The mediating role of AI anxiety between AI literacy and attitudes towards AI in nursing students: a cross-sectional study

Abstract

Background

AI is a key driver in the Fourth Industrial Revolution, significantly impacting healthcare. Nursing students’ acceptance and readiness to use AI are crucial for its successful integration. Understanding their perceptions is vital to prepare them for the intelligent healthcare era.

Aim

To examine the mediation effect of AI anxiety between AI literacy and attitudes towards AI in college nursing students.

Design

A cross-sectional study.

Methods

Nursing students from 10 universities across 7 provinces in China were recruited using convenience sampling from May to June 2025. The data collection tools included the Self-Designed Demographic Questionnaire, the General Attitudes Towards Artificial Intelligence Scale, the Artificial Intelligence Literacy Scale and the Artificial Intelligence Anxiety Scale. Data were collated and analyzed using SPSS Statistics 26.0 software and AMOS 29.0 software.

Results

The study found that nursing students possess moderate levels of AI literacy and AI anxiety, while generally maintaining positive attitudes towards AI. Mediation analysis indicated that AI anxiety statistically mediated the association between AI literacy and nursing students’ attitudes towards AI.

Conclusions

AI literacy was significantly associated with nursing students’ attitudes towards AI, and AI anxiety statistically mediated this association. Therefore, nursing education may benefit from a dual-path approach, designed to simultaneously enhance cognitive competencies (AI literacy) and manage affective responses (AI anxiety).

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12909-026-10522-1.

Introduction

Artificial intelligence (AI) serves as a crucial catalyst for the Fourth Industrial Revolution, with its applications permeating various societal sectors, notably the healthcare industry [1–5]. As a vital component of the healthcare system, the nursing sector is increasingly impacted by the integration of AI technologies. For example, AI-powered nursing tools, including those used for predictive analytics, clinical decision support, electronic reminders, and advanced interactions, enable nurses to implement preventive strategies early [6]. Nursing students’ acceptance and willingness to adopt AI technologies will have a profound impact on the successful dissemination and implementation of AI in clinical practice. As the foundation of the future healthcare workforce, nursing students must develop proficiency in AI technologies to meet the demands of the intelligent healthcare era. Therefore, it has become imperative to understand nursing students’ perceptions and readiness regarding AI [7].

The technology acceptance model (TAM) is a widely used theoretical framework for predicting the adoption and use of new technologies in specific healthcare contexts [8], including nursing [9]. In the nursing student population, several scholars have investigated the relationship between attitudes towards AI and the intention to use it, employing the TAM as their theoretical framework [9–11]. From the perspective of the TAM, attitudes are antecedents of behavioural intention, suggesting that understanding the factors shaping nursing students’ attitudes towards AI may facilitate their future adoption of AI technologies. However, general attitudes towards AI are both positive and negative [12]. Some authors found that although nursing students generally hold positive attitudes towards AI, the majority lack substantial professional knowledge and practical skills for applying AI technologies [13, 14]. This suggests low AI literacy in this population, which may influence attitudes and behavioural intentions.

The European Union’s Digital Education Action Plan (2021–2027), initiated in September 2020, seeks to improve student learning outcomes by integrating AI into educational systems while providing support for educators and school administrators [15]. The promotion of digital education in the healthcare sector is equally imperative. For instance, the U.S. National Science Foundation has prioritised the application of AI technologies to advance both education and public health initiatives [16]. However, researchers from South Korea [7] and Turkey [17] reported that only 3.9% and 11.9% of nursing students, respectively, had received formal AI training. A meta-analysis of 8,491 students across medical, dental, and nursing disciplines further revealed that the majority had limited AI knowledge, with significant regional disparities [14]. These differences underscore a global digital divide in access to AI skills training. Against this backdrop, AI literacy has garnered increasing research attention as a key metric for assessing a nursing student’s ability to adapt to the age of AI.

AI literacy is the ability to accurately identify, use, and evaluate AI-related products while adhering to ethical standards [18, 19]. Many authors have found that healthcare students exhibit relatively low levels of AI literacy [19–21]. One study found that AI literacy can subsequently influence nursing students’ professional development [22]. Therefore, enhancing AI literacy is essential for adapting to the data-driven digital transformation of the nursing profession and is a core competency for future career development.

The ABC model of attitudes, also known as the tripartite model, was proposed by Hovland and Rosenberg in 1960 as a core theoretical framework in social psychology for deconstructing the composition of attitudes [23]. This model conceptualises an attitude as comprising three interrelated components: affect (emotion), behavioural tendency, and cognition [24]. AI anxiety is a psychological reaction of fear and unease towards uncontrollable AI systems, characterised by apprehensive or fearful emotions that inhibit individuals from interacting with AI [25]. Grounded in the ABC model, AI literacy can be regarded as the cognitive basis of attitude formation, whereas AI anxiety represents its affective component; because these components are interrelated, insufficient AI literacy may be associated with AI anxiety, which in turn may relate to less positive attitudes towards AI. Therefore, investigating AI anxiety in nursing students and how it arises from AI literacy can help us better understand their attitudes towards AI and more accurately predict their AI-related behavioural intentions.

Recently, scholars have begun to explore the negative correlation between AI anxiety and attitudes towards AI among nursing students [26, 27], as well as the positive contribution of AI literacy to attitudes towards AI [19, 28]. Moreover, researchers suggest that low AI literacy can cause uncertainty and a lack of control when using AI, potentially triggering AI anxiety [20, 29]. This negative affective state, in turn, may undermine positive attitudes towards AI adoption. Accordingly, the ABC model suggests that AI literacy may influence attitudes towards AI both directly through cognitive evaluation and indirectly through AI anxiety, providing a theoretical rationale for examining the mediating role of AI anxiety in this study. However, researchers have focused on direct pairwise relationships between these variables, while the underlying mechanism of how AI literacy influences attitudes towards AI remains underexplored. In summary, this study draws on the TAM and ABC models as theoretical frameworks to explore the potential mechanisms linking AI literacy, AI anxiety, and attitudes towards AI. This may contribute to a more comprehensive understanding of the psychological pathways through which nursing students’ attitudes towards AI are formed. The following hypotheses are proposed:

• H1: AI literacy is positively associated with nursing students’ attitudes towards AI.

• H2: AI literacy is negatively associated with nursing students’ AI anxiety.

• H3: AI anxiety is negatively associated with nursing students’ attitudes towards AI.

• H4: AI anxiety statistically mediates the association between AI literacy and attitudes towards AI.

The theoretical framework of this study is presented in Fig. 1. This model helps elucidate the intrinsic pathway through which the affective factor (AI anxiety) operates within the relationship between AI literacy and attitudes towards AI. The findings are expected to provide a precise theoretical basis for developing targeted intervention strategies to promote AI use and practical application among nursing students.

Fig. 1.

The theoretical framework of this study

Methods

Study design and setting

This study used a cross-sectional design and was conducted in accordance with the STROBE guidelines (see Table S1).

Participants and sample

A convenience sampling method was employed to recruit nursing students from 10 universities and colleges across seven provinces in China during May–June 2025. These institutions were selected to reflect diversity in geographic location (including seven provinces across northern, central, southern, and eastern China) and institutional type (comprehensive universities, medical universities, and vocational colleges), aiming to capture a broad range of nursing education contexts. Two inclusion criteria were used: (a) full-time enrolment in a nursing program at the associate, bachelor’s, or master’s degree level; and (b) provision of informed consent and voluntary participation in the study. The exclusion criterion was students who were on a leave of absence during the investigation period. Based on the Kendall sample size calculation method [30], the formula used was n = [Max(number of variables) × (5–10)] × [1 + (10%–30%)]. In this study, the maximum number of variables for statistical analysis was 21. These included: (a) eight items from the self-designed demographic questionnaire assessing nursing students’ characteristics and AI engagement; and (b) three total scale scores and ten subscale scores derived from three validated scales. Using a multiplier of 10 and accounting for a 10% sample dropout rate, the required sample size was determined to be 231. Additionally, the sample size needed for structural equation modelling (SEM) should exceed 200 [31]. Therefore, this study requires a sample of 231 students.

Data collection

This study used an online questionnaire survey administered via Questionnaire Star. An introductory statement was provided at the beginning of the questionnaire to convey the significance of the research to the participating nursing students, employing consistent language throughout. Participants were requested to complete a structured questionnaire, which included the self-designed demographic questionnaire, the artificial intelligence literacy scale (AILS), the artificial intelligence anxiety scale (AIAS), and the general attitudes towards artificial intelligence scale (GAAIS).

Self-designed demographic questionnaire

The researchers designed a self-administered questionnaire with eight items to investigate the characteristics of nursing students and their engagement with AI. The instrument consisted of a section for sociodemographic information (age, gender, residence, educational background, student leadership experience, and family structure [only-child status]), as well as an AI competency assessment of their AI training exposure and practical AI application experience.

Artificial intelligence literacy scale

The AILS, developed by Wang et al., is a 7-point Likert-type scale comprising 12 items used to assess the level of AI literacy [32]. This instrument is structured into four subdimensions: awareness, usage, evaluation, and ethics. Participants are asked to express their level of agreement or disagreement with the research questions, with response options ranging from 1 (strongly disagree) to 7 (strongly agree). The scale includes three reverse-scored items. Scores on the scale range from 12–84, with higher scores indicating greater AI literacy. In this study, the overall Cronbach’s α coefficient for the scale was 0.956, with subscale α coefficients ranging from 0.851 to 0.954, indicating good internal consistency reliability. The KMO value was 0.909, and Bartlett’s sphericity test yielded significant results (p < 0.001), indicating the data’s suitability for factor analysis.

Artificial intelligence anxiety scale

The AIAS, developed by Wang and Wang in 2019 [33], assesses AI-induced anxiety as an affective state, capturing negative emotional responses towards AI. Participants are required to reflect on their current experiences using a 7-point Likert-type scale. This scale comprises 21 items divided into four subdimensions: learning (eight items), job replacement (six items), sociotechnical blindness (four items), and AI configuration (three items). The scale’s scores range from 21 to 147, with higher scores indicating greater levels of AI anxiety. Wang and Wang reported high reliability coefficients for each dimension: 0.974 for learning, 0.917 for job replacement, 0.917 for sociotechnical blindness, and 0.916 for AI configuration. In this study, the reliability and validity testing of the AIAS yielded an overall Cronbach’s α coefficient of 0.983, with sub-dimension alphas ranging from 0.955 to 0.979. The KMO value was 0.954, and Bartlett’s test of sphericity was significant (p < 0.001), indicating high internal consistency among items. Factor analysis further confirmed the scale’s strong construct validity.

The general attitudes towards artificial intelligence scale

The GAAIS was developed by Schepman and Rodway and revised by Huang et al. [12]. This scale was used to assess attitudes towards AI, reflecting individuals’ evaluative judgments and behavioural intentions regarding the use of AI technologies. This 5-point Likert scale assesses attitudes towards AI through negative and positive subdimensions and includes a total of 15 items. Negative attitude items are reverse scored to ensure that higher scores on the negative subdimension reflect more positive attitudes towards AI. The original internal consistency reliability was 0.88 for the positive GAAIS and 0.83 for the negative GAAIS. For this study, the GAAIS showed excellent scale properties, with strong internal consistency (Cronbach’s α = 0.869), sampling adequacy (KMO = 0.896), and a significant factor structure (Bartlett’s test, p < 0.001). These findings confirm the scale’s reliability and validity for assessing attitudes towards AI in this population, as well as the suitability of the data for factor analysis.

Validity and reliability

Table 1 presents the summary results of the reliability and validity indices. McDonald’s ω coefficients ranged from 0.887 to 0.984, and Cronbach’s α coefficients for all dimensions of the three scales exceeded 0.7 (ranging from 0.851 to 0.983), indicating good internal consistency. Confirmatory factor analysis (CFA) was conducted using AMOS to evaluate the structural validity of the scales. Following the classic study by Hu & Bentler, we used the following fit indices: Chi-square to degrees of freedom ratio (χ²/df, < 3 = good fit, 3–5 = acceptable), Root mean square error of approximation (RMSEA, < 0.05 = good fit, 0.05–0.08 = acceptable), Standardized Root Mean Square Residual (SRMR, < 0.05 = good fit, 0.05–0.08 = acceptable), Comparative fit index (CFI, > 0.9 = good fit), and Tucker‑Lewis index (TLI, > 0.9 = good fit) to assess model fit. The results showed that the χ2/df, RMSEA, and SRMR values for all three scales were within acceptable ranges, and the model fit indices CFI and TLI both exceeded 0.9, indicating good model fit. In addition, the average variance extracted (AVE) for each subscale exceeded the recommended threshold of 0.5 (ranging from 0.587 to 0.920), and the composite reliability (CR) exceeded the recommended threshold of 0.7 (ranging from 0.809 to 0.978). All of these indicators together demonstrate that the scales provide satisfactory validity evidence.

Table 1.

Reliability, validity indices, and model fit statistics

Study Variables	N	ω	α	CR	AVE	χ2/df	RMSEA	SRMR	CFI	TLI
AILS	12	0.963	0.956			3.517	0.074	0.023	0.982	0.973
Awareness	3	0.934	0.894	0.876	0.705
Usage	3	0.959	0.935	0.946	0.854
Evaluation	3	0.971	0.954	0.955	0.876
Ethics	3	0.910	0.851	0.809	0.594
AIAS	21	0.984	0.983			3.776	0.078	0.032	0.976	0.964
Learning	8	0.982	0.979	0.978	0.847
Job replacement	6	0.969	0.962	0.959	0.797
Sociotechnical blindness	4	0.967	0.955	0.954	0.837
AI configuration	3	0.981	0.972	0.972	0.920
GAAIS	15	0.887	0.869			3.400	0.073	0.075	0.960	0.948
Positive	8	0.939	0.924	0.920	0.594
Negative	7	0.936	0.917	0.906	0.587

Abbreviations: AILS Artificial Intelligence Literacy Scale, AIAS Artificial Intelligence Anxiety Scale, GAAIS General Attitudes Towards Artificial Intelligence Scale, N Number of items, ω McDonald’s omega for reliability statistics, α Cronbach’s alpha for reliability statistics, AVE Average variance extracted, CR Composite reliability, χ2/df Chi-square to degrees of freedom ratio, RMSEA Root mean square error of approximation, SRMR Standardized Root Mean Square Residual, CFI Comparative fit index, TLI Tucker-Lewis index

Data analysis

Data processing was carried out with SPSS 26.0 and AMOS 29.0. SPSS version 26.0 was used for descriptive statistics, group comparisons, correlation analysis, and multivariate analysis. Normally distributed continuous variables were expressed as mean ± SD, and categorical variables as frequencies and percentages. Independent sample t‑test and one‑way ANOVA were applied to compare AILS, AIAS, and GAAIS scores across sociodemographic factors and AI‑related experience levels. Pearson correlation analysis was conducted to examine relationships among the study variables. Multiple linear regression was separately performed to identify factors significantly associated with AI literacy, AI anxiety, and attitudes towards AI. The variables that remained statistically significant in any of these regression models were selected based on widely accepted theories of the digital divide and TAM [8, 34] and entered as raw observed covariates in the subsequent mediation model.

AMOS 29.0 was used to conduct reliability and validity testing and to construct SEM to examine hypothesized mediating paths among variables. In our SEM model, AI literacy and AI anxiety were defined as two latent variables; each latent construct was measured via its own set of four subscale composite scores. Attitudes towards AI were treated as an observed endogenous outcome variable. Its positive and negative subdimensions represent two complementary facets of a single attitudinal construct, so we adopted its aggregated total score to enhance model parsimony and stability. A structural residual term was attached to this composite total score in AMOS to capture unexplained variance in the outcome after controlling for all predictors and covariates. Maximum likelihood estimation was used for SEM analyses. Multivariate normality of continuous indicators was examined using Mardia’s test. Where multivariate non-normality was present, the bias-corrected percentile bootstrap method (5,000 resamples, 95% CI) was used to obtain robust significance tests for path coefficients and mediating associations. Model fit was primarily evaluated using the following key indices: χ²/df, CFI, TLI, SRMR, and RMSEA. A model was deemed to have an acceptable fit if the χ²/df was less than 5, the CFI and TLI were greater than 0.9, the SRMR was less than 0.08, and the RMSEA was less than 0.08 [35]. The rules for model respecification stipulated three simultaneous criteria for adding covariances: modification index (MI) > 10, theoretical and literature justification, and significant preliminary chi-square or correlation results. Only two covariance types were allowed: covariances between exogenous demographic controls and residual covariances of measurement errors within the same latent construct. No untheorized structural paths were added. The threshold for statistical significance was established at p < 0.05. The reason for using SEM is the increasing importance of data validity and the avoidance of measurement errors in recent years. Conventional statistical methods that rely on averaged indicators practically ignore the measurement errors of each indicator. By introducing the concept of latent variables, SEM is in a privileged position in this regard [36].

Results

Characteristics of the sample

Ultimately, 19 invalid questionnaires were screened out. A total of 455 valid questionnaires (no missing values) were collected, with a valid response rate of 95.99% (455/474 × 100%). The study encompassed a sample of 455 nursing students, whose sociodemographic characteristics and engagement with AI are detailed in Table 2.

Table 2.

Sociodemographic and artificial intelligence usage characteristics of nursing students (n = 455)

Characteristics	Categories	n(%) or Mean ± SD
Age(years)		20.53 ± 2.25
Gender	Male	68(14.9)
	Female	387(85.1)
Residence	Countryside	343(75.4)
	City	112(24.6)
Educational backgrounds	Master’s degree in progress	14(3.1)
	Bachelor’s degree in progress	167(36.7)
	Associate’s degree in progress	274(60.2)
Student leadership experience	Yes	253(55.6)
	No	202(44.4)
Only child status	Yes	64(14.1)
	No	391(85.9)
AI training exposure	Yes	46(10.1)
	No	409(89.9)
practical AI application experience	Yes	407(89.5)
	No	48(10.5)

Abbreviations: AI Artificial intelligence

Scores of AI literacy, AI anxiety, and general attitudes towards AI

The findings revealed that the mean total AILS score among nursing students in college was 56.77 (SD = 6.99). Participants’ AI anxiety, as assessed by the AIAS, had a mean total score of 63.91 (SD = 28.17). The mean total score on the GAAIS was 51.25 (SD = 7.72). Table 3 shows the scores for the subscales of AI literacy, AI anxiety, and attitudes towards AI.

Table 3.

AILS, AIAS and GAAIS scores of nursing students(n = 455)

Variable	X ± SD	(Min–max)
AILS
Awareness	13.94 ± 1.35	8–15
Usage	13.60 ± 1.48	6–15
Evaluation	15.87 ± 3.80	3–21
Ethics	13.35 ± 1.43	9–15
Total score	56.77 ± 6.99	30–66
AIAS
Learning	22.36 ± 11.19	8–56
Job replacement	19.48 ± 8.73	6–42
Sociotechnical blindness	12.98 ± 5.90	4–28
AI configuration	9.08 ± 4.62	3–21
Total score	63.91 ± 28.17	21–147
GAAIS
Positive	27.84 ± 5.84	8–40
Negative	23.42 ± 5.64	7–35
Total score	51.25 ± 7.72	15–75

Abbreviations: AILS Artificial Intelligence Literacy Scale, AIAS Artificial Intelligence Anxiety Scale, GAAIS General Attitudes Towards Artificial Intelligence Scale

Associations between sociodemographic factors and AI experience and scores on the AILS, AIAS, and GAAIS

Firstly, analysis of the associations among sociodemographic factors, AI engagement, and AILS scores showed that age, residence, and previous AI exposure (both training exposure and practical application experience) were significantly associated with higher AILS total and specific subscale scores (p < 0.05). Secondly, analysis of the relationships among sociodemographic characteristics, AI engagement of the nursing students, and AIAS scores revealed distinct patterns (p < 0.05). Age was positively correlated only with the job replacement and sociotechnical blindness subscales. In contrast, leadership experience and AI training exposure were associated with significantly lower total AIAS scores and lower scores on multiple subscales, while urban residence showed no significant association with AIAS scores. Practical AI experience was associated with significantly lower scores on the AI configuration subscale. Significant differences were also observed across educational levels (associate, bachelor’s and postgraduate) in the total score and several subscales. Thirdly, the analysis of factors associated with GAAIS scores showed that older age, urban residence, and practical AI experience were associated with significantly higher scores on the positive dimension and the total scale. Significant variations were also observed by educational level in these domains. Conversely, female students and those with leadership experience reported significantly higher scores on the negative dimension (p < 0.05 for all). Group comparisons of AI literacy, AI anxiety, and attitudes towards AI scores across sociodemographic and AI-related characteristics are summarised in Table 4 (only total scale scores), while complete subscale-level detailed results are provided in Supplementary Table S2.

Table 4.

Correlations and Group Comparisons of AI Literacy, AI Anxiety, and Attitudes Towards AI Scores by Sociodemographic and AI-Related Characteristics Among Nursing Students (n = 455)

Characteristic	AILS	AIAS	GAAIS
Age
r	0.104	0.072	0.114
pa	0.026	0.124	0.015
Gender
Male	57.5 ± 7.63	69.18 ± 34.28	51.50 ± 7.96
Female	56.63 ± 6.88	62.98 ± 26.90	51.21 ± 7.68
t	0.994	1.416	0.286
pb	0.321	0.160	0.775
Residence
Countryside	56.22 ± 6.70	64.52 ± 27.81	50.42 ± 6.98
City	58.45 ± 6.76	62.04 ± 29.27	53.80 ± 9.21
t	−2.952	0.809	−3.569
pb	0.003	0.419	< 0.001
Educational backgrounds
Master’s degree in progress	55.00 ± 10.77	66.71 ± 25.90	55.86 ± 7.79
Bachelor’s degree in progress	56.90 ± 5.77	70.62 ± 24.84	52.34 ± 7.83
Associate’s degree in progress	56.78 ± 7.44	59.67 ± 29.43	50.36 ± 7.50
F	0.476	0.33	6.115
pc	0.622	< 0.001	0.002
Student leadership experience
Yes	57.00 ± 7.51	59.15 ± 27.65	51.48 ± 7.99
No	56.48 ± 6.30	69.86 ± 27.74	50.97 ± 7.37
t	0.797	−4.100	0.697
pb	0.426	< 0.001	0.486
Only child status
Yes	57.53 ± 6.63	62.66 ± 31.45	51.38 ± 7.92
No	56.64 ± 7.05	64.11 ± 27.63	51.23 ± 7.69
t	0.943	−0.382	0.137
pb	0.346	0.702	0.891
AI training exposure
Yes	59.17 ± 5.99	51.26 ± 27.63	53.91 ± 10.12
No	56.50 ± 7.05	65.33 ± 27.90	50.95 ± 7.35
t	2.476	−3.245	1.927
pb	0.014	0.001	0.060
Practical AI application experience
Yes	57.04 ± 6.71	63.14 ± 27.78	51.74 ± 7.86
No	54.42 ± 8.80	70.38 ± 30.80	47.08 ± 4.69
t	2.002	−1.686	5.964
pb	0.050	0.092	< 0.001

The bold values represent statistically significant difference at p < 0.05

Abbreviations: AI Artificial intelligence, AILS Artificial Intelligence Literacy Scale, AIAS Artificial Intelligence Anxiety Scale, GAAIS General Attitudes towards Artificial Intelligence Scale

^aPearson’s correlation test; ^bIndependent samples t test; ^cOne-way variance analysis

Lastly, the results of the Pearson correlation analysis showed that scores of AI literacy and its subscales (awareness, usage, and evaluation) among nursing students were negatively correlated with AIAS and its dimensions (r = −0.275 to −0.104, p < 0.05). Additionally, AI literacy and all its subscale scores were positively correlated with GAAIS and its dimensions (r = 0.108 to 0.487, p < 0.05). In addition, AIAS and its dimensions were negatively correlated with GAAIS total scores and negative dimension scores (r = −0.310 to −0.513, p < 0.05). Of note, the GAAIS negative subscale is reverse scored, such that higher scores indicate more positive attitudes towards AI. Details are presented in Table 5.

Table 5.

Correlation between AILS, AIAS and GAAIS scores(n = 455)

Variables		1	2	3	4	5	6	7	8	9	10	11	12	13
AILS	1.Awareness	1.000
	2.Usage	0.619**	1.000
	3.Evaluation	0.623**	0.745**	1.000
	4.Ethics	0.533**	0.499**	0.643**	1.000
	5.Total score	0.772**	0.838**	0.953**	0.762**	1.000
AIAS	6.Learning	−0.197**	−0.244**	−0.265**	−0.148**	−0.264**	1.000
	7.Job replacement	−0.104*	−0.148**	−0.181**	−0.083	−0.167**	0.788**	1.000
	8.Sociotechnical blindness	−0.112*	−0.169**	−0.203**	−0.068	−0.182**	0.763**	0.939**	1.000
	9.AI configuration	−0.184**	−0.259**	−0.275**	−0.118*	−0.264**	0.724**	0.779**	0.865**	1.000
	10.Total score	−0.164**	−0.221**	−0.249**	−0.118*	−0.238**	0.920**	0.947**	0.945**	0.874**	1.000
GAAIS	11.Positive	0.322**	0.363**	0.428**	0.280**	0.429**	−0.085	0.010	0.010	−0.066	−0.040	1.000
	12.Negative	0.177**	0.232**	0.215**	0.108*	0.222**	−0.505**	−0.435**	−0.449**	−0.511**	−0.513**	0.096*	1.000
	13.Total score	0.372**	0.444**	0.481**	0.290**	0.487**	−0.433**	−0.310**	−0.320**	−0.424**	−0.405**	0.687**	0.658**	1.000

Abbreviations: AILS Artificial Intelligence Literacy Scale, AIAS Artificial Intelligence Anxiety Scale, GAAIS General Attitudes towards Artificial Intelligence Scale

^*p <.05

**p <.01

Analysis of the mediating role of AI anxiety between AI literacy and attitudes towards AI

An SEM was established using AMOS software, with AI literacy as the independent variable, attitudes towards AI in nursing students as the dependent variable, and AI anxiety as the mediating variable. Variables that were statistically significant in univariate analyses for AILS, AIAS, and GAAIS were entered into a multivariate linear regression analysis. The following variables were identified as meaningful and included as control variables in the mediation model: residence, student leadership experience, educational backgrounds, AI training exposure, and practical AI application experience.

The univariate normality of all continuous composite observed variables corresponding to latent constructs was evaluated via skewness and kurtosis. The skewness values ranged from −1.581 to 1.319, falling within the acceptable range of −3 to + 3. Kurtosis values ranged from −0.390 to 2.836, consistent with the cutoff criterion of −7 to + 7. Mardia’s multivariate skewness and kurtosis further revealed multivariate non-normality. To alleviate estimation bias induced by non-normal data, bias-corrected percentile bootstrap method with 5,000 resamples was adopted for path coefficient and mediating association significance testing.

Sequential model adjustments were performed based on preset standards, relying on modification indices, chi-square tests for categorical demographics, subdimension correlation results, and theoretical and literature evidence. Two demographic covariances and three within-construct residual covariances were added step by step. Supplementary Table S3 provides full documentation of all modification stages, stage-wise model fit indices, and supporting rationales for each added correlation pair. The final structural equation model (see Fig. 2) demonstrated an acceptable fit to the data. The overall fit of the final structural equation model was examined using multiple fit indices (see Table 6). The results showed that the model fitted the data well: χ²/df = 3.301 (< 5, acceptable given the large sample size), RMSEA = 0.071 (< 0.08), SRMR = 0.057 (< 0.08), CFI = 0.954 (> 0.90), TLI = 0.940 (> 0.90), goodness-of-fit index (GFI) = 0.936 (> 0.90), normed fit index (NFI) = 0.936 (> 0.90), incremental fit index (IFI) = 0.955 (> 0.90), relative fit index (RFI) = 0.916 (> 0.90), and parsimony goodness-of-fit index (PGFI) = 0.615 (> 0.50). All indices met the recommended criteria, supporting the structural model for further path analysis.

Fig. 2.

Final structural equation model with standardized coefficients. Notes: 1) Panels A, B, and C Separately illustrate the measurement models, core structural pathways, and exogenous covariates included in the model. All paths shown are statistically significant at p < 0.05. 2) Single-headed arrows denote standardized factor loadings (within measurement model) and standardized structural path coefficients (structural model and covariate paths). Double-headed curved arcs represent error correlations between observed indicators (Panel A) and correlation coefficients among exogenous covariates (Panel C). 3) Values adjacent to error terms (e1-e11) indicate standardized error variances. 4) Pale beige rectangles = observed indicators of latent variables; light blue-grey rectangles = exogenous covariates; grey ovals = latent variables (AI Literacy, AI Anxiety); grey rectangle = observed outcome variable (Attitudes towards AI)

Table 6.

Goodness-of-fit indices for the structural equation model (n = 455)

Model	χ2/df	RMSEA	SRMR	GFI	NFI	IFI	CFI	RFI	TLI	PGFI
Recommended value	≤ 5	≤ 0.08	≤ 0.08	≥ 0.90	≥ 0.90	≥ 0.90	≥ 0.90	≥ 0.90	≥ 0.90	> 0.50
Final model	3.301	0.071	0.057	0.936	0.936	0.955	0.954	0.916	0.940	0.615

χ²/df chi-square/degrees of freedom, RMSEA Root mean square error of approximation, SRMR Standardized root mean square residual, GFI Goodness-of-fit index, NFI Normed fit index, IFI Incremental fit index, CFI comparative fit index, RFI Relative fit index, TLI Tucker-Lewis index, PGFI Parsimony goodness-of-fit index

We tested the significance of the direct and indirect paths using a bias-corrected percentile bootstrap method with 5,000 resamples. The standardised path coefficients and their 95% confidence intervals (CIs) are presented in Fig. 2 (which also includes the five control variables and their paths) and Table 7. Effects were considered statistically significant if the 95% CI did not include zero. As presented in Table 7, the total effect of AI literacy on attitudes towards AI was significant (β = 0.507, 95% CI [0.442, 0.569]). The direct effect of AI literacy on attitudes towards AI was significant (β = 0.414, 95% CI [0.356, 0.487]), accounting for 81.66% of the total effect. The indirect effect via AI anxiety was significant (β = 0.093, 95% CI [0.054, 0.138]), accounting for approximately 18.34% of the total effect. Both direct and indirect pathways were statistically significant, indicating that AI anxiety statistically mediated the association between AI literacy and attitudes towards AI.

Table 7.

Direct, Indirect, and Total Effects of AI Literacy on Attitudes towards AI

Effect	Structural paths	Estimate	Bootstrap Std. Error	Bootstrap 95% CI		Ratio of the total effects
				Lower	Upper
Direct effect(a)	X → M	−0.263	0.060	−0.379	−0.146	——
Direct effect(b)	M → Y	−0.352	0.040	−0.428	−0.270	——
Total effect (c = c’ + a × b)	X → Y	0.507	0.032	0.442	0.569	——
Indirect effect(a × b)	X → M → Y	0.093	0.022	0.054	0.138	18.34%
Direct effect(c’)	X → Y	0.414	0.033	0.356	0.487	81.66%

X AI literacy, M AI anxiety, Y attitudes towards AI. All estimates are standardized path coefficients. The bootstrap 95% confidence intervals (CIs) are based on the bias-corrected percentile method with 5,000 resamples

Discussion

This study investigated the current levels of AI literacy, AI anxiety, and attitudes towards AI among nursing students. First, this study found that nursing students’ average score on the AI literacy indicator was 56.77 ± 6.99, a moderate score indicating a need for further enhancement in their AI-related knowledge and competencies. This result is consistent with previous findings [20, 37]. In the univariate analysis, we found that AI literacy among nursing students was linked to several key factors: age, residence, exposure to AI training, and practical AI application experience. Although the study cohort exhibited a limited age span, AI literacy still showed a significant positive correlation with age. It may suggest that this association reflects not chronological age itself, but the cumulative factors associated with it, particularly education advancement and the accumulation of relevant experiences. Previous research has shown that students in advanced years, notably 2nd-year and 3rd-year students, displayed a higher level of AI literacy than 1st-year students [21]. Our study also revealed significantly higher level among city students regarding use, evaluation, and overall AI literacy. This is consistent with Zhao et al. [34]. Therefore, this disparity of AI literacy may be related to the fact that city students typically have earlier and more frequent access to digital infrastructure, such as smartphones, computers, and high-speed internet. Such an environment fosters the subconscious accumulation of digital skills and technological awareness. Furthermore, students with prior AI training or hands-on experience demonstrated higher levels of AI literacy. This aligns with the previous findings [19, 38]. To help address these disparities, educators could implement two targeted strategies. First, educators may design tailored training based on students’ backgrounds. They may offer foundational courses for junior or rural students with limited digital exposure, and provide advanced modules on clinical AI applications and ethics for senior students. Second, integrating experiential learning, such as project-based activities and AI-enhanced clinical simulations, may help translate knowledge into practical skills.

Second, the average AI anxiety level among nursing students was 63.91 ± 28.17, which was lower than that reported previously [27]. This disparity may stem from a digital divide in access to AI training and education between regions and countries [14, 39]. A significant positive correlation was observed between age and anxiety concerning job replacement and sociotechnical blindness, indicating that senior students harbour stronger fears about AI’s potential to displace jobs and overlook humanistic aspects of care [40, 41]. Additionally, significant differences in AI anxiety scores were identified across students with varying educational backgrounds and student leadership experience. Graduate students typically develop more systematic and critical thinking skills, whereas those with leadership experience often have stronger decision-making abilities and a sense of responsibility. These students may be more likely to perceive the benefits of AI and to prospectively discern its complex ethical, social, and professional implications. This prudent cognitive appraisal may manifest as elevated anxiety. Practical experience only reduced anxiety related to AI configuration. This shows hands-on skills address operational concerns yet can barely assuage broader macro anxieties about ethics and occupational prospects. To bridge the variations in AI anxiety among nursing students across academic levels and leadership backgrounds, educators could implement the following targeted educational measures. For senior students and students with leadership experience concerned about job displacement and ethical risks, targeted AI ethics and nursing career courses can be provided. These courses may help transform their anxiety into critical engagement.

Third, the overall attitude of nursing students towards AI was comprehensively evaluated by the GAAIS in this study, with a mean total score of 51.25 ± 7.72. A total score above the theoretical midpoint suggests that the nursing student cohort hold generally positive and accepting attitudes towards AI. This finding aligns with those of many other scholars, indicating that nursing students, as well as most medical students, generally hold optimistic attitudes about AI [14, 26, 28, 39]. A nuanced analysis of the subscales reveals the structure of this attitude. The positive subscale (eight items) yielded a mean item score of 3.48, indicating a moderate-to-high level of agreement with favourable statements about AI. Crucially, the negative subscale (seven items), for which items were reverse scored, had a mean item score of 3.35. This signifies a relatively low level of agreement with negative statements concerning AI. A South Korean study using the same scale reported higher scores among nursing students in both the positive and negative dimensions than our findings. It is worth noting that this study had a limited sample size of 24 participants [28]. This attitudinal structure provides a highly advantageous starting point for nursing education. It suggests that the student cohort already possesses a psychological foundation for accepting AI technology and is not mired in widespread technophobia. Our results align with Gordon et al.’s scoping review of attitudes towards AI in clinical medicine, which shows health profession students hold generally optimistic yet cautious views of AI [42]. Furthermore, although negative attitudes are not dominant, they persist. This indicates that although students recognise the value of AI, they also maintain a certain degree of vigilance towards its potential risks. Moreover, our study identified significant correlations between attitudes towards AI and key demographic variables of age, residence, and educational level through correlation analyses. These findings highlight the value of differentiated nursing AI education to guide students’ attitudes toward AI. In response to variations in attitudes towards AI among nursing students across age, residence and educational level, the following considerations may inform educational practice. First, given students’ generally positive attitudes, systematic curricula may be used to shape their cognitive framework. Second, students with negative attitudes may be guided toward a more positive orientation. This may help foster their intention to use AI, support rational evaluation of technical risks, and enhance their willingness to adopt AI tools in future practice.

Lastly, this study revealed interconnections among AI literacy, AI anxiety, and attitudes towards AI through both correlation and mediation analyses. First, AI literacy was positively correlated with attitudes towards AI and negatively correlated with AI anxiety, which aligns with our hypotheses. This suggests higher AI literacy is linked to more positive attitudes and lower anxiety, possibly related to decreased uncertainty. Second, the negative correlation between AI anxiety and attitudes towards AI suggests that emotional factors play a role in technology acceptance. Most importantly, the mediation analysis indicated the partially mediated effect of AI anxiety. Specifically, AI anxiety statistically mediated the association between AI literacy and attitudes towards AI, revealing both direct and indirect correlational pathways. This pattern is consistent with the ABC model, suggesting that both cognitive (AI literacy) and affective (AI anxiety) components may be relevant to understanding nursing students’ attitudes towards AI. These findings have implications for nursing education. Educators may consider implementing a dual-path approach. On the cognitive path, they may strengthen AI literacy through systematic curriculum development. On the affective path, they may address AI anxiety by incorporating AI ethics education and showcasing successful examples of AI-assisted nursing care. Although AI anxiety showed a significant statistical association with attitudes towards AI, our findings suggest it may not serve as the primary pathway. Future research may explore other potential mediators, such as AI self-efficacy, perceived usefulness, and ethical awareness, to further elucidate the mechanisms underlying nursing students’ attitudes towards AI.

Limitations

However, this study also has the following limitations. Firstly, although participants were recruited from 10 universities across seven provinces to include varied geographic and institutional settings, convenience sampling relying on our research collaborative networks may introduce systematic selection bias and restrict external validity. Findings cannot be readily generalized to nursing students outside this network or to other populations and settings. While the sample covers diverse nursing education environments, future research should adopt probability sampling and conduct multicenter investigations including a broader range of institutions. Secondly, the cross-sectional design limits causal inference, precluding the establishment of causal pathways or the temporal ordering of the variables. Thus, the detected mediation model represents statistical associations rather than causal mechanisms. Further research should employ longitudinal or experimental designs to test causal relationships, thereby providing a richer foundation for intervention programs. Thirdly, the study relied on self-reported data, which could be influenced by response bias and personal interpretation. Future studies should complement self-reports with objective measures to enhance validity.

Conclusion

This study elucidates the current status of AI literacy, AI anxiety, and attitudes towards AI among nursing students in China. The key findings reveal significant associations of AI literacy and AI anxiety with students’ attitudes towards AI, with AI anxiety statistically mediating the association between AI literacy and attitudes towards AI. This suggests that differentiated, targeted AI education programs via a dual-path approach may be worthy of consideration for nursing educators. This involves systematically developing AI literacy courses to strengthen the cognitive path while addressing AI anxiety to manage the emotional path.

Authors’ contributions

Liu Sai: Conceptualization, Methodology, Data curation, Writing—Original draft preparation, Writing—Review & Editing, Funding acquisition. Wang Liping: Conceptualization, Methodology, Formal analysis, Investigation, Writing—Review & Editing, Funding acquisition. Nie Manhua: Investigation, Supervision. Xiao Yanchao: Investigation, Supervision.

Funding

This work was supported by the (Natural Science Foundation of Hunan Province) under Grant [number 2025JJ81165 and 2024JJ9201]; this research also was funded through the 2025 ‘Young Talent’ Initiative, awarded by the Hunan Nursing Association.

Data availability

The data of this study is provided within the manuscript and available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

We received ethical approval (NO: LYEC2025-0062) from the Ethics Committee of the Second Xiangya Hospital of Central South University. Before completing the questionnaires, all participants were informed of the purpose and method of the study in detail, and their consent was obtained and informed consent signed. To ensure participant anonymity and data security, the online questionnaire was administered via Questionnaire Star with data encryption enabled. No personally identifiable information (e.g., names, or student ID numbers) was collected. Raw data were accessible only to the research team and were stored on a password-protected server. The research procedures adhered strictly to the ethical principles outlined in the Helsinki Declaration of 1964 and all subsequent amendments.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.