mandates with investment in student digital literacy and wellbeing infrastructure. This study extends technostress theory to the generative AI context in higher education. It establishes perceived ChatGPT authenticity as a novel antecedent of AI trust, and repositions trusting intention as a technostress creator rather than inhibitor in high-adoption educational environments. This pattern may extend to other settings in which AI adoption outpaces institutional support.