Heavy Reliance on AI Chatbots Linked to Academic Anxiety in Nursing Students

Nursing students in southern China who report a strong dependence on large language models such as AI chatbots also tend to report higher levels of academic anxiety, according to a cross-sectional survey published in Frontiers in Psychology. The study, conducted among 473 nursing students in Guangdong Province between March and April 2026, found that the direct link between dependence on AI tools and anxiety about academic performance was statistically significant but small. More of the association traveled through two intermediate experiences: academic burnout, the emotional exhaustion that builds up around coursework and clinical training, and a pattern the researchers call artificial intelligence addiction, in which use of AI tools becomes excessive and difficult to control.

The research team, led by Taiping Li of Guangdong University of Science and Technology, recruited participants through teaching staff, clinical placement supervisors and online channels using convenience sampling. Of 500 nursing students invited to take part, 473 provided valid questionnaires, a retention rate of 94.6 percent. The sample included students on clinical placements, who retained their student status, so it represented learners at different stages of nursing education, from classroom study to hospital-based training. About 80 percent of respondents were female, and most were between 18 and 23 years old. The authors emphasize that the survey captured a single moment in time, which means the results describe associations rather than proven causes.

To measure the four key constructs, the researchers drew items from previously published self-report instruments. Large language model dependence was assessed with 11 items adapted from an instrument developed by Li and colleagues, including statements such as believing that large language models possess all the functions a user expects. Artificial intelligence addiction was measured with nine items from an instrument by El-Sayed and colleagues, covering patterns such as using AI tools longer or more frequently than originally intended. Academic anxiety was assessed with six items from the Academic Anxiety Scale by Bülbül and Odacı, and academic burnout with eight items from a scale by Jan and Parveen, including feeling drained and emotionally exhausted by studies. All items were administered in Chinese on five-point Likert scales, and an English-language instrument was translated and back-translated to ensure semantic equivalence.

The statistical approach was partial least squares structural equation modeling, a technique suited to predictive models with multiple interlocking pathways. Age and sex were included as covariates with direct paths to academic anxiety, although neither showed a statistically significant association. The researchers evaluated direct and indirect pathways using bootstrapping with 5,000 resamples, reporting standardized path coefficients, confidence intervals and p-values. The measurement side of the model performed well: Cronbach’s alpha values ranged from 0.914 to 0.944, composite reliability from 0.935 to 0.954, and average variance extracted from 0.560 to 0.721, all meeting conventional thresholds. Heterotrait-monotrait ratios between 0.485 and 0.733 supported the idea that the four constructs, though related, are empirically distinct.

The structural results told a layered story. Large language model dependence showed a strong positive association with artificial intelligence addiction, with a standardized coefficient of 0.580 and a large effect size. AI addiction, in turn, was positively associated with academic burnout, which was strongly linked to academic anxiety. The direct path from AI dependence to academic anxiety was positive but modest, with a coefficient of 0.138, a confidence interval of 0.006 to 0.278, and a p-value of 0.048, just inside the conventional significance threshold. The authors caution against overinterpreting this small direct effect, noting that its contribution to explained variance was limited. Overall, the model accounted for 49.2 percent of the variance in academic anxiety, 34.1 percent in academic burnout and 33.6 percent in AI addiction.

The indirect pathways were where the model gained most of its explanatory weight. The specific indirect association from AI dependence to academic anxiety through academic burnout was significant, with a coefficient of 0.205 and a confidence interval of 0.123 to 0.303. A second significant pathway ran through artificial intelligence addiction followed by academic burnout, with a coefficient of 0.098 and a confidence interval of 0.049 to 0.147. By contrast, the pathway from AI dependence to anxiety through AI addiction alone was not statistically significant, suggesting that problematic AI engagement matters for anxiety mainly insofar as it feeds into study-related exhaustion. An additional analysis found that AI addiction had a significant indirect association with anxiety through burnout, reinforcing burnout’s central position in the model.

Model fit evidence was mixed, and the authors are explicit about this limitation. The standardized root mean square residual was 0.068, below the 0.08 reference value, but the normed fit index was 0.812, falling short of the 0.90 benchmark. The researchers state that these indices do not support an unqualified conclusion that the model demonstrated acceptable overall fit. They also report a common method bias assessment: the first unrotated factor in Harman’s single-factor test explained 41.884 percent of variance, below the commonly used 50 percent threshold, but the authors acknowledge that because all data came from self-reports at a single time point, common method variance could not be completely ruled out.

The theoretical framing draws on perspectives from problematic technology use and social stress theory. Large language model dependence is defined as a tendency to rely on AI tools for cognitive, academic or work-related tasks, especially when users believe the tools can satisfy most of their expected functions. AI addiction, by contrast, describes a maladaptive pattern marked by prolonged use, difficulty controlling engagement and continued use despite negative consequences. The distinction matters because frequent or instrumental AI use is not inherently problematic, but the study’s findings suggest that heavy reliance may coexist with poorly regulated use, which in turn is associated with reduced learning motivation and emotional exhaustion. In nursing education, where independent judgment, critical thinking and professional reasoning are essential, the authors argue that overreliance on AI-generated answers may leave students less confident in their own abilities when facing tasks without technical help.

The authors are careful to stress what the design cannot show. Because the data were collected at a single point in time, the findings represent statistical indirect associations within a theory-specified model, not evidence of temporal ordering or causal mediation. Reverse explanations remain plausible: students experiencing greater academic anxiety may seek out more AI assistance to manage uncertainty, and students experiencing burnout may find independent task completion more demanding and lean more heavily on AI support. Shared influences such as academic pressure or pre-existing psychological distress could also shape all four constructs simultaneously. The proposed sequence, in which dependence fosters addiction, addiction feeds burnout and burnout heightens anxiety, is theoretically specified but open to alternative orderings that only longitudinal or experimental designs could test.

Several further limitations shape how the results should be read. Convenience sampling through institutional and online channels may have attracted students with greater interest in AI, introducing selection bias, and the sample was restricted to nursing students in one Chinese province, limiting generalizability to other disciplines, regions and cultures. AI-use frequency, duration and purpose were not measured, so their potential confounding influence could not be evaluated, and demographic adjustment was limited to age category and sex. The study also did not evaluate any educational or psychological intervention, so it cannot demonstrate that changing AI use or reducing burnout would improve academic anxiety. Still, the authors suggest practical directions worth exploring, including AI literacy education covering critical evaluation of AI-generated information, verification against authoritative sources and appropriate disclosure of AI assistance, along with attention to academic workload, supervisory support and access to counseling. The measures used do not establish clinical diagnoses or thresholds for intervention, but the findings add a cautious counterpoint to research emphasizing only the benefits of AI-assisted learning, indicating that in demanding professional training environments, the way students relate to AI tools may be intertwined with their psychological well-being.

Subject of Research: Associations between large language model dependence and academic anxiety among Chinese nursing students, mediated statistically by AI addiction and academic burnout.

Article Title: The association between large language model dependence and academic anxiety among nursing students: statistical indirect associations through artificial intelligence addiction and academic burnout

Article References: Li, T., Ye, Q., Dong, J., & Yan, Y. (2026). The association between large language model dependence and academic anxiety among nursing students: statistical indirect associations through artificial intelligence addiction and academic burnout. Frontiers in Psychology, 17, Article 1917190. https://doi.org/10.3389/fpsyg.2026.1917190

Image Credits: AI Generated

DOI: 10.3389/fpsyg.2026.1917190

Keywords: large language models, nursing students, academic anxiety, academic burnout, artificial intelligence addiction, PLS-SEM, cross-sectional study, China, nursing education, problematic AI use, mental health, structural equation modeling