Artificial intelligence attitudes and mental health in Chinese university students: Nonlinear relationships and the mediating role of self-efficacy

Abstract

BACKGROUND

When students perceive artificial intelligence (AI) as a threat to their career prospects (an environmental factor), this perception may elicit self-doubt (a cognitive factor), which in turn may lead to avoidance of academic challenges (a behavioral response), ultimately contributing to anxiety or depression (psychological outcomes). However, to date, the literature lacks empirical evidence supporting this viewpoint within university student samples.

AIM

To explore the associations between multidimensional AI attitudes and mental health outcomes among Chinese university students, and to examine the nonlinear relationships and the mediating role of self-efficacy in these associations.

METHODS

A cross-sectional survey was conducted among 2736 undergraduate students from China. Linear regression, restricted cubic spline (RCS) analysis, and path analysis were applied to examine the relationships between AI attitudes (anxiety, job insecurity, self-doubt, and procrastination), self-efficacy, and mental health outcomes (anxiety, depression, stress, loneliness, suicidal ideation, and non-suicidal self-injury).

RESULTS

Students’ AI attitudes showed a differentiated structure, characterized by low anxiety (9.5%), high job insecurity (50.4%), and moderate levels of self-doubt (25.6%) and procrastination (36.8%). All four AI attitude dimensions positively predicted poor mental health, with procrastination emerging as the strongest predictor. The RCS analysis revealed heterogeneous nonlinear relationships: AI-related anxiety showed nonlinearity for most outcomes, self-doubt and procrastination only for severe outcomes, while job insecurity exhibited stable linearity. Path analysis confirmed that self-efficacy played a partial mediating role in the AI attitude-mental health link.

CONCLUSION

AI attitudes are associated with complex linear and nonlinear impacts on university students’ mental health, with self-efficacy serving as a key mediating mechanism. These correlational findings provide an empirical reference for targeted mental health interventions and AI literacy education in higher education settings.

Key Words: Artificial intelligence attitudes; University students; Mental health; Self-efficacy; Nonlinear relationship

Core Tip: This study explored the associations between artificial intelligence (AI) attitudes and mental health outcomes among 2736 college students. It identified distinct linear and nonlinear relationships across four dimensions of AI attitudes, and confirmed that general self-efficacy played a partial mediating role in these associations.

INTRODUCTION

In contemporary society, the accelerating pace of life, intensifying competition, and growing uncertainties have placed university students under increasing pressure, making mental health issues more prominent[1,2]. Mental health not only affects individuals’ quality of life and subjective well-being but also has significant implications for academic performance, social functioning, and future career development[3,4]. Anxiety, depression, perceived stress, and loneliness are prevalent among university students[5,6]; concurrently, suicidality (SI) and non-suicidal self-injury (NSSI) have drawn growing public concern[7]. Hence, an in-depth exploration of the factors influencing university students’ mental health and the underlying mechanisms is of considerable practical value.

The rapid proliferation of artificial intelligence (AI) and the experiences associated with its use may affect students’ mental health through multiple pathways[8]. In particular, attitudes toward AI use—characterized by heightened anxiety, perceived job insecurity, persistent self-doubt, and increased procrastination—may exert adverse effects on mental health via several mechanisms[9]. Theoretically, these effects can be understood through Social Cognitive Theory (SCT; Bandura, 2001)[10], which posits a dynamic interplay among personal, behavioral, and environmental factors. For instance, when students perceive AI as a threat to their career prospects (an environmental factor), this perception may trigger self-doubt (a cognitive factor), which in turn may lead to avoidance of academic challenges (a behavioral response), ultimately contributing to anxiety or depression (psychological outcomes). Nevertheless, empirical evidence supporting this theoretical account within university student populations remains scarce.

To address this gap, the present study focused on university students who use AI and assessed AI-related attitudes using validated scales that capture technology anxiety, perceived job insecurity, persistent self-doubt, and procrastination. Grounded in SCT, we propose that AI-related attitudes are associated with mental health outcomes and that self-efficacy serves as a mediating mechanism: More negative AI attitudes are expected to be linked to lower self-efficacy and, subsequently, higher levels of mental health problems (i.e., anxiety, depression, stress, loneliness, SI, and NSSI). Specifically, we will test the following hypotheses: H1: AI-related attitudes are positively associated with the six mental health outcomes; and H2: Self-efficacy mediates these associations.

MATERIALS AND METHODS

Participants and procedures

This study was conducted from April to May 2025 using a convenience sample drawn from multiple provinces across China, with 78% of valid respondents coming from three universities (Jiangxi Police College, Nanchang Medical College, and Nanchang University) in Nanchang, Jiangxi Province. Participants voluntarily registered for the study. Prior to data collection, researchers provided a detailed explanation of the study objectives, procedures, and participants’ rights, and informed consent was obtained from all individuals. Data were collected via Wenjuanxing, an online survey platform in China, with participation strictly voluntary and informed consent reaffirmed before questionnaire completion. The survey was designed to take approximately 10–15 minutes and was administered during classroom sessions or evening self-study periods organized by the universities. Participants received a compensation of 2–5 yuan upon completing the questionnaire. Ethical approval was obtained from the Medical Ethics Committee of the Health Science Center at Yangtze University (Approval No. 202502001).

A total of 3674 questionnaires were distributed to undergraduates nationwide via convenience sampling. After quality screening, 338 questionnaires containing contradictory or logically inconsistent responses were eliminated. An additional 600 participants who reported no AI usage in the previous month were further excluded. This exclusion was necessary because all core measurement items focused on psychological perceptions generated by personal AI experience, and respondents without relevant usage experience could not provide valid responses to these items. Ultimately, 2736 valid questionnaires were retained for statistical analysis.

Measures

Basic demographic variables: Sociodemographic information was collected, including age, gender, academic year, residence, whether the participant was an only child, and field of study.

Depression, anxiety, and stress: To evaluate psychological distress, the study employed the Chinese version of the Depression Anxiety and Stress Scale-21 (DASS-21)[11]. This instrument comprises 21 items across three subscales—depression, anxiety, and stress—with seven items per subscale. Example items include: “I could not feel enthusiastic about anything” (depression), “I felt scared for no good reason” (anxiety), and “I tended to overreact to things” (stress). Respondents rate each item on a four-point scale indicating how well the statement described their experiences over the preceding week. Higher subscale scores indicate greater distress. The DASS-21 has been widely used in Chinese population studies[12]. In this study, internal consistency was satisfactory, with Cronbach’s α values of 0.87 for stress, 0.84 for anxiety, and 0.85 for depression.

Loneliness: Loneliness was assessed using the Three-Item Loneliness Scale[13], which employs a three-point response format: “hardly ever”, “some of the time”, and “often”. One A representative item asks, “How frequently do you feel that you lack companionship?” Subscale scores are summed to yield a total ranging from 3 to 9, with higher scores denoting greater loneliness. The Chinese version has demonstrated strong reliability and validity[14]. In the present analysis, the scale showed a Cronbach’s α of 0.79.

Suicidal ideation/NSSI: Suicidal ideation and non-suicidal self-injury (NSSI) were assessed using two single-item indicators. For suicidal ideation: “In the past 12 months, have you had any thoughts of suicide?” For NSSI: “In the past 12 months, have you engaged in NSSI?” Both items were rated on a 5-point Likert scale (1 = never, 5 = always). These measures are widely used and have demonstrated reliability and validity in prior research on college student mental health[15,16].

Self-efficacy: The General Self-Efficacy Scale (GSES) was used to assess participants’ perceived self-efficacy. We employed the Chinese version, which comprises 10 items rated on a 4-point Likert scale (1 = strongly disagree to 4 = strongly agree). Total scores range from 10 to 40, with higher scores indicating greater self-efficacy. The GSES has demonstrated strong reliability and validity in Chinese populations. In the present study, internal consistency, assessed via Cronbach’s α, was 0.86.

Attitudes toward AI use: In this study, we assessed participants’ use of AI through a set of structured questions based on previous studies[17,18]. The procedure was as follows: First, participants were asked whether they use AI in daily life; if yes, they responded to four single-item indicators corresponding to four constructs focusing on attitudes toward AI use. Consistent with common questionnaire design strategies in the existing literature, one reverse-coded item was adopted to avoid response-set bias, whereby participants might uniformly select the same Likert score for all items. The four constructs and their items are: (1) Increased anxiety: “Using AI to complete tasks, my anxiety has decreased compared with before” (reverse-scored); (2) Job insecurity: “AI progress has increased my worry about future job competitiveness”; (3) Self doubt: “Seeing others proficient at AI makes me feel self-doubt”; and (4) Increased procrastination: “The convenience of AI makes me procrastinate more on important tasks”. All items are rated on a 5-point Likert scale (1 = strongly disagree to 5 = strongly agree). Item 1 is reverse-scored; items 2–4 are forward-scored.

Statistical analysis

Descriptive analysis: Descriptive analyses were conducted at the outset. Qualitative data are summarized as n (%), while quantitative data are presented as the mean ± SD.

Linear regression: We used linear regression to examine relationships among key variables, including attitudes toward AI use, self-efficacy, and psychological outcomes (stress, anxiety, depression, loneliness, SI, and NSSI). The analysis was performed in two steps. Model 1 regressed each psychological outcome on demographic covariates (e.g., age, sex, education, and income) and attitudes toward AI use.

Non-liner relationship: We investigated non-linear relationships between attitudes toward AI use and the set of mental health outcomes (stress, anxiety, depression, loneliness, SI, and NSSI). The primary predictor was the attitude score toward AI use. We modeled potential nonlinear associations using restricted cubic splines (RCS) with four knots, adjusting for demographic covariates. To test result robustness, sensitivity analyses with three knots (10^th, 50^th, and 90^th quantiles) were performed. The nonlinear curve patterns, inflection points, and statistical significance of nonlinear relationships remained nearly identical across the 3-knot and 4-knot models for all exposures and outcomes, indicating reliable and consistent nonlinear findings (Supplementary Figure 1).

Mediation analysis: We assessed whether self-efficacy mediates the relationship between attitudes toward AI use and mental health outcomes. Path analysis was employed to estimate direct, indirect, and total effects. Model fit was evaluated using indices such as NFI, CFI, TLI, RMSEA, and SRMR; indirect effects were tested with bootstrap confidence intervals (e.g., 5000 resamples). Covariates (demographic factors) were controlled for in the mediation model.

All statistical analyses were conducted using R version 4.5.1. A two-tailed test was employed, with statistical significance set at P < 0.05.

RESULTS

The study sample comprised 2736 undergraduate students (Table 1). Among them, 1607 were female (58.7%) and 1129 were male (41.3%). In terms of grade level, there were 592 freshmen (21.6%), 688 sophomores (25.1%), 968 juniors (35.4%), 419 seniors (15.3%), and 69 fifth-year students (2.5%). Regarding academic fields, 490 students majored in science (17.9%), 565 in engineering (20.7%), 473 in humanities (17.3%), 610 in medicine (22.3%), and 598 in other disciplines (21.9%). With respect to family structure, 2012 participants were only children (73.5%), while 724 were not (26.5%). As for residence, 1372 students came from urban areas (50.1%) and 1364 from rural areas (49.9%). The mean age of the sample was 20.49 years (SD = 1.46). The psychometric indicators, presented as the mean ± SD, were as follows: Anxiety 4.21 ± 3.44, depression 3.85 ± 3.46, stress 3.73 ± 3.27, and loneliness 1.76 ± 1.52. The mean score for self-efficacy was 14.72 (SD = 6.27). For self-harm indicators, the mean values were 0.15 ± 0.49 for NSSI and 0.22 ± 0.54 for SI. In addition, the combined endorsement rates were as follows: Increased anxiety 9.5% (agree 7.5% + strongly agree 2.0%), job insecurity 50.4% (agree 36.3% + strongly agree 14.1%), self-doubt 25.6% (agree 18.7% + strongly agree 6.9%), and increased procrastination 36.8% (agree 27.2% + strongly agree 9.6%) (Figure 1).

Figure 1  Distribution of attitudes toward AI use, ranging from “strongly disagree” to “strongly agree”.

Table 1 Social-demographic characteristics of the participants (n = 2736), mean ± SD.

Characteristic	n (%)
Sex
Men	1129 (41.3)
Women	1607 (58.7)
Grade
Freshman	592 (21.6)
Sophomore	688 (25.1)
Junior	968 (35.4)
Senior	419 (15.3)
Fifth-year	69 (2.5)
Academic field
Science	490 (17.9)
Engineering	565 (20.7)
Literature	473 (17.3)
Medicine	610 (22.3)
Others	598 (21.9)
Only children
Yes	2012 (73.5)
No	724 (26.5)
Residence
City	1372 (50.1)
Countryside	1364 (49.9)
Age	20.49 ± 1.46
Stress	3.73 ± 3.27
Depression	3.85 ± 3.46
Anxiety	4.21 ± 3.44
Loneliness	1.76 ± 1.52
GSES	14.72 ± 6.27
NSSI	0.15 ± 0.49
SI	0.22 ± 0.54

GSES: General Self-Efficacy Scale; NSSI: Non-suicidal self-injury; SI: Suicidality.

The linear regression results from Model 2 (Figure 2) indicated that the four predictors—increased anxiety, job insecurity, self-doubt, and increased procrastination—were significantly or near-significantly and positively associated with all six mental-health outcomes. Among these predictors, increased procrastination yielded the largest effects, with standardized coefficients of 0.568 for anxiety, 0.511 for depression, 0.472 for stress, 0.211 for loneliness, 0.029 for suicidal ideation, and 0.029 for NSSI (all P < 0.001). Self-doubt also exerted strong effects, with coefficients of approximately 0.45 to 0.50 for anxiety, depression, and stress, 0.221 for loneliness, and around 0.036 for both suicidal ideation and NSSI. Job insecurity showed moderate effects, with coefficients of 0.423 for anxiety, 0.417 for depression, 0.374 for stress, 0.139 for loneliness, 0.036 for suicidal ideation, and 0.026 for NSSI. Increased anxiety, though comparatively smaller in magnitude, still demonstrated significant effects, with coefficients of 0.319 for anxiety, 0.372 for depression, 0.313 for stress, 0.118 for loneliness, 0.033 for suicidal ideation, and 0.020 for NSSI. Overall, procrastination emerged as the strongest and most consistent predictor across all outcomes.

Figure 2 Linear regression results for the associations among AI attitudes, self-efficacy, and psychological outcomes. Model 1: Psychological outcomes regressed on AI attitudes and basic demographic covariates. Model 2: Psychological outcomes regressed on AI attitudes, basic demographic covariates, and self-efficacy.

Based on the dose-response analysis in Figure 3, nonlinear evidence varies by predictor and outcome: Increased anxiety shows significant nonlinearity for anxiety, depression, stress, and SI (P = 0.002, 0.025, 0.035, and 0.002, respectively) and for NSSI (P = 0.021), but not for loneliness (P = 0.127). Job insecurity shows no significant nonlinearities across outcomes (all P > 0.05). Self-doubt exhibits significant nonlinearity for SI and NSSI (P = 0.045, 0.013); other outcomes are non-significant (loneliness 0.077). Increased procrastination has a nonlinear effect on NSSI (P = 0.028); none of the other outcomes show significant nonlinearities (loneliness is borderline at P = 0.050). Overall, most relationships appear linear, with notable dose-response nonlinearities, particularly for increased anxiety across several outcomes.

Figure 3  Nonlinear relationships between AI attitudes and mental health outcomes, estimated using restricted cubic splines.

We constructed a mediation model (Figure 4) in which self-efficacy was specified to mediate the relationships between four predictors—increased anxiety, self-doubt, increased procrastination, and job insecurity—and six mental-health outcomes (anxiety, depression, stress, loneliness, SI, and NSSI). However, job insecurity was excluded from all mediation models based on strict statistical criteria: Preliminary Pearson correlation tests demonstrated a non-significant bivariate association between job insecurity and self-efficacy (P > 0.05). A significant correlation between the independent variable and the mediator is a necessary prerequisite for mediation effects to exist; without a meaningful linkage between job insecurity and self-efficacy, self-efficacy cannot carry any indirect effect of job insecurity on mental health outcomes, and conducting mediation analysis for this dimension would produce uninterpretable null indirect effects. In contrast, increased anxiety, self-doubt, and increased procrastination all yielded significant negative correlations with self-efficacy (all P < 0.001), satisfying the core statistical precondition for mediation testing. Subsequent mediation analyses suggest indirect effects through self-efficacy across several paths. This indicates a partially mediated role of self-efficacy in transmitting the impact of the predictors to the outcomes. Overall, all models show that 2 < CMIN/DF < 3; CFI > 0.95; TLI > 0.94; RMSEA < 0.050; and NFI > 0.95, supporting the plausibility of the mediation structure. For increased anxiety (Figure 4A), the mediation proportion was 40.30% for anxiety, 38.36% for depression, 42.86% for stress, 27.08% for suicidal ideation, and 31.03% for NSSI. For self-doubt (Figure 4B), mediation proportions were 17.50% (anxiety), 16.62% (depression), 17.10% (stress), 18.03% (suicidal ideation), and 11.01% (NSSI). For increased procrastination (Figure 4C), mediation proportions ranged from 11.89% to 15.25% across all five outcomes, representing the weakest mediating magnitude among the three predictors.

Figure 4 Path analysis results showing the mediating role of self-efficacy in the relationships between AI attitudes and mental health outcomes. A: Self-efficacy mediates the relationship between increased anxiety and mental health outcomes; B: Self-efficacy mediates the relationship between self-doubt and mental health outcomes; C: Self-efficacy mediates the relationship between increased procrastination and mental health outcomes. All models demonstrated good fit, with 2 < CMIN/DF < 3, CFI > 0.95, TLI > 0.94, RMSEA < 0.050, and NFI > 0.95. All models were adjusted for basic demographic variables. ^cP < 0.001. NSSI: Non-suicidal self-injury.

DISCUSSION

This study investigated the associations between college students' multi-dimensional attitudes toward AI technology, self-efficacy, and mental health outcomes, as well as the mediating role of self-efficacy in these relationships. The findings provide novel insights into the mental health implications of AI exposure among college students amid rapid global AI advancement. The results revealed that college students’ attitudes toward AI form a complex, multi-dimensional structure. Specifically, these attitudes showed marked differentiation: Only 9.5% of students reported AI-related anxiety on the dimension of technological anxiety, indicating that the overall acceptance and adaptability of current college students to AI technology are relatively high. In contrast, the proportion of students endorsing employment insecurity reached as high as 50.4%, making it the most prominent negative psychological response in AI-related cognition. This reflects that perceived career competition linked to AI technological iteration is correlated with widespread concern among college students. Meanwhile, 25.6% of students reported self-doubt associated with peers’ AI proficiency, and 36.8% reported procrastination correlated with the convenience brought by AI. These findings suggest that AI technology is correlated with individual psychological states via external occupational perceptions, and is linked to negative psychological and behavioral tendencies in self-cognition and daily routines through social comparison and shifts in behavioral habits.

The linear regression results demonstrated that all four dimensions of AI attitudes were significantly associated with mental health outcomes, with procrastination showing the strongest and most consistent effects across all outcomes (anxiety, depression, stress, loneliness, suicidal ideation, and NSSI). This finding suggests that AI-related procrastination shows stronger cross-sectional associations with mental health outcomes relative to cognitive or emotional AI attitudes. Procrastination is correlated with task backlogs[19], missed deadlines, and feelings of failure, which are further correlated with elevated psychological distress[20,21], including anxiety, depression, and stress. Additionally, procrastination may exacerbate feelings of loneliness[22], as delayed task completion may reduce social engagement and opportunities for peer interaction.

Self-doubt and job insecurity also exerted significant effects on mental health outcomes. Self-doubt showed strong cross-sectional correlations with anxiety, depression, and stress; persistent self-doubt is correlated with reduced perceptions of self-worth and competence, which are linked to negative emotional states[23]. Job insecurity[24], as a future-oriented concern, is correlated with higher anxiety, depression, and stress, as perceptions of career uncertainty are associated with lower life satisfaction and self-identification. Although AI-related anxiety had a smaller effect size, it was still significantly associated with mental health outcomes, indicating that even a small proportion of students who experience AI-related anxiety may face elevated psychological distress, which should not be overlooked. Notably, the regression coefficients linking AI-related attitudes to continuous suicidal ideation and NSSI scores were modest (β = 0.02–0.05), illustrating the distinction between statistical significance and practical significance. The limited magnitude may be partly attributable to the use of single-item measurement for AI-related perceptions. In addition, self-harm-related symptoms are shaped by a broad range of risk factors[25,26], and AI-related attitudes merely constitute one set of psychosocial predictors. Nevertheless, subtle increases in symptom scores at the population level still carry public health implications for university student mental health.

The RCS analysis (Figure 3) uncovered heterogeneous nonlinear relationships that add critical nuance to the mental health impacts of AI attitudes. AI-related anxiety displayed significant nonlinear cross-sectional correlations with most outcomes: The strength of association with poor mental health was weak at low levels and strengthened sharply beyond a critical threshold, with chronic technological anxiety correlated with sustained psychological tension. Nonlinear associations between self-doubt, AI-related procrastination, and severe mental health outcomes were also observed; elevated self-harm indicators were correlated disproportionately with high levels of these negative attitudes, alongside diminished self-concept and accumulated task delays. In stark contrast, job insecurity exhibited stable linearity across all outcomes: AI-driven career worry appears to be a chronic stressor for students, with distress risk rising proportionally to perceived insecurity and no clear threshold effect[27,28]. These findings challenge the simplistic linear assumption of technology-attitude-mental health links, demonstrating that different dimensions of AI attitudes operate through distinct modes.

The mediation analysis (Figure 4) confirmed that self-efficacy plays an important mediating role in the relationships between AI attitudes and mental health outcomes. This finding supports SCT[10], which emphasizes that self-efficacy is a key factor influencing individuals’ emotional and behavioral responses to environmental stimuli[29,30]. Specifically, negative AI attitudes are correlated with lower self-efficacy: Career displacement worries are associated with reduced confidence in future AI-era competitiveness; social-comparison self-doubt is correlated with diminished self-efficacy; and AI-related procrastination is linked to feelings of failure and weaker task confidence[31]. Lower self-efficacy[29], in turn, shows consistent cross-sectional correlations with elevated mental health distress. Meanwhile, the varying proportions of mediation effects across the three AI attitude dimensions further elaborate the heterogeneous partial mediation mechanism. Increased anxiety exhibited the largest mediation proportion (> 27% for all outcomes), which implies that worries triggered by AI technology directly undermine students’ self-perceived competence, and that reduced self-efficacy substantially translates technological anxiety into psychological distress. In contrast, self-doubt and increased procrastination yielded only small-to-moderate mediation proportions (mostly below 20%). Although social comparison and delayed task completion can erode self-efficacy, their direct adverse impacts on mental health remain dominant[32], so self-efficacy accounts for only a minor share of the total association. Such differentiated mediation magnitudes highlight that distinct negative AI-related cognitions operate through divergent cognitive pathways, which extends the theoretical explanation of SCT in AI contexts.

Notably, job insecurity failed to correlate with self-efficacy and was therefore excluded from mediation modelling, which carries distinct theoretical implications rooted in SCT[10]. Unlike AI-related anxiety, self-doubt, and procrastination—intrapersonal cognitive and behavioral perceptions that directly undermine students’ self-judgment of academic and AI operational competence—job insecurity refers to distal, future-oriented worries about career competition, which are psychologically decoupled from individuals’ immediate self-evaluation of personal capabilities. College students tend to separate long-term labor market risks triggered by AI from their current self-confidence in learning and AI proficiency, which explains the absence of a correlation between job insecurity and self-efficacy and the lack of an indirect mediating pathway. This heterogeneous mechanistic pattern across four AI attitude dimensions further highlights that different negative AI-related cognitions and behaviors affect mental health through differentiated psychological pathways.

This study has several limitations that guide future research. A major limitation is the use of single items to assess AI attitudes, SI, and NSSI, which leads to limited reliability and construct validity and potential measurement bias. Although single items are widely used for large-sample rapid screening, their precision is inferior to that of validated multi-item scales. Future work should adopt standardized multi-dimensional scales. The cross-sectional design precludes causal inferences; longitudinal studies are needed to track dynamic changes in variables and to verify causal directions. The convenience sampling from three Jiangxi universities limits generalizability; future research should adopt stratified random sampling across Chinese regions and institutional types and include cross-cultural samples. We excluded 600 non-AI users because the core items were designed for respondents with AI usage experience; however, this exclusion may introduce selection bias and narrow the scope to which our results can be generalized. Some theoretically relevant confounders, including prior mental health history, AI literacy, social media use, and academic performance, were not assessed. Although available demographic factors were adjusted for, residual confounding may still exist. Only self-efficacy’s mediation was examined; expanding the framework to include moderators or mediators (e.g., coping style, social support, and AI literacy) will help build a more comprehensive influence model. Finally, future research should explore positive AI attitudes (enthusiasm and empowerment) to provide a balanced understanding of AI’s psychological impacts.

CONCLUSION

This study explored the links between four dimensions of university students’ AI attitudes and mental health outcomes (anxiety, depression, stress, loneliness, suicidal ideation, and NSSI), as well as the nonlinear relationships and the mediating role of self-efficacy in these associations, based on a sample of 2736 Chinese undergraduate students. Key results showed that students’ AI attitudes presented a differentiated structure (9.5% AI-related anxiety, 50.4% job insecurity, 25.6% self-doubt, and 36.8% AI-induced procrastination), with all four negative dimensions positively predicting poor mental health—AI-induced procrastination being the strongest and most consistent predictor. Heterogeneous nonlinear relationships were identified: AI-related anxiety showed nonlinearity for most mental health outcomes, self-doubt and AI-induced procrastination only for severe outcomes (suicidal ideation/NSSI), while job insecurity exhibited a stable linear association across all outcomes. Additionally, self-efficacy played a partial mediating role in the AI attitude-mental health link, acting as a core cognitive mechanism that translates negative AI perceptions into psychological distress.

These correlational findings elucidate the complex, multi-dimensional cross-sectional associations between AI attitudes and university students’ mental health, moving beyond the simplistic linear association assumption and validating self-efficacy as a critical intermediary. Theoretically, this research enriches the empirical literature on AI and young adult mental health from a Chinese context and extends the application of SCT to the AI era. Practically, it provides targeted implications for university mental health practice and AI literacy education: We define students selecting “Agree/Strongly Agree” on AI attitude items as high-risk subgroups (prevalence ranging from 9.5% to 50.4%). Based on these cross-sectional associations and differentiated mediation proportions, targeted mental health interventions may prioritize AI-related procrastination and perceived job insecurity through vocational AI literacy training, establish early warning systems for high levels of AI-related anxiety and self-doubt to prevent the nonlinear escalation of severe psychological distress, and adopt self-efficacy enhancement as a cross-cutting protective strategy. All proposed interventions can be integrated into existing campus psychological service systems with low implementation cost and high feasibility.

ACKNOWLEDGEMENTS

We would like to thank all participants who generously shared their time to participate in this survey.