AI Dependence Is Not the Anxiety Culprit, Landmark Study Finds

As artificial intelligence tools become woven into nearly every corner of university life, a growing chorus of educators and psychologists has warned that students may be becoming dangerously dependent on AI, with anxiety as the inevitable price. A new three-wave study of 1,451 Chinese university students, published in Frontiers in Psychiatry, upends that narrative. The research, led by Xu Liu of Xinxiang University and Universiti Putra Malaysia, together with Zhenggui Gu and Lulu Shi, found that perceived AI dependence was not directly linked to later anxiety symptoms at all. Instead, the strongest predictor of anxiety was a familiar psychological process: experiential avoidance, the tendency to escape uncomfortable thoughts and feelings rather than engage with them.

The study arrives at a moment when AI use in higher education has shifted from novelty to infrastructure. Large language models now assist students with writing, decision-making, feedback, and academic problem-solving, prompting UNESCO to call for a human-centered approach to generative AI in education. Anxiety among students is itself a serious public-health concern: a meta-analysis of 64 studies covering more than 100,000 college students estimated the pooled prevalence of anxiety symptoms at 39.0 percent and depression symptoms at 33.6 percent. Whether AI reliance adds to this burden has been an open question, and the new research offers the most time-ordered answer to date.

To disentangle the relationships, the team used a three-wave time-lagged survey design, a methodological step up from the cross-sectional snapshots that dominate the field. In January 2026, students reported their perceived AI dependence, measured with a validated five-item scale capturing insecurity without AI support, reassurance dependence on AI feedback, replacement anxiety, and competitive pressure, along with their general digital literacy. In March 2026, they completed the Acceptance and Action Questionnaire-II, a widely used index of experiential avoidance and psychological inflexibility. In June 2026, they completed the seven-item Generalized Anxiety Disorder scale, or GAD-7. Separating the predictor, the putative psychological process, and the outcome across three time points reduces same-time measurement bias and establishes a basic temporal order, though the authors are careful to note that the design establishes time-lagged associations rather than strict causation.

The statistical engine behind the study was partial least squares structural equation modeling, or PLS-SEM, conducted in SmartPLS 4 with 5,000 bootstrap subsamples. This variance-based approach was chosen a priori because the model included new constructs, three product-indicator interaction terms testing moderation, and specific indirect effects, alongside diagnostics for explained variance and out-of-sample prediction. The measurement model performed well: Cronbach’s alpha values ranged from 0.845 to 0.939, composite reliability from 0.885 to 0.951, and average variance extracted from 0.608 to 0.733, while heterotrait-monotrait ratios confirmed that the constructs were empirically distinct. Harman’s single-factor test and full collinearity checks suggested that common method bias was unlikely to dominate the findings.

The results were striking. Experiential avoidance measured in March was strongly and positively associated with anxiety symptoms measured in June, with a standardized path coefficient of 0.677 (p < 0.001) and an effect size of f² = 0.827, the largest in the model. By contrast, the direct path from perceived AI dependence to anxiety symptoms was not statistically significant (beta = -0.030, p = 0.147). The model explained 46.6 percent of the variance in anxiety symptoms, but only 2.5 percent of the variance in experiential avoidance, indicating that AI dependence and digital literacy are far from the whole story on the antecedent side. Positive but small Q²_predict values showed limited out-of-sample predictive relevance, so the authors interpret the structural paths as explanatory associations rather than strong prediction.

The most unexpected finding was the direction of the AI dependence pathway. Rather than increasing experiential avoidance, perceived AI dependence was negatively associated with it (beta = -0.114, p < 0.001), producing a statistically significant but theoretically unexpected negative time-ordered indirect association with anxiety symptoms (beta = -0.077, p < 0.001). Before estimation, the team conducted a scoring audit and found no reversed items, miscoding, or anchor inversion, ruling out an identified coding artifact. The authors stress that this negative result should be read as boundary-setting evidence against a simple harmful-use story, not as proof that AI dependence is protective. One plausible explanation is that the dependence scale captures psychological reliance rather than pathological addiction: students who feel insecure without AI may also be actively engaged with AI-supported academic tasks, relying on AI as a cognitive scaffold while still confronting uncertainty rather than avoiding it.

Digital literacy, the other candidate protective factor, also failed to behave as theory predicted. Although higher digital literacy was negatively associated with experiential avoidance (beta = -0.100, p < 0.001), none of the three moderation tests was significant: the interactions between digital literacy and AI dependence predicting avoidance (beta = -0.021, p = 0.498) and anxiety (beta = 0.019, p = 0.334), and between digital literacy and avoidance predicting anxiety (beta = -0.002, p = 0.937), all missed the mark. The authors suggest that the 10-item Digital Literacy Scale, which measures broad competence in locating, evaluating, communicating, and creating digital content, may be too general to capture AI-specific skills such as evaluating AI-generated outputs, detecting hallucinations, recognizing automation bias, and maintaining human judgment when algorithmic advice is available. Digital literacy may also carry mixed implications, increasing both capability and exposure to AI-related academic competition.

Theoretically, the study reframes the mental-health implications of AI in higher education. The constructs of AI use, AI literacy, AI dependence, AI reassurance seeking, and AI-driven competence threat may have different psychological consequences, and mental-health risk cannot be inferred from AI exposure alone. The findings align with decades of automation research showing that humans can misuse, overuse, or defer excessively to automated systems, but they also echo the acceptance and commitment therapy literature, in which experiential avoidance and psychological inflexibility are transdiagnostic processes robustly linked to anxiety across clinical and nonclinical populations. In other words, AI may be a new technological context, but the psychological vulnerability process is familiar and well grounded.

Practically, the authors argue that universities should avoid two extreme messages: treating AI as purely beneficial or equating AI dependence with addiction. Campus mental-health promotion should integrate responsible AI education with support for psychological flexibility, uncertainty tolerance, and autonomous learning. Counseling centers, teaching units, and digital-learning offices could collaborate to identify students who rely on AI for reassurance rather than learning, and interventions that reduce avoidance and build engagement with difficult tasks may be especially valuable in AI-mediated learning environments.

The study has clear limits that the authors confront directly. Baseline anxiety and baseline experiential avoidance were not measured, so the strong association between avoidance and later anxiety may partly reflect pre-existing distress, and the findings cannot establish within-person change. All measures were self-reports, recruitment was purposive at two Chinese universities and restricted to students with prior AI experience, and AI-use frequency was not assessed. The authors call for repeated-measures designs using cross-lagged panel, random-intercept, or latent growth models, along with item-level invariance testing and qualitative work to distinguish functional AI reliance from problematic reassurance dependence, and more specific moderators such as AI literacy, prompt evaluation, and algorithmic awareness. Even so, the central message stands: in the AI era, the anxiety story on campus is less about the machines and more about how students relate to their own discomfort.

Subject of Research: Perceived artificial intelligence dependence and anxiety symptoms among university students: a three-wave study of experiential avoidance and digital literacy

Article Title: Perceived artificial intelligence dependence and anxiety symptoms among university students: a three-wave study of experiential avoidance and digital literacy

Article References: Liu, X., Gu, Z., & Shi, L. (2026). Perceived artificial intelligence dependence and anxiety symptoms among university students: a three-wave study of experiential avoidance and digital literacy. Frontiers in Psychiatry, 17, Article 1934192. https://doi.org/10.3389/fpsyt.2026.1934192

Image Credits: AI Generated

DOI: 10.3389/fpsyt.2026.1934192

Keywords: Perceived, artificial, intelligence, dependence, anxiety, symptoms, university, students, three-wave, experiential, avoidance, digital