Artificial intelligence self-efficacy and attitudes among nursing students: a multicenter network analysis of educational stratification

Abstract

Background

The integration of artificial intelligence (AI) in nursing education is crucial for enhancing technological adaptability. However, the relationship between nursing students’ AI self-efficacy, attitudes, and educational level (associate vs. bachelor’s degrees) remains underexplored. Previous studies predominantly rely on linear models, which inadequately capture the complex dynamics of these factors.

Methods

This multicenter, cross-sectional study was conducted from February to March 2025 across 93 medical institutions in 13 provinces in China. A total of 1,113 nursing students (748 with associate degrees and 365 with bachelor’s degrees) participated. Data were collected using the Artificial Intelligence Self-Efficacy Scale (AISES) and the General Attitudes toward AI Scale (GAAIS). A network model was constructed through LASSO-regularized partial correlation analysis to assess node strength, bridge strength, and predictability. Differences between the two educational groups were analyzed using the Network Comparison Test.

Results

The 24-node network revealed 156 valid connections (56.5%). AI_4 (AI tone matches humans) emerged as the central hub (strength = 2.145), while AI_1 (AI interaction vivid) demonstrated the strongest bridging (bridge strength = 3.255). The Network Comparison Test indicated no significant difference in global network structure (M = 0.255, P = 0.476) or global strength (S = 0.407, P = 0.287) between educational groups. However, significant local differences were found: the bachelor’s degree group showed stronger connections related to technical transparency (E = 0.255), suggesting a more integrated understanding of AI systems. Conversely, the associate degree group exhibited a significant negative association between programming confidence and risk anxiety (E = 0.087), indicating a potential competence-anxiety tension that warrants targeted support.

Conclusions

The AI self-efficacy and attitudes of nursing students form a “cognitive-affective-skill” network, with anthropomorphic interactions serving as a key factor. While the cross-sectional design precludes causal inference, these findings suggest the need for stratified educational strategies: enhancing safety simulations for associate degree students and integrating technical principles for bachelor’s degree students. Such targeted interventions could foster technological empowerment and promote educational equity.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12909-026-08685-y.

Introduction

In recent years, the rapid development of artificial intelligence (AI) technology has profoundly impacted various industries worldwide, particularly in healthcare and education. AI technologies, such as deep learning and natural language processing, have not only revolutionized medical diagnosis and treatment but are also being gradually introduced into nursing education to enhance teaching effectiveness, improve clinical skills, and enrich learning experiences [1, 2]. As AI continues to gain momentum globally, educational institutions are increasingly focusing on how to cultivate AI literacy among nursing students. AI literacy is defined as a set of competencies that enables individuals to critically evaluate AI technologies, communicate and collaborate effectively with AI, and use AI as a tool online, at home, and in the workplace [3]. Research underscores the importance of developing “AI-ready nursing talents”, where self-efficacy and positive attitudes lay the foundation for building AI competencies in nursing students [4, 5]. The urgency of this educational goal is highlighted by recent findings from the clinical workforce. Studies indicate that practicing nurses often face significant challenges, including gaps in digital competencies [6], moderate readiness coupled with operational uncertainty [7], and ethical concerns regarding professional displacement [8]. Thus, integrating AI literacy into nursing education—encompassing basic principles, applications, and ethical considerations—is critical to proactively address these workforce anxieties and equip future healthcare professionals for informed practice.

According to Bandura’s [9] social cognitive theory, self-efficacy is defined as an individual’s belief in their ability to perform specific tasks. In the context of technology adoption, this belief directly influences willingness to adopt AI and actual usage behaviors. This theoretical assertion is supported by recent evidence from the practicing workforce. For instance, Zeng et al. [10] elucidated that AI literacy mediates the pathway from self-efficacy to attitude, ultimately shaping nurses’ AI use intention. Furthermore, in specialized fields such as maternal and child healthcare, network analysis has revealed that self-efficacy interacts closely with usage demands to determine technology acceptance [11]. In nursing students, attitudes toward AI reflect their acceptance and emotional identification with this emerging technology. Studies have shown that nursing students’ AI self-efficacy significantly impacts their willingness to adopt and apply technology, and the quality of their clinical decision-making. Attitudinal dimensions, such as trust in AI and risk perception, are closely linked to their practical application skills in future careers [12–14].

Despite growing research on the impact of AI self-efficacy and attitudes on nursing students’ learning outcomes and technology acceptance, there are several limitations in current studies [4, 15, 16]. Predominantly, existing studies focus on the independent effects of single dimensions using traditional linear models, which fail to capture the complex, synergistic interactions between cognitive, affective, and skill-based factors [17–19]. While this constitutes a critical methodological gap, a more profound conceptual gap remains. Specifically, there is a scarcity of research that moves beyond asking if differences exist between educational levels, to investigate the specific mechanisms that underlie these differences. It is unclear how the internal psychological network structures connecting AI self-efficacy and attitudes-such as the unique patterns of node centrality, bridge influence, and community structure-differ between associate and bachelor’s degree students. Elucidating these education-level-specific network mechanisms is the core of our research question, as it is essential for developing precisely stratified and targeted educational interventions.

To address these gaps, this study is grounded in an integrative theoretical framework that synergizes three pivotal theories to comprehensively explain the network dynamics of AI self-efficacy and attitudes (Figure S1). First, Bandura’s Social Cognitive Theory [9] provides the foundational construct of self-efficacy, positing that individuals’ confidence in their capabilities influences their motivation and behavioral performance. Second, Borsboom’s Network Theory [20] offers the methodological rationale, conceptualizing psychological attributes as causal systems of interacting variables, thus justifying the use of network analysis to move beyond linear models and map complex interactions. Third, Rogers’ Diffusion of Innovations Theory [21] helps explain the attitudinal components and how perceptions of AI influence its adoption within the social system of nursing education. These theories are not employed in isolation; they form a cohesive structure where Bandura’s theory defines the core constructs (“what”), Borsboom’s theory provides the analytical lens (“how”), and Rogers’ theory contextualizes the findings within the process of technological innovation (“why”). This integrated framework ensures a theoretically robust approach to investigating the complex interplay between AI self-efficacy and attitudes.

This study aims to address the identified research gaps by employing psychological network analysis to explore the interactive relationships between AI self-efficacy and attitudes among nursing students at different educational levels. A multicenter cross-sectional survey was conducted across 93 medical schools in China. Data were collected using the Artificial Intelligence Self-Efficacy Scale (AISES) and the General Attitudes Toward Artificial Intelligence Scale (GAAIS). Through LASSO-regularized partial correlation and bridge centrality analysis, this study constructs a network model to identify core factors, bridge factors, and their interaction mechanisms, thereby providing insights for developing stratified educational strategies.

Methods

Aims

To provide new insights for nursing educators into the relationship between AI self-efficacy and attitudes among nursing students. Through network analysis, the study will identify the core and bridge factors influencing these relationships and propose targeted intervention strategies to enhance students’ acceptance and application of AI technology. Additionally, by examining the differences in AI self-efficacy and attitudes across different educational levels, the study will offer a theoretical foundation for developing tailored teaching approaches, thereby optimizing the integration of AI technology in nursing education.

Research context

This cross-sectional study recruited nursing students from medical colleges across mainland China via convenience sampling between February 1 and March 31, 2025. Eligible participants were full-time nursing students (associate or bachelor’s programs) with prior exposure to medical AI through coursework, internships, or self-study, who provided informed consent. Exclusion criteria encompassed: diagnosis of anxiety/depression within the past six months, prolonged medical leave (> 6 months) in the preceding year, dual enrollment in computer science or biomedical engineering programs, or prior participation in analogous AI-related surveys. Questionnaires completed in < 3 min or > 30 min, or those with logical inconsistencies, were excluded to ensure data integrity. Ethical approval was granted by the Medical Research Ethics Committee of West China Second University Hospital, Sichuan University (Ethics No: 2025-050).

Research tools

The study employed three instruments: a general information survey, the AISES, and the GAAIS. The general information survey, designed by the researchers, captured demographic variables such as gender, age, ethnicity, family type, religious beliefs, household registration, educational level, grade, academic stage, and prior AI-related learning experience.

The AISES, developed by Wang et al. [22] who explicitly state that the scale can be ‘adopted or adapted’ for research purposes, assessed nursing students’ AI self-efficacy through 22 items rated on a 7-point Likert scale (1 = “strongly disagree” to 7 = “strongly agree”), with higher scores indicating greater self-efficacy. The scale included four dimensions: assistance (e.g., AI makes learning easier), anthropomorphic interaction (e.g., AI interaction vivid), comfort with AI (e.g., AI interaction calm), and technological skills (e.g., AI use risk-free). Our Chinese version retains all original items with linguistic modifications only, following the World Health Organization (WHO) guidelines for cross-cultural adaptation, the scale was translated into Chinese and validated by six multidisciplinary experts, achieving excellent content validity (I-CVI: 0.83-1.00; S-CVI: 0.992). Reliability was confirmed in both pilot (Cronbach’s α: 0.969 total, 0.888–0.953 subscales) and main studies (Cronbach’s α: 0.982 total, 0.944–0.977 subscales).

The GAAIS, developed by Schepman et al. [23], was used under academic fair use principles to measure nursing students’ attitudes toward AI. The original 20-item structure and all scale items were strictly preserved during the linguistic adaptation process, with items rated on a 5-point Likert scale (1 = “strongly disagree” to 5 = “strongly agree”). Items 1–12 represented positive attitudes while items 13–20 represented negative attitudes (reverse-scored), with higher scores indicating more positive attitudes. Following WHO guidelines for cross-cultural adaptation, the Chinese version was validated by six senior experts, demonstrating excellent content validity (I-CVI: 0.83-1.00; S-CVI: 0.986). Reliability was confirmed in both pilot (Cronbach’s α: 0.931 positive, 0.945 negative) and main studies (Cronbach’s α: 0.957 positive, 0.958 negative).

Survey methods and quality control

This study applied a multi-step process for data collection and quality control. Instead of relying solely on traditional power analysis based on simple proportions, the sample size was determined to meet the rigorous stability requirements for psychological network analysis. According to Epskamp et al. [24], a sample size exceeding 1,000 is generally recommended to accurately estimate network parameters and ensure replicability. Our final valid sample of 1,113 participants satisfied this criterion, yielding a Correlation Stability (CS) coefficient of 0.75 in the subsequent analysis, which far exceeds the recommended threshold of 0.50 for stable networks.

To ensure the applicability of the AISES and GAAIS among Chinese nursing students, a systematic cross-cultural validation process was conducted. First, expert evaluation ensured the content validity of the translated scales. A pilot test with 80 nursing students from four universities in Chengdu assessed the clarity, accuracy, and completion time of the questionnaire, followed by localization optimization. The formal survey was conducted using the Wenjuanxing platform (https://www.wjx.cn/), employing international-standard data encryption, intelligent logic jumps, and real-time quality control. A multi-stage sampling strategy was implemented: uniquely identified electronic questionnaires were distributed through official university channels and social media (WeChat), forwarded by class teachers to eligible students.

A comprehensive quality control system was established for all stages of data collection and processing. Ethical compliance was ensured through electronic informed consent and dual confirmation. Technical measures such as IP restrictions, mandatory fields, and page design (3–5 questions per page) were used to control data quality. The completion process was monitored in real-time, and during data processing, two researchers performed double-blind verification of raw data to ensure accuracy. A multi-stage stratified sampling strategy across 93 universities in 13 provinces was employed to ensure representativeness and data reliability.

Statistical analysis

Data analysis was performed using R software (version 4.4.3) with packages such as qgraph, Network Comparison Test (NCT), and Bootnet. Descriptive statistics were calculated for all variables, with continuous variables expressed as mean ± standard deviation and categorical variables as frequencies and percentages. To evaluate the discriminative power of the AISES and GAAIS items, standard deviations were calculated as measures of variability, and inter-item correlations were tested to ensure item independence.

Network structure was estimated using the qgraph package, which visualized partial correlation matrices and employed LASSO-regularized analysis to construct sparse AI self-efficacy-attitude networks for nursing students [25]. Regularization helped eliminate noise and emphasize core item associations. The extended Bayesian information criterion (EBIC) was applied to optimize the γ parameter (default 0.5), balancing model fit and sparsity [24]. The spring layout algorithm was used to arrange 24 nodes (22 AISES + 2 GAAIS items) with central items placed at network cores and closely related items adjacent. NCT facilitated group comparisons, and edges (green = positive, red = negative) represented partial correlations while controlling for covariates, revealing the multidimensional structures of AI competency mechanisms.

Network characteristics were characterized by three key indicators: node strength, bridge strength, and node predictability. Node strength is the sum of the absolute weights of all edges connected to a node, reflecting its overall influence in the network. Bridge strength represents the sum of the edge weights connecting a node to other dimensional nodes, highlighting interactions between different dimensions [26]. Node predictability quantifies how well a node can be predicted by other nodes in the network [27]. These indicators together helped reveal the complex interactions between AI self-efficacy and attitudes among nursing students.

To ensure the reliability and robustness of the network analysis results, multiple validation methods were used [24]. Edge accuracy was analyzed by calculating 95% non-parametric confidence intervals (CI) for edge weights using 1,000 bootstrap samples, where narrower intervals indicated higher precision. Centrality stability was evaluated via CS coefficients, with values > 0.25 considered acceptable and > 0.5 as ideal stability. Finally, bootstrap difference tests compared edge/node strength differences, with 95% CIs excluding zero (P < 0.05) confirming significance. This multimodal validation framework guaranteed methodologically robust network outcomes.

Results

Sample characteristics and measurement tool analysis

A total of 1,113 valid questionnaires (88.83% validity) were collected from 93 medical colleges across 13 Chinese provinces. The sample was predominantly female (87.0%), aged 20–22 (41.0%), of Han ethnicity (84.9%), and from intact two-parent families (83.3%). Educationally, 67.2% were associate degree students, with 50.7% in their third year. Key findings include 65.8% from rural areas, 47.9% clinical interns, and polarized AI exposure: 38.5% had no experience, 18.6% received systematic training, and 61.5% had varying degrees of exposure (Table S1).

Regarding the measurement tools, item analysis of the AISES (Table S2) and GAAIS (Table S3) scales indicated that the standard deviations of all 24 items were within a reasonable range, with no items showing insufficient information (Data is provided within the supplementary information files). Inter-item correlation coefficients were all below 0.25, confirming good discriminability and independence among the items.

Network characteristics analysis

Network structure Estimation and centrality features

Figure 1A illustrates the AI self-efficacy and attitudes network with 24 items across five dimensions: Assistance (AS_1–7), Anthropomorphic Interaction (AI_1–5), Comfort with AI (CF_1–6), Technological Skills (TS_1–4), and Attitude with AI (AA_1–2). The network includes 156 non-zero edges (56.52% density) with an average edge weight of 0.040, indicating moderate associations. Strength centrality analysis identified AI_4 (AI tone matches humans) as the most influential node (Strength = 2.145), followed by AS_2 (AI is helpful for learning, Strength = 1.232) and AS_1 (AI makes learning easier, Strength = 0.683) (Fig. 1B and Table S4). Bridge strength analysis highlighted AI_1 (AI interaction vivid, Strength = 3.255) as the key cross-dimensional connector, ahead of AA_1 (Positive attitude, Strength = 1.028) and AI_2 (AI expresses uniquely, Strength = 0.890) (Fig. 1C and Table S4). Predictability analysis showed that AS_2 (AI is helpful for learning) had the highest predictability (R²=92.3%), followed by AS_3 (AI aids learning effectively, R²=90.2%) and AI_4 (AI tone matches humans, R²=89.3%) (Fig. 2A-C and Table S4).

Fig. 1.

AI self-efficacy and attitudes network diagram (N = 1113). Note: A The network structure comprising 24 nodes, constructed using the qgraph package based on a partial correlation matrix. B Strength centrality for the 24 nodes in the network.C Bridge strength centrality for the 24 nodes in the network

Fig. 2.

AI self-efficacy and attitudes network structure with node predictability and influence estimates (N = 1113). Note: A Network structure with the 24 nodes interactions and attributes. B Expected Influence of each node.C Bridge Expected Influence (1-step)

Network feature differences across educational levels

The comparison between associate and bachelor’s degree groups revealed a pattern of “global structural similarity with specific local variations”. The Independent Groups Gaussian Network Comparison Test (IG-NCT) revealed no significant structural differences in the global network organization (M = 0.255, P = 0.476, Fig. 3A) or global strength (associate degree group = 11.754, bachelor’s degree group = 11.347, S = 0.407, P = 0.287, Fig. 3B) between the two groups. This suggests that the overall cognitive-affective architecture of AI literacy is consistent across educational levels.

Fig. 3.

Comparison of network structure and influence by education level. Note: (A)Network for associate degree or below (N = 748), (B) Network for bachelor’s degree or higher (N = 365), (C) Strength and Expected Influence by education level

However, significant local differences were identified in specific edge connections. Edge invariance analysis identified 11 significant differences (P < 0.05) between groups (Fig. 3 and Table S5). The bachelor’s degree group had stronger connections in several areas, including the association between TS_3 (AI usage no mysteries) and TS_4 (AI jargon clear) (E = 0.255, P = 0.004), and between AS_5 (AI programming confidence) and AI_2 (AI expresses uniquely) (E = 0.199, P = 0.015). According to Epskamp et al. [24], these edge differences represent moderate effects in the context of regularized partial correlation networks. Additionally, the bachelor’s degree group showed stronger connections in CF_3 (AI interaction comfortable) and CF_6 (AI smooth happy interaction) (E = 0.178, P = 0.005), and AS_4 (AI increases learning interest) and AS_6 (AI saves significant time) (E = 0.163, P = 0.026).

Conversely, the associate degree group had stronger connections in CF_6 and TS_4 (E = 0.235, P = 0.003), and AI_1 (AI interaction vivid) and AI_2 (AI expresses uniquely) (E = 0.209, P = 0.019). A particularly notable finding was the weak negative correlation observed in the associate degree group between AS_5 (AI programming confidence) and TS_1 (AI use risk-free) (E = 0.087, P = 0.013). While the magnitude of this edge is small (E < 0.1), its directionality is theoretically significant, suggesting a unique “competence-risk tension” specific to this group.

The centrality invariance test results revealed significant centrality differences between the two groups (Fig. 3C). The associate degree group exhibited significantly higher centrality for AI_1 (“AI interaction vivid”, strength: 1.021 vs. 0.860, P = 0.037, C = 0.221), while the bachelor’s degree group showed stronger hub effects for CF_4 (AI interaction peaceful, strength: 0.887 vs. 1.222, P = 0.019, C=-0.262).

Bridge strength analysis indicated no significant overall differences in AI self-efficacy and attitudes between the two groups (t=-1.081, df = 23, P = 0.291; MD = 0.026, 95% CI: -0.024-0.075; Cohen’s d = 0.158). Node centrality difference analysis revealed that the associate degree group had significantly higher centrality for CF_6 (AI smooth happy interaction, 0.493 vs. 0.210, Δ = 0.284), TS_4 (AI jargon clear, 0.411 vs. 0.152, Δ = 0.258), TS_1 (AI use risk free, 0.326 vs. 0.152, Δ = 0.174), and AS_5 (AI programming confidence, 0.601 vs. 0.387, Δ = 0.214), while the bachelor’s degree group had higher centrality for CF_2 (AI interaction easy, 0.326 vs. 0.161, Δ=-0.165) (Fig. 4).

Fig. 4.

Comparison of bridge strength and expected influence by education level. Note: (A)Comparison of bridge strength by education level, (B)Comparison of bridge expected influence (1-step) by education level

Network accuracy and stability analysis

This study employed multiple validation methods to rigorously evaluate the network analysis results of AI self-efficacy and attitudes. Bootstrap analysis results showed that the average 95% CI for network edge weights was [0.0027, 0.0866], with a median connection strength of 0.0397, indicating moderately weak overall network connectivity (Fig. 5A). Most edge connections were statistically significant (lower bound ≥ 0.0027) with narrow CIs, such as the connection between AI_4 (AI tone matches humans) and AI_5 (AI text mirrors human speech), with a strength of 0.607 (95% CI: [0.523, 0.679]), confirming the high precision and stability of core network parameter estimates. However, edges close to zero, such as some connections in the assistance dimension, exhibited wider CIs, suggesting some uncertainty in the estimation of these minor connections. This differentiated CI pattern validates the stability of the core network structure while providing necessary statistical references for interpreting weak connections.

Fig. 5.

Accuracy and stability analysis of the AI self-efficacy and attitudes network (N = 1113). Note: A Accuracy analysis of the edge weights. B Stability analysis of the centrality indices

Centrality stability analysis revealed that the CS coefficients for strength centrality and bridge strength were both 0.750, significantly higher than the recommended threshold of 0.5, confirming that the network centrality structure remains highly stable even under sample fluctuations (Fig. 5B). This result further validates the core status of key nodes (e.g., AI_4 “AI tone matches humans” and AS_2 “AI is helpful for learning”) in the network.

Network feature difference testing

This study analyzed network features using bootstrap tests, revealing significant structural differences in the AI self-efficacy and attitudes network. Edge weight analysis showed that the strongest connection was between AI_4 (AI tone matches humans) and AI_5 (AI text mirrors human speech) in the anthropomorphic interaction dimension (edge weight = 0.607, 95% CI: [0.523, 0.679]), indicating high consistency in nursing students’ perception of AI’s anthropomorphic traits. Connections in the technical skills dimension, such as TS_1 (AI use risk-free)-TS_2 (AI use no damage fear, 0.501, 95% CI: [0.386, 0.595]) and TS_3 (AI usage no mysteries)-TS_4 (AI jargon clear, 0.458, 95% CI: [0.363, 0.535]), showed moderate-strength associations, reflecting the internal consistency of students’ evaluation of AI’s technological aspects. Several connections in the assistance and comfort dimensions also reached statistical significance (e.g., AS_2-AS_3, 0.353; CF_5-CF_6, 0.342), highlighting key areas for targeted interventions (Figure S2).

Node strength centrality testing further confirmed that AI_4 in the Anthropomorphic Interaction dimension (AI tone matches humans, strength = 2.145) and AS_2 in the Assistance dimension (AI is helpful for learning, strength = 1.232) exhibited significant differences from most nodes (P < 0.01), indicating their central roles in the network (Figure S3).

Bridge strength analysis revealed that AI_1 in the anthropomorphic interaction dimension (AI interaction vivid, bridge strength = 3.255) and AA_1 in the attitude with AI dimension (Positive attitude, bridge strength = 1.028) exhibited significant cross-dimensional connectivity (P < 0.05), further validating the influential bridge strength of AA_1 (Figure S4).

Discussion

Core features of AI self-efficacy and attitudes network

This study is the first to apply psychological network analysis to map the complex structure of AI self-efficacy and attitudes among nursing students. The network model revealed moderate associations between dimensions (56.52% non-zero edges, average weight = 0.040), aligning with the interaction patterns identified by Gunawan et al. [16]. This structure suggests that AI competence is not an isolated skill but an integrated system formed by the interaction of cognitive, affective, and technical components, supporting Borsboom’s [20] network theory.

Strength centrality analysis indicated that the “Anthropomorphic Interaction” dimension (referring to how much AI resembles human tone) had the highest centrality (Strength = 2.145). This echoes findings by Balakrishnan et al. [28], highlighting that human-like qualities are crucial for technology adoption. AS_2 (AI is helpful for learning) and AS_1 (AI makes learning easier) ranked next in importance, reinforcing OR’s [29] theory on the alignment between technology and tasks. This suggests that AI design in nursing education should prioritize optimizing naturalistic interactions and practical learning support.

Bridge strength analysis identified AI_1 (AI interaction vivid) as the strongest connector between different psychological domains (Strength = 3.255). Extending Jones et al.‘s [30] theory to educational technology, this finding - coupled with its link to positive attitudes (AA_1) - implies that making AI interactions more vivid can significantly improve students’ overall attitudes. This supports a network-oriented approach [31] network-oriented methodology, confirming the value of targeted interventions that optimize key nodes.

Predictability analysis showed that AS_2 (AI is helpful for learning) explained the highest variance (R² = 0.923), acting as a stabilizing anchor in the network. Consistent with Bandura’s [9] social cognitive theory, this indicates that when students perceive AI as functionally useful, their self-efficacy improves. Enhancing students’ cognitive evaluations of AI’s utility can trigger confidence in their technical abilities, creating a positive cycle of motivation and performance. This high predictability provides empirical support for prioritizing “efficacy” in nursing education technology [19, 32].

Furthermore, the central role of the positive attitude node (AA_1) identified in our network resonates with the practical considerations of AI among practicing nurses. For instance, a recent study of Neonatal Intensive Care Unit (NICU) nurses identified ‘perceived worries’ about AI as a significant barrier to its adoption [33]. This connection suggests that the positive attitudes towards AI cultivated during nursing education could serve as a crucial psychological resource for students, potentially buffering against similar worries and facilitating smoother AI integration when they enter the clinical workforce. Therefore, fostering positive attitudes in education is not only about enhancing current learning efficacy but also about preparing students for future professional challenges.

Mechanisms of educational level differences

The core pattern uncovered in this study—“global similarity in network architecture coupled with local specificity in connections” among nursing students at different educational levels-provides the foundational rationale for developing stratified educational strategies. The NCT results indicated that the network structure invariance test (M = 0.255, P = 0.476) and global strength invariance test (S = 0.407, P = 0.287) were not statistically significant, suggesting that the overall organization and connectivity strength of AI self-efficacy and attitudes are comparable between associate degree and bachelor’s degree groups. This consistency aligns with Kwak et al. [4] and Lin et al. [34], suggesting that formal education contributes to a stable cognitive framework regarding technology, likely due to standardized curricula on basic AI principles [35, 36].

However, specific differences (11 edges, P < 0.05) revealed distinct strengths. The bachelor’s group demonstrated systematic advantages in technological skills, assistance, and comfort with AI. For instance, the stronger TS_3–TS_4 connection (E = 0.255) reflects their greater ability to integrate technical principles with terminology, consistent with higher education’s emphasis on systematic knowledge construction [37]. Structured training helps reduce the “black box” effect of AI, thereby increasing their sense of control [38]. Furthermore, the link between programming confidence and the creative use of AI (AS_5–AI_2, E = 0.199) suggests that stronger technical skills facilitate innovation, resonating with Rogers’ theory [21] and the mechanism whereby higher-order skills drive technological re-creation [39].

In contrast, the associate degree group exhibited relative strengths in operational fluency (CF_6–TS_4, E = 0.235) and vivid anthropomorphic interaction (AI_1–AI_2, E = 0.209). These may reflect associate curricula’s focus on practice-based learning. Situated learning theory posits that embedding technical instruction within real-world tasks enables direct mapping from semantic knowledge to operational behaviors [40]. For example, frequent practice with AI-enabled nursing simulations (e.g., voice assistant scenarios) may facilitate fluency and enhance multidimensional interaction skills [41]. Cognitive load theory further suggests that such training reduces context-switching costs, improving efficiency in interaction resource allocation [42]. However, the weak negative edge between programming confidence (AS_5) and risk-free use confidence (TS_1) in the associate group (E = 0.087) highlights a competence-risk tension. This highlights a tension between competence and risk: when technical skills do not match the perceived complexity of a task, students may feel anxious despite having high self-efficacy [43]. This suggests that associate-level curricula need to bridge this gap with safety-oriented scenarios, whereas bachelor curricula can focus on principle-based understanding to buffer against anxiety.

Centrality differences further revealed distinct network “hubs”. In the associate group, AI_1 (AI interaction vivid) exhibited significantly higher centrality (ΔStrength = 0.161, P = 0.037), suggesting a perceptual sensitivity fostered by situational training. Repeated practice with AI-based communication tools may elevate students’ expectations for naturalistic interaction, thereby increasing AI_1’s influence as a network hub [44, 45]. Conversely, bachelor students showed higher centrality for “AI interaction peaceful” (CF_4, ΔStrength = 0.335, P = 0.019), reflecting a sense of emotional security derived from their theoretical understanding [46, 47].

Although no overall bridge strength differences were found (P = 0.29 > 0.05), node-specific differences aligned with educational emphases: associate students showed higher bridge centrality in operational nodes (e.g., TS_4, AS_5), while bachelor students showed higher centrality in affective nodes (e.g., CF_2). This differentiation aligns with the principle that nursing technology education should adapt to learner characteristics [48, 49]. As Aliaga et al.‘s [50] adaptive education framework suggests, stratified interventions are warranted: associate programs should reinforce operational competence through positive feedback, whereas bachelor programs should emphasize emotional support to foster holistic technology integration.

Network structure validation

This study presents the first quantifiable psychological network model of AI self-efficacy and attitudes in nursing education, constructed through systematic network analysis. Bootstrap analysis yielded a 95% CI for core edge weights of [0.0027, 0.0866], with a significant connection (0.607, 95% CI: [0.523, 0.679]) between AI_4 and AI_5 in the anthropomorphic interaction dimension, aligning with Epskamp et al.‘s [24] core structure principles. Centrality stability analysis confirmed the stability of key nodes, such as AI_4 and AS_2, against sampling fluctuations (CS = 0.75), supporting Robinaugh et al.‘s [51] theory on stable intervention targets. Caution is advised in interpreting weak associations due to the wide confidence intervals of minor connections, highlighting the precision stratification of core-periphery structures and its alignment with DE et al.‘s [52] hierarchical framework.

Network analysis revealed a hierarchical connection pattern: the strong AI_4-AI_5 link (0.607) confirms that students perceive human-like traits in AI with high consistency [53], while moderate connections between TS_1-TS_2 (0.501) and TS_3-TS_4 (0.458) reflect the interplay between technical confidence and operational anxiety [54, 55]. Node strength centrality difference tests reaffirmed AI_4 (strength = 2.145) and AS_2 (strength = 1.232) as core hubs, resonating with Bandura’s [9] self-efficacy theory. Bridge strength analysis identified AI_1 (AI interaction vivid, bridge strength = 3.255) as the strongest cross-dimensional connector, supporting Powers et al.‘s [56] theory regarding how vivid technical features strongly trigger affective responses.

Research value and implications for stratified education

The theoretical contribution of this study lies in its pioneering use of psychological network analysis to reconstruct the psychological mechanisms of technology acceptance in nursing education. The study establishes a dynamic interaction model of AI self-efficacy and attitudes, overcoming the limitations of traditional linear causal models. It validates the applicability of Borsboom’s [20] psychological network theory in the field of educational technology and reveals the networked synergistic effect of cognitive, emotional, and skill-based factors. The identification of key hub nodes, such as AI_4 (AI humanized tone matching) and AS_2 (AI assists learning), suggests that the generation of self-efficacy is not driven by a single factor but depends on the nonlinear connections between multiple factors. The cross-dimensional role of bridge nodes, like AI_1 (AI interactive liveliness), extends the boundaries of bridge node theory and offers a new paradigm for the cognitive-emotional integration in educational technology. Furthermore, the high predictive power of AS_2 (R²=0.923) supports Bandura’s [9] social cognitive theory, confirming that perceived utility is critical for building self-efficacy. Critically, our findings translate directly into design strategies: the identified hub and bridge nodes provide precise targets for the framework shown in Figure S1.

From a practical standpoint, the findings offer an actionable framework for the digital transformation of nursing education. First, leveraging the core role of AI_4, we recommend embedding emotion-sensitive interactions (e.g., virtual assistants that recognize emotions and offer empathetic feedback). This reduces anxiety and enhances engagement [57], applying Bandura’s affective principles. Operationally, nursing educators could integrate such intelligent dialogue modules into existing high-fidelity simulators and evaluate the intervention’s impact by tracking changes in students’ anxiety levels (e.g., using the State-Trait Anxiety Inventory) and willingness to interact before and after exposure. Second, we propose linking technology functions with learning scenarios, such as AI simulators that provide real-time error detection [58, 59]. This design directly builds “mastery experiences” (Bandura) and makes the technology’s benefits observable (Rogers). These strategies operationalize the most influential nodes to drive adoption.

Regarding educational stratification, we provide two level‑specific pathways. For associate degree students, we suggest a “safe operation—immediate feedback” pathway. This should include AI-driven diagnosis simulations and automated feedback aligned with International Nursing Association for Clinical Simulation and Learning (INACSL) standards. Research shows that such structured practice improves confidence in foundational procedures [60, 61]. In our model, this approach builds mastery and reduces fear, directly mitigating the negative tension between programming confidence and risk-free operation confidence (E = 0.087).

For bachelor’s degree students, we recommend a visual AI decision‑making dashboard that provides algorithmic transparency (e.g., feature‑importance, case‑based explanations) and supportive prompts to scaffold higher‑order reasoning; research indicates that explainable AI (XAI) enhances understandability and trust in educational contexts, thereby supporting cognitive control during complex decision tasks [62, 63]. Within our framework, this pathway leverages social persuasion and cognitive control (Bandura) and increases perceived relative advantage and observability (Rogers), with AI_1 acting as a bridge between affective and cognitive clusters.

Strategically, this “data-driven precision education path” ensures practical relevance. Associate programs should emphasize high-frequency simulations to help students translate knowledge into action faster, while undergraduate programs should use analytic tools to strengthen emotional security and advanced decision-making. In network terms, targeting hubs (AI_4/AS_2) and bridges (AI_1) will reduce the influence of anxiety-linked nodes and boost efficacy-linked nodes, producing broad improvements consistent with Borsboom’s theory [20]. For evaluation, we recommend tracking measurable outcomes (e.g., Objective Structured Clinical Examination (OSCE) error rates, time-to-decision) alongside network metrics (e.g., global strength) to verify these mechanisms and refine interventions.

Finally, the psychological network structure uncovered in this study underscores that nursing students’ competencies constitute a complex, interconnected system. This perspective aligns with findings from research on other core competencies in nursing education. For example, a recent study demonstrated a significant positive correlation between nursing students’ emotional intelligence and their critical thinking skills [64]. Collectively, this evidence suggests that competencies—be it AI literacy, emotional intelligence, or critical thinking-cannot be developed in isolation; they collectively form a dynamic network underlying professional competence. Therefore, future educational interventions should move beyond traditional siloed skill training and increasingly adopt the network perspective demonstrated in this study. By identifying and precisely strengthening key nodes, educators can more effectively foster the development of the entire integrated competency network.

Limitations and future directions

While this study provides novel insights through network analysis, several methodological limitations warrant consideration. First, the cross-sectional design prevents causal inference regarding how educational experiences shape the network structure over time, particularly regarding the developmental trajectories of key nodes like AI_4 and AS_2.

Second, despite the multicenter design enhancing geographical diversity, the reliance on convenience sampling and recruitment via social media (WeChat) may have introduced selection bias. This approach, while practical for reaching a broad audience, likely attracted participants with a pre-existing interest in or higher familiarity with AI technology. Furthermore, participant contribution was not uniform across the 93 institutions, with a subset of institutions contributing disproportionately to the sample. This limits the generalizability of our findings to the entire population of Chinese nursing students.

Third, the exclusive focus on Chinese nursing students limits cross-cultural generalizability; the central role of “anthropomorphic interaction” observed in this study may be culturally specific and requires validation in other contexts. Finally, this study relies on self-reported perceptions, which may not fully reflect actual competence or behavioral usage of AI tools.

Future research should address these limitations through: [1] longitudinal designs to track network evolution during educational interventions [2], multi-center international studies to establish cultural invariance of the network structure, and [3] integration of objective behavioral data (e.g., actual AI tool usage logs or OSCE performance) to complement self-report measures. Methodologically, developing dynamic network models that incorporate temporal and contextual factors would significantly advance the field.

Conclusion

This study constructs a psychological network model of AI self-efficacy and attitudes in nursing education, providing empirical evidence for competency development strategies. Key findings include: [1] Identification of AI_4 (anthropomorphic interaction) and AS_2 (learning assistance) as central nodes (strength = 2.145 and 1.232, respectively) with AI_1 serving as a cross-dimensional bridge (bridge strength = 3.255) [2], Observed differences between bachelor’s students’ affective integration and associate-degree students’ competence-anxiety tension [3], Quantifiable network parameters suggesting intervention thresholds. While these findings offer preliminary insights for curriculum design, their application should consider the methodological constraints, particularly the need for longitudinal validation. This work lays the groundwork for future research on network-based AI education frameworks.

Abbreviations

AI

Artificial intelligence

AISES

Artificial Intelligence Self-Efficacy Scale

GAAIS

General Attitudes toward AI Scale

Authors’ contributions

Qin Zeng: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Software, Writing – original draft.Jun Zhu: Conceptualization, Methodology, Project administration, Supervision, Writing – review & editing.Yuji Wang: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Writing – review & editing.Shaoyu Su: Conceptualization, Methodology, Project administration, Supervision.Yan Huang: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Supervision, Visualization, Writing – review & editing.

Funding

This work was supported by the Higher Education Teaching Reform Project (11th Batch) of Sichuan University [grant number SCU11199]. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Data availability

Data is provided within the supplementary information files.

Declarations

Ethics approval and consent to participate

Ethical approval for this study was obtained from the Ethics Committee of West China Second University Hospital, Sichuan University (Approval No.: Medical Research 2025 Ethics Review [050]). This research was conducted in strict accordance with the principles of the Declaration of Helsinki established by the World Medical Association. All participants provided electronic informed consent after comprehensive disclosure of study procedures prior to commencement. The study rigorously adhered to ethical principles of confidentiality and non-maleficence to safeguard participants’ rights.

Consent for publication

All participants provided written informed consent for publication of the anonymized data collected in this study.

Competing interests

The authors declare no competing interests.