Chinese version of the Nurses’ AI Ethical Awareness Scale: translation, cross-cultural adaptation, and psychometric evaluation among hospital nurses Authors: Mali Zhang, Yanan Wei, Fangfang Feng, He Tian, Yuandi Yang, Li Yan, Yu Cai, Dan Wang, Chenqian Zhu, Cong Yu Published: 2026 ABSTRACT AbstractBackgroundAs artificial intelligence (AI) becomes increasingly integrated into healthcare, nurses need to recognise and respond to its ethical implications. However, validated Chinese-language instruments for assessing nurses’ ethical awareness of AI use remain limited. This study aimed to translate, culturally adapt, and psychometrically evaluate the Chinese version of the Nurses’ AI Ethical Awareness Scale.MethodsA methodological study was conducted among 455 nurses from a tertiary hospital in Shenzhen, China, between December 2025 and January 2026. The original scale was translated and cross-culturally adapted using a Brislin-style procedure. Participants were randomly divided into an exploratory factor analysis (EFA)/model-refinement subsample (n = 227) and a confirmatory factor analysis (CFA) subsample (n = 228). Item analysis, EFA, candidate-version comparison, CFA, reliability testing, and validity assessment were performed.ResultsThe data were suitable for factor analysis (KMO = 0.9509; Bartlett’s test χ² = 4943.71, P < 0.001). The scale was reduced from 21 to 15 items, forming three closely related content domains. In the CFA subsample, the correlated three-factor and second-order models showed better relative fit than the one-factor and alternative two-factor models. The bifactor model showed the most favourable relative fit among the tested models, with favourable incremental fit indices (CFI = 0.9587, TLI = 0.9421), although RMSEA remained elevated (RMSEA = 0.0967). In the correlated three-factor CFA model, standardised factor loadings ranged from 0.7201 to 0.9455. In the full sample, internal consistency was high for the final 15-item scale (Cronbach’s α = 0.9669). Discriminant validity was only partially supported, whereas bifactor indices suggested that the total score captured most of the common variance (ωh = 0.9459; ECV = 0.8596).ConclusionsThe adapted 15-item Chinese version showed high internal consistency and preliminary validity evidence among hospital nurses. Its three closely related content domains were largely explained by a dominant general factor, supporting total-score interpretation as the primary approach and cautious use of subscale scores. Because privacy, confidentiality, and data-governance items were not retained, the scale should be interpreted as assessing practice-facing aspects of AI ethical awareness. Further multicentre validation is needed.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12912-026-05036-6.Keywords: Artificial intelligence, Nursing ethics, Nurses, Cross-cultural adaptation, Psychometrics, Scale validation, China Introduction Artificial intelligence (AI) is increasingly being incorporated into healthcare and is reshaping nursing-related activities such as risk warning, clinical decision support, documentation, workflow management, patient monitoring, and resource allocation [1]. Although these applications may improve efficiency and service delivery, they also raise ethical concerns regarding privacy, transparency, accountability, fairness, patient autonomy, professional responsibility, and the preservation of human-centred care [2–6]. These concerns are particularly important in nursing because nursing practice is relational, patient-centred, and ethically grounded. Nurses need to recognise not only whether AI systems are useful, but also whether their use protects patient dignity, supports informed participation, maintains professional accountability, and remains compatible with human-centred nursing values. This issue is especially relevant in China, where hospital digitalisation and Internet Plus Nursing Services have increased nurses’ exposure to digital platforms, remote care processes, and data-mediated workflows [7]. At the same time, high clinical workload remains a prominent feature of Chinese hospitals; a large-scale observational study reported that nurses in large general hospitals cared for an average of eight patients during the day and 23 patients at night [8]. Recent Chinese evidence also suggests that although healthcare and nursing-related professionals generally hold positive attitudes toward AI in nursing, many have limited knowledge of AI in nursing, and 95.7% considered it necessary to strengthen medical ethics related to AI in nursing [9]. In such rapidly digitalising and high-volume clinical environments, nurses may face ethical tensions when balancing efficiency, patient safety, professional accountability, and human-centred care. Therefore, a context-appropriate measure of nurses’ AI ethical awareness is needed to support ethics education, organisational assessment, and responsible AI implementation in Chinese nursing settings. In the present study, nurses’ AI ethical awareness was conceptualised as nurses’ ability to recognise, reflect on, and respond to ethical issues that may arise when AI is used in nursing and healthcare contexts. Operationally, this construct was assessed as nurses’ self-reported awareness of ethical responsibilities relevant to AI-supported nursing practice, including patient rights and informed decision-making, professional accountability, fairness and bias awareness, safety oversight, responsible implementation, and continuing ethical learning. Privacy and data governance were also regarded as important AI-related ethical concerns during translation and item evaluation; however, their empirical performance and factorial contribution were examined rather than assumed. Existing studies concerning AI in nursing have focused mainly on attitudes toward AI, willingness to use AI systems, technology acceptance, digital literacy, readiness, perceived implementation barriers, or broader ethical and regulatory considerations [10, 11]. These constructs are important but are not equivalent to ethical awareness. A nurse may be willing to use AI or may possess adequate digital competence, yet still lack sufficient awareness of ethical risks such as opaque decision-making, algorithmic bias, over-reliance on AI recommendations, unclear accountability, insufficient disclosure to patients, or weakened patient participation. Without adequate ethical awareness, AI implementation may become efficiency-driven while insufficiently addressing patient protection and professional integrity. Valid and culturally appropriate measurement tools are therefore required to assess nurses’ AI ethical awareness, identify educational needs, evaluate ethics-focused training, and support ethically oriented AI implementation in nursing settings [12, 13]. However, instruments developed in one linguistic, cultural, and healthcare-system context cannot be assumed to function equivalently in another. Differences in language, healthcare organisation, professional norms, ethical expectations, regulatory emphasis, and clinical exposure to AI systems may influence how nurses interpret items related to accountability, transparency, consent, privacy, and safety. Therefore, cross-cultural adaptation should be accompanied by empirical testing of item performance and factorial structure rather than assuming structural equivalence. The Nurses’ AI Ethical Awareness Scale was originally developed and validated by El-Sayed et al. [14] as a profession-specific instrument measuring nurses’ ethical awareness regarding AI use in clinical practice. The original scale retained 21 items across six domains, including data privacy and confidentiality, consent and autonomy, transparency and accountability, bias and equity, safety and professional integrity, and education and sustainability. Compared with instruments measuring AI literacy, technology acceptance, digital readiness, or general attitudes toward AI, this scale is more closely aligned with the present construct because it focuses specifically on ethical awareness in AI-supported nursing practice. Nevertheless, its linguistic equivalence, cultural appropriateness, item performance, and factor structure require re-evaluation before use in Chinese nursing contexts. Therefore, this study aimed to translate and culturally adapt the Nurses’ AI Ethical Awareness Scale into Chinese and to evaluate its psychometric properties among hospital nurses. Specifically, this study examined item performance, internal structure, internal consistency, convergent validity, discriminant validity, and score interpretation. By providing preliminary evidence for a context-appropriate Chinese version, this study may support systematic assessment of nurses’ AI ethical awareness and inform nursing education, research, and organisational efforts toward ethically responsible AI implementation. Methods Study design and participants This methodological study aimed to translate, culturally adapt, and psychometrically evaluate the Chinese version of the Nurses’ AI Ethical Awareness Scale among hospital nurses. A convenience sample of registered nurses was recruited from Shenzhen Second People’s Hospital between December 2025 and January 2026. Eligible participants were nurses who: (1) held a valid nursing license in the People’s Republic of China; (2) had been engaged in clinical nursing, nursing management, or nursing education in the study hospital for at least 1 year; and (3) voluntarily agreed to participate and provided informed consent. Nurses were excluded if they were internship nurses or visiting nurses, or if they had been continuously absent from work for more than 3 months due to sick leave, maternity leave, or other reasons. A total of 455 questionnaires were submitted, and all were retained after data-quality screening. The initial Chinese version contained 21 items; therefore, the final sample size exceeded the commonly recommended minimum of 5–10 participants per item for psychometric evaluation [15]. Instrument translation and cross-cultural adaptation Permission to translate and use the original scale was obtained from the original author. The Chinese version was developed using a Brislin-style translation and cross-cultural adaptation procedure, with reference to established recommendations for cross-cultural scale adaptation and validation [12]. First, two bilingual translators independently translated the original English version into Chinese. One translator was a nursing doctoral-level expert familiar with nursing terminology, and the other was an English language professional. The two forward translations were then compared and synthesised into a preliminary Chinese version after discussion within the research team. Second, two additional bilingual translators who had not been exposed to the original scale independently back-translated the preliminary Chinese version into English. The back-translated versions were compared with the original scale to identify discrepancies in meaning, wording, and conceptual equivalence. Third, an expert panel was convened to review the translated and back-translated versions. The panel consisted of six experts, including one expert with a doctoral degree in English, one nursing doctoral-level expert, and four senior clinical nurses with senior professional titles. The panel compared the Chinese version with the original instrument and revised the wording to improve semantic equivalence, conceptual clarity, cultural appropriateness, and clinical readability. During this process, the panel paid particular attention to AI-related ethical terms, including accountability, transparency, AI-assisted decision-making, consent, privacy, and data governance, because these concepts may be understood differently across linguistic, professional, and healthcare-system contexts. Several terms required contextual refinement for Chinese clinical nursing practice. For example, wording related to accountability was reviewed to reflect nurses’ professional responsibility when AI-assisted recommendations are used, rather than only legal liability. Wording related to AI-assisted decision-making was refined to avoid implying that AI could replace nurses’ clinical judgement, and privacy- and data-governance-related wording was reviewed to distinguish bedside confidentiality responsibilities from hospital-level information-system governance. These refinements were intended to preserve the conceptual meaning of the original items while improving readability and contextual relevance in routine Chinese hospital nursing settings. Disagreements were resolved through discussion until consensus was reached. A pilot test was then conducted among 40 registered nurses who met the inclusion criteria but were not included in the formal psychometric sample. The pilot test evaluated the comprehensibility, acceptability, clarity, and response burden of the preliminary Chinese version. Based on the pilot feedback, the research team refined the wording, simplified ambiguous expressions, and improved item clarity before formal data collection. No item was deleted at this stage. An initial Chinese version containing 21 items was then produced for subsequent psychometric item screening and structural validation. Measures Sociodemographic and professional characteristics A researcher-designed questionnaire was used to collect participants’ background information, including sex, age group, highest education, professional title, years of nursing practice, department, and previous exposure to AI-related systems, such as intelligent warning systems, clinical decision support systems, and intelligent scheduling systems. Chinese version of the Nurses’ AI Ethical Awareness Scale The original Nurses’ AI Ethical Awareness Scale developed by El-Sayed et al. [14] consisted of 21 items across six domains: data privacy and confidentiality, consent and autonomy, transparency and accountability, bias and equity, safety and professional integrity, and education and sustainability. Each item was rated on a 5-point Likert scale from 1 = strongly disagree to 5 = strongly agree, with higher scores indicating greater ethical awareness regarding AI use. The initial Chinese version retained the 21 translated items before psychometric item screening. Data collection Data were collected using an electronic questionnaire. Eligible nurses were informed of the study purpose, content, and procedures through their head nurses. The anonymity and confidentiality of the survey were emphasised before participation. Nurses who agreed to participate accessed the questionnaire by scanning a Quick Response code and completed it independently. To improve data quality, the electronic questionnaire required complete responses before submission and allowed only one submission per participant. After data collection, responses were screened for eligibility, completeness, duplicate entries, logical inconsistencies, and invariant response patterns across scale items. No submitted questionnaires met the exclusion criteria, and all 455 responses were retained for analysis. Statistical analysis All statistical analyses were conducted using Python 3.9. Descriptive statistics were used to summarise participants’ sociodemographic and professional characteristics, and categorical variables were presented as frequencies and percentages. Exploratory factor analysis, confirmatory factor analysis, reliability testing, validity assessment, and bifactor-index calculation were performed according to the prespecified psychometric evaluation workflow. Confirmatory factor analysis and bifactor models were estimated using semopy (version 2.3.11), an open-source Python package for structural equation modelling [16], to support transparent and reproducible script-based implementation of the prespecified structural models. The interpretation of the results was based on the prespecified models, fit indices, and bifactor indices rather than on software-specific preferences. Psychometric evaluation was conducted in several stages. To strengthen structural validation, the full sample was randomly divided into two approximately equal subsamples using stratified random sampling based on the initial 21-item total score distribution. One subsample was used as the EFA/model-refinement subsample, and the other was used as the independent CFA subsample. In the EFA/model-refinement subsample, item analysis was first performed for the initial 21-item Chinese version using corrected item–total correlations and Cronbach’s α if item deleted. Data suitability for exploratory factor analysis was assessed using the Kaiser–Meyer–Olkin measure and Bartlett’s test of sphericity. Parallel analysis was conducted to examine the likely number of latent factors. Exploratory factor analysis was performed using the minimum residual extraction method with promax rotation because correlations among latent dimensions were expected. Items were flagged for potential removal when they showed weak corrected item–total correlations, low primary factor loadings, problematic cross-loadings, increased Cronbach’s α if deleted, conceptual redundancy, or limited theoretical interpretability. Commonly used psychometric reference values were considered, including corrected item–total correlations below 0.40, primary factor loadings below 0.40, and secondary loadings of 0.30 or above. These criteria were used as decision aids rather than automatic deletion rules, and final item retention was determined by jointly considering statistical performance, content coverage, structural clarity, parsimony, and theoretical interpretability [15]. After preliminary item screening, several candidate item-retention versions were compared in the model-refinement subsample to examine their relative structural performance. This comparison was used to support short-form refinement and item-retention decisions, together with EFA results, item performance, content representativeness, parsimony, and theoretical interpretability. The final retained version was subsequently evaluated in the independent CFA subsample. For the final retained version, confirmatory factor analysis was conducted in the independent CFA subsample to compare a one-factor model, correlated three-factor model, second-order three-factor model, and bifactor model. Supplementary two-factor models formed by merging highly correlated factors were also examined. Confirmatory models were estimated using maximum likelihood estimation. The 5-point Likert items were treated as approximately continuous variables in the main confirmatory analyses because each item had five ordered response categories and this approach is commonly used in psychometric validation studies. Model fit was primarily evaluated using χ²/df, comparative fit index, Tucker–Lewis index, root mean square error of approximation, Akaike information criterion, and Bayesian information criterion. Additional fit indices, including goodness-of-fit index, adjusted goodness-of-fit index, and normed fit index, were also reported for comparability with previous psychometric studies. Comparative fit index and Tucker–Lewis index values ≥ 0.90 were considered acceptable, and values ≥ 0.95 indicated good fit. Root mean square error of approximation values < 0.08 were considered acceptable; elevated root mean square error of approximation values were interpreted cautiously in combination with other fit indices and theoretical interpretability [17]. Internal consistency was assessed using Cronbach’s α. Convergent validity was evaluated using standardised factor loadings, composite reliability, and average variance extracted. Standardised loadings ≥ 0.50, composite reliability ≥ 0.70, and average variance extracted ≥ 0.50 were considered acceptable. Discriminant validity was assessed using inter-factor correlations, the Fornell–Larcker criterion, and the heterotrait–monotrait ratio. Heterotrait–monotrait ratio values < 0.85 were interpreted as meeting the stricter criterion, whereas values < 0.90 were considered acceptable under a more lenient criterion [18]. Because confirmatory factor analysis results suggested closely related dimensions, bifactor indices were further calculated to inform dimensionality and score interpretation, including omega total, omega hierarchical, explained common variance, item explained common variance, percentage of uncontaminated correlations, and omega hierarchical subscale values. These indices were used to evaluate the strength of the general factor and the independent interpretability of specific factors, rather than to claim a definitively optimal structural model [19]. All statistical tests were two-sided where applicable, and a P value < 0.05 was considered statistically significant. Ethical considerations This study was approved by the Medical Ethics Committee of Shenzhen Second People’s Hospital (approval number: 2026-013-01PJ). All procedures involving human participants were performed in accordance with relevant guidelines and regulations, including the Declaration of Helsinki. All participants voluntarily took part after providing informed consent. The survey was anonymous, and all collected data were used only for research purposes. No personally identifiable information was collected. Results Participant characteristics A total of 455 nurses were enrolled. Using stratified random splitting based on the initial 21-item total score distribution, 227 participants were assigned to the EFA subsample and 228 to the CFA subsample. As shown in Table 1, the two subsamples were comparable across sociodemographic and professional variables. In the total sample, most participants were female (95.6%) and held a bachelor’s degree (92.5%). The largest age group was 31–35 years (25.9%), and the most common professional title was nurse-in-charge (52.1%). Regarding AI-related systems, 36.9% had heard of such systems but had never used them, whereas 45.7% reported occasional or frequent use. Table 1Participant characteristics of the total, EFA, and CFA samplesVariableTotal sample (n = 455)EFA sample (n = 227)CFA sample (n = 228) Sex Female435 (95.6%)219 (96.5%)216 (94.7%) Male20 (4.4%)8 (3.5%)12 (5.3%) Age group 18–2557 (12.5%)27 (11.9%)30 (13.2%) 26–3091 (20.0%)50 (22.0%)41 (18.0%) 31–35118 (25.9%)56 (24.7%)62 (27.2%) 36–4069 (15.2%)35 (15.4%)34 (14.9%) 41–4569 (15.2%)29 (12.8%)40 (17.5%) ≥4651 (11.2%)30 (13.2%)21 (9.2%) Highest education Junior college13 (2.9%)7 (3.1%)6 (2.6%) Bachelor’s degree421 (92.5%)206 (90.7%)215 (94.3%) Master’s degree or above21 (4.6%)14 (6.2%)7 (3.1%) Professional title Nurse61 (13.4%)32 (14.1%)29 (12.7%) Nurse practitioner105 (23.1%)55 (24.2%)50 (21.9%) Nurse-in-charge237 (52.1%)110 (48.5%)127 (55.7%) Associate chief nurse or above52 (11.4%)30 (13.2%)22 (9.6%) Years of nursing practice ≤5 years103 (22.6%)50 (22.0%)53 (23.2%) 6–10 years92 (20.2%)49 (21.6%)43 (18.9%) 11–15 years91 (20.0%)46 (20.3%)45 (19.7%) 16–20 years48 (10.5%)22 (9.7%)26 (11.4%) >20 years121 (26.6%)60 (26.4%)61 (26.8%) Department Internal medicine116 (25.5%)57 (25.1%)59 (25.9%) Surgery227 (49.9%)118 (52.0%)109 (47.8%) ICU/Emergency29 (6.4%)11 (4.8%)18 (7.9%) Operating room/Anesthesia43 (9.5%)21 (9.3%)22 (9.6%) Outpatient/Medical technology38 (8.4%)18 (7.9%)20 (8.8%) Other2 (0.4%)2 (0.9%)0 (0.0%) Previous exposure to AI-related systems Never79 (17.4%)42 (18.5%)37 (16.2%) Heard of but never used168 (36.9%)81 (35.7%)87 (38.2%) Occasionally used145 (31.9%)74 (32.6%)71 (31.1%) Frequently used63 (13.8%)30 (13.2%)33 (14.5%)Note: Values are presented as n (%). EFA = exploratory factor analysis; CFA = confirmatory factor analysis Participant characteristics of the total, EFA, and CFA samples Note: Values are presented as n (%). EFA = exploratory factor analysis; CFA = confirmatory factor analysis Data factorability was supported in the EFA subsample, with a KMO value of 0.9509 and a significant Bartlett’s test of sphericity, χ² = 4943.71, P < 0.001. Parallel analysis suggested the presence of a strong general factor underlying the item pool (Table S1). Item screening and reduction The initial 21-item Chinese version demonstrated excellent internal consistency in the EFA/model-refinement subsample (Cronbach’s α = 0.9625). Full item-level diagnostics for all 21 items, including corrected item–total correlations, Cronbach’s α if item deleted, EFA loadings, cross-loading indicators, and final item decisions, are provided in Supplementary Table S2. Key item-level diagnostics and final decisions are summarised in Table 2. Table 2Item-level diagnostics and final decisions for the initial 21-item pool in the EFA subsampleItem code and abbreviated contentCITCMaximum EFA loadingCross-loadingFinal decisionQ1. De-identifying patient identifiers before non-direct AI use0.3880.208NoDeletedQ2. Perceived compliance of AI systems with data-protection requirements0.6180.260NoDeletedQ3. Avoiding AI use that may compromise patient privacy0.6930.687NoDeletedQ4. Awareness of patient-data leakage risks during AI use0.4610.613NoDeletedQ5. Patients’ and families’ right to know AI involvement0.6860.627NoRetainedQ6. Explaining AI use to patients and obtaining consent0.8010.855NoRetainedQ7. Patients’ right to refuse AI-assisted care decisions0.7760.961NoRetainedQ8. Documenting AI-assisted recommendations in medical records0.7290.871NoRetainedQ9. Explainability of AI recommendations0.7970.430YesDeletedQ10. AI as an auxiliary tool; clinicians make final decisions0.8240.613NoRetainedQ11. Nurses’ responsibility for professional decisions with AI0.8380.602NoRetainedQ12. Awareness of bias from incomplete AI training data0.7810.825NoRetainedQ13. Regular checking of AI systems for bias and potential harm0.8820.794NoRetainedQ14. Considering individual patient differences when using AI advice0.8750.665NoRetainedQ15. Checking AI recommendations for errors or unreasonable advice0.8830.703NoRetainedQ16. Prioritising professional judgement when AI advice conflicts0.7340.301NoRetainedQ17. Reporting safety problems or adverse events related to AI0.8530.358YesRetainedQ18. Ensuring AI use complies with nursing workflow and patient-safety rules0.8640.755NoDeletedQ19. Monitoring appropriate AI use and new developments0.8200.829NoRetainedQ20. Supporting reasonable AI use under safety and ethical safeguards0.8120.987NoRetainedQ21. Including AI ethics in nursing training and continuing education0.6780.958NoRetainedNote: CITC = corrected item–total correlation; EFA = exploratory factor analysis. Item content is abbreviated for readability; full English and Chinese wording of the final retained 15 items is provided in Supplementary Table S9. Cross-loading was defined as a second loading ≥0.30. Final item decisions were based on item-level diagnostics, EFA loading patterns, candidate-version model comparison, content representativeness, and theoretical interpretability rather than on a single statistical threshold. Full EFA loadings and additional item-level diagnostics, including second loadings and Cronbach’s α if item deleted, are provided in Supplementary Tables S2A and S2C Item-level diagnostics and final decisions for the initial 21-item pool in the EFA subsample Note: CITC = corrected item–total correlation; EFA = exploratory factor analysis. Item content is abbreviated for readability; full English and Chinese wording of the final retained 15 items is provided in Supplementary Table S9. Cross-loading was defined as a second loading ≥0.30. Final item decisions were based on item-level diagnostics, EFA loading patterns, candidate-version model comparison, content representativeness, and theoretical interpretability rather than on a single statistical threshold. Full EFA loadings and additional item-level diagnostics, including second loadings and Cronbach’s α if item deleted, are provided in Supplementary Tables S2A and S2C Item reduction was conducted by jointly considering statistical performance, content representativeness, candidate-version model comparison, and theoretical interpretability. Statistical thresholds were treated as indicators of potential item problems rather than automatic deletion rules. Q1 was deleted because it showed a low corrected item–total correlation, weak EFA loading, and increased Cronbach’s α if deleted. Q2 was deleted because of weak EFA loading and limited contribution to the emerging factor structure. Q4 was deleted because Cronbach’s α increased after deletion and its item performance was relatively weaker. Q9 was deleted because it showed cross-loading and structural ambiguity. Not all deleted items showed poor standalone item statistics. Q3 showed acceptable corrected item–total correlation and EFA loading, but was removed during candidate-version refinement because the 15-item solution showed better structural clarity, parsimony, and content non-redundancy after its removal. Q18 also showed acceptable standalone item performance, but was removed during candidate-version refinement because of conceptual overlap with retained items related to professional accountability, clinical judgement, workflow appropriateness, and safety oversight. These decisions were therefore based on the overall measurement structure, parsimony, and content non-redundancy rather than on item-level statistics alone. In contrast, Q16 and Q17 showed weaker loading patterns in the initial 21-item EFA, and Q17 also showed cross-loading. However, both items were retained because they represented important content related to professional judgement, safety reporting, and clinical accountability, and both showed satisfactory standardised loadings in the final 15-item CFA model (Q16 = 0.7201; Q17 = 0.9344). Consequently, the final Chinese version retained 15 items: Q5–Q8, Q10–Q17, and Q19–Q21. This retained version was then selected for subsequent structural model comparison and confirmatory validation. Candidate-version comparison and confirmatory structural validation After preliminary item screening, four candidate item-retention versions were compared in the model-refinement subsample to examine their relative structural performance (Table 3A). This comparison was interpreted alongside EFA loading patterns, item diagnostics, content coverage, parsimony, and theoretical interpretability, rather than as a replacement for conventional EFA interpretation. Compared with the 16-, 17-, and 18-item alternatives, the final 15-item version showed the most favourable relative performance, including lower χ²/df, AIC, and BIC values and better CFI, TLI, and RMSEA values. Considering model fit, parsimony, structural clarity, and content coverage, this version was selected for independent confirmatory evaluation in the CFA subsample. Table 3Model fit of candidate versions and competing structural models 3A. CFA-type model-based comparison of candidate retained item versions in the EFA/model-refinement subsample Candidate version Deleted items χ² df χ²/df CFI TLI RMSEA AIC BIC 18-item versionQ1, Q3, Q4591.5461324.4810.9010.8850.12472.788206.36117-item versionQ1, Q2, Q3, Q4544.7351164.6960.9050.8890.12869.201195.92416-item versionQ1, Q2, Q3, Q4, Q9437.5861014.3330.9210.9060.12166.145186.018Final 15-item versionQ1, Q2, Q3, Q4, Q9, Q18317.541873.6500.9400.9280.10863.202176.226 3B. Confirmatory comparison of competing structural models for the final 15-item version in the independent CFA subsample Model χ² df χ²/df CFI TLI RMSEA GFI AGFI NFI AIC BIC One-factor model615.563906.8400.8640.8410.1600.8440.8190.84454.600157.481Correlated three-factor model341.957873.9310.9340.9200.1140.9140.8960.91463.000176.169Second-order three-factor model341.957873.9310.9340.9200.1140.9140.8960.91463.000176.169Bifactor model234.221753.1230.9590.9420.0970.9410.9170.94187.945242.266Note: CFI = comparative fit index; TLI = Tucker–Lewis index; RMSEA = root mean square error of approximation; GFI = goodness-of-fit index; AGFI = adjusted goodness-of-fit index; NFI = normed fit index; AIC = Akaike information criterion; BIC = Bayesian information criterion. Table 3A presents CFA-type model-based comparisons in the EFA/model-refinement subsample; full EFA loadings are reported in Supplementary Table S2A. Table 3B presents confirmatory structural models for the final 15-item version in the independent CFA subsample. The second-order three-factor model produced identical fit indices to the correlated three-factor model in this specification. Because the bifactor model showed better incremental and residual fit indices but higher AIC/BIC, it was used to inform dimensionality and score interpretation rather than to establish a definitively optimal model Model fit of candidate versions and competing structural models Note: CFI = comparative fit index; TLI = Tucker–Lewis index; RMSEA = root mean square error of approximation; GFI = goodness-of-fit index; AGFI = adjusted goodness-of-fit index; NFI = normed fit index; AIC = Akaike information criterion; BIC = Bayesian information criterion. Table 3A presents CFA-type model-based comparisons in the EFA/model-refinement subsample; full EFA loadings are reported in Supplementary Table S2A. Table 3B presents confirmatory structural models for the final 15-item version in the independent CFA subsample. The second-order three-factor model produced identical fit indices to the correlated three-factor model in this specification. Because the bifactor model showed better incremental and residual fit indices but higher AIC/BIC, it was used to inform dimensionality and score interpretation rather than to establish a definitively optimal model In the CFA subsample, competing structural models were evaluated for the final 15-item version (Table 3B). The one-factor model showed poorer fit than the multidimensional models. Supplementary two-factor models formed by merging highly correlated factors showed poorer incremental and residual fit than the correlated three-factor model (Table S3), suggesting that simple factor merging did not adequately represent the data structure. The correlated three-factor model showed acceptable incremental fit indices but an elevated RMSEA. The second-order three-factor model yielded the same fit indices as the correlated three-factor model, which is expected under this model specification and was therefore not interpreted as independent evidence of better fit. The bifactor model showed the strongest overall performance among the tested models, with CFI = 0.9587 and TLI = 0.9421. These values indicated favourable incremental fit, with the CFI reaching the conventional good-fit threshold and the TLI exceeding the acceptable-fit threshold. Although the RMSEA remained above the conventional threshold, it was lower than that of the one-factor and correlated three-factor models (RMSEA = 0.0967). Given the favourable incremental fit indices, theoretical interpretability, and subsequent bifactor indices, the elevated RMSEA was not interpreted in isolation as evidence against the adapted structure. Although the bifactor model showed the most favourable χ²/df, CFI, TLI, RMSEA, GFI, AGFI, and NFI values, it had higher AIC and BIC values, reflecting greater model complexity. Therefore, the bifactor model was used to inform dimensionality and score interpretation rather than to designate a definitively optimal structural model. Bifactor indices further supported a dominant general factor. Omega total was 0.9775, omega hierarchical was 0.9459, and the explained common variance of the general factor was 0.8596. The percentage of uncontaminated correlations was 0.6476. These findings indicated that most reliable common variance was attributable to the general factor. Accordingly, the total score was considered the most defensible primary scoring approach for the final 15-item Chinese version, whereas the three content domains were interpreted as closely related descriptive domains rather than fully independent subscales. Psychometric properties of the final 15-item scale The final 15-item solution comprised three interpretable content factors. Factor 1, Professional accountability and safety oversight (Q10, Q11, Q12, Q13, Q14, Q15, Q16, and Q17), reflected nurses’ awareness of professional responsibility, bias monitoring, safety checking, and clinician oversight in AI-assisted care. Factor 2, Patient rights, disclosure, and informed decision-making (Q5, Q6, Q7, and Q8), captured awareness of patients’ rights to be informed, to consent, to refuse, and to have AI-assisted care appropriately documented. Factor 3, Responsible implementation and continuous ethical learning (Q19, Q20, and Q21), reflected support for ethically governed AI use, continued attention to appropriate application, and incorporation of AI ethics into nursing education and continuing professional development. In the correlated three-factor model in the CFA subsample, standardised factor loadings ranged from 0.7201 to 0.9455 (Table 4 and Table S4). Convergent validity was supported by high composite reliability values (CR = 0.9039–0.9601) and AVE values (0.7024–0.7736). In the full sample, the final 15-item scale demonstrated high internal consistency, with a total Cronbach’s α of 0.9669; Cronbach’s α values for the three factors were also satisfactory (Table S8). Table 4Core psychometric properties of the final 15-item scale in the CFA subsample 4A. Standardized factor loadings Factor Items Standardized loading range F1Q10, Q11, Q12, Q13, Q14, Q15, Q16, Q170.7201–0.9455F2Q5, Q6, Q7, Q80.7642–0.8821F3Q19, Q20, Q210.7657–0.9373 4B. Composite reliability and convergent validity Factor Number of items CR AVE F180.9600.752F240.9040.702F330.9110.774 4C. Inter-factor correlations and HTMT ratios Factor pair Standardized correlation HTMT F1–F20.8470.849F1–F30.9020.873F2–F30.7900.786Note: F1 = Professional accountability and safety oversight; F2 = Patient rights, disclosure, and informed decision-making; F3 = Responsible implementation and continuous ethical learning. CR = composite reliability; AVE = average variance extracted; HTMT = heterotrait–monotrait ratio. Standardised loadings are from the correlated three-factor model in the CFA subsample. Cronbach’s α values calculated in the full sample are reported in Supplementary Table S8 Core psychometric properties of the final 15-item scale in the CFA subsample Note: F1 = Professional accountability and safety oversight; F2 = Patient rights, disclosure, and informed decision-making; F3 = Responsible implementation and continuous ethical learning. CR = composite reliability; AVE = average variance extracted; HTMT = heterotrait–monotrait ratio. Standardised loadings are from the correlated three-factor model in the CFA subsample. Cronbach’s α values calculated in the full sample are reported in Supplementary Table S8 Inter-factor correlations ranged from 0.7895 to 0.9018, indicating substantial overlap among the three content factors. Discriminant validity was therefore only partially supported. Although all HTMT values were below 0.90, the HTMT value for the relationship between Professional accountability and safety oversight and Responsible implementation and continuous ethical learning exceeded the stricter 0.85 criterion. This pattern was consistent with the Fornell–Larcker results, which did not fully support discriminant validity (Table S5). Bifactor indices further supported the dominance of a general factor. Omega hierarchical was 0.9459, and the explained common variance of the general factor was 0.8596 (Table S6). The PUC value was 0.6476, suggesting that some residual multidimensionality remained. Omega hierarchical subscale values were low, indicating limited independent reliability of the specific factors after controlling for the general factor. Item-level bifactor indices are presented in Table S7. Taken together, these findings suggest that the total score may be more robust and interpretable than the subscale scores, whereas subscale scores should be used cautiously as supplementary descriptive information. Supplementary analyses in the full sample showed a similar overall pattern of model fit and internal consistency (Table S8). Discussion Principal findings This study translated, cross-culturally adapted, and psychometrically evaluated the Nurses’ AI Ethical Awareness Scale among hospital nurses in China. The findings provide preliminary support for a shortened 15-item Chinese version with high internal consistency in the full sample and satisfactory convergent validity in the CFA subsample. However, the structural results also indicated important complexity. Although the retained items could be organised into three interpretable content factors, the high inter-factor correlations, partial support for discriminant validity, and bifactor indices suggested that these factors were closely related and underpinned by a strong general factor. Accordingly, the Chinese version may be more robustly interpreted at the total-score level, whereas subscale scores should be used cautiously as supplementary descriptive information. Factor structure, model fit, and score interpretation The original English version of the Nurses’ AI Ethical Awareness Scale reported a 21-item, six-factor structure [14]. In contrast, the present Chinese adaptation supported a more parsimonious 15-item structure comprising three closely related content domains: professional accountability and safety oversight, patient rights, disclosure, and informed decision-making, and responsible implementation and continuous ethical learning. This difference should not be interpreted simply as a failure to replicate the original structure. Rather, it may reflect the combined influence of translation, cultural adaptation, clinical context, and the way nurses in this sample understood AI-related ethical responsibilities in clinical practice. The model-fit results were interpreted using multiple sources of evidence rather than relying on a single fit index. In the CFA subsample, the bifactor model showed the most favourable overall performance among the tested models, with CFI = 0.9587 and TLI = 0.9421, supporting favourable incremental fit. Although RMSEA remained above the conventional threshold, it was lower in the bifactor model than in the one-factor and correlated three-factor models. Empirical methodological work has cautioned that fixed RMSEA cut-off values should not be treated as rigid universal decision rules, because RMSEA may be influenced by model specification, degrees of freedom, sample size, model complexity, and other modelling conditions [20]. Therefore, the elevated RMSEA in this study was interpreted as indicating residual model misfit and the need for further validation, but not as sufficient grounds to reject the adapted 15-item structure. The bifactor indices further clarified score interpretation. Omega hierarchical was 0.9459, and the explained common variance of the general factor was 0.8596, indicating that most reliable common variance was attributable to a dominant general AI ethical awareness factor. These findings support the use of the total score as the primary and most defensible scoring approach for the Chinese version. However, the bifactor results should not be interpreted as proving strict unidimensionality. Rather, they suggest that the three content domains are best understood as closely related descriptive domains within an overarching construct of nurses’ AI ethical awareness, and that domain-level scores should be interpreted cautiously until further validation is available. Because all items were self-reported and administered in a similar Likert-type format, common response tendencies or shared method variance may also have contributed to the strength of the general factor. Therefore, the dominant general factor should be regarded as evidence supporting total-score interpretation, not as definitive proof that AI ethical awareness is strictly unidimensional. The emergence of three closely related content domains may also be conceptually plausible in the nursing context. In real clinical situations, nurses may not evaluate AI-related ethical issues as sharply separated categories. Nursing clinical judgement is generally contextual, relational, and practice-based rather than a purely technical response to isolated information [21]. Evidence from Chinese nursing ethics research also suggests that moral sensitivity and ethical decision-making are shaped by hospital ethical climate, supporting the view that ethical judgement in nursing is embedded in organisational and practice contexts rather than formed as isolated abstract reasoning [22]. Concerns about accountability, bias monitoring, safety checking, patient rights, transparency, informed decision-making, and responsible implementation are often interwoven in practical care decisions. When nurses judge whether AI use is appropriate, they may consider whether AI-generated recommendations are explainable, whether patient preferences are respected, whether professional responsibility remains clear, and whether AI use may compromise safety or fairness. This integrated nature of ethical judgement may partly explain why the retained domains remained interpretable while showing substantial overlap. This interpretation may be particularly relevant in hospital settings where nurses encounter AI-related technologies mainly through workflow-embedded systems, such as clinical decision support, electronic documentation, intelligent warning systems, and digitally mediated nursing services. In these situations, ethical concerns are rarely presented as isolated theoretical principles; instead, they are embedded in daily decisions about trust, responsibility, communication, documentation, and patient safety. Broader AI ethics theory emphasises that autonomy, justice, explicability, beneficence, and non-maleficence should be considered together when evaluating AI systems [23]. In healthcare, machine-learning implementation also raises practical ethical challenges related to safety, trust, accountability, and the distribution of responsibility [24]. These perspectives support the interpretation that nurses’ AI ethical awareness may operate as an integrated form of practical judgement rather than as a set of sharply separated ethical categories. This finding can also be understood in relation to the broader landscape of AI-related nursing and ethical-awareness instruments. Existing AI-related nursing scales have mainly focused on attitudes toward AI technologies, technology acceptance, or readiness for AI use, rather than on ethical awareness as a profession-specific construct. For example, instruments assessing attitudes toward the use of AI technologies in nursing have been developed and culturally adapted, including a Chinese version, but their primary focus is attitude toward AI use rather than ethical awareness in AI-supported nursing care [25, 26]. More general AI ethical-awareness instruments have examined ethical awareness across broader AI-related populations and contexts rather than specifically addressing bedside nursing responsibilities [27]. Compared with these instruments, the Nurses’ AI Ethical Awareness Scale is distinctive because it centres on ethical responsibilities embedded in AI-supported nursing care, including patient rights, disclosure, professional accountability, safety oversight, bias awareness, responsible implementation, and continuing ethical learning. The present findings therefore suggest that, in the Chinese hospital nursing context, AI ethical awareness may be better represented as a dominant general construct with closely related practical domains than as a set of sharply independent ethical subdimensions. Item-level interpretation and cultural adaptation The reduction from 21 to 15 items was based on multiple sources of evidence, including item-level diagnostics, EFA loading patterns, cross-loadings, candidate-version model comparison, content representativeness, and theoretical interpretability. Several deleted items showed clear item-level or structural concerns. Q1 and Q2 showed weak EFA loadings, Q4 showed increased Cronbach’s α if item deleted and contributed less clearly to the retained structure, and Q9 showed problematic cross-loading and structural ambiguity. These items were therefore removed because they did not contribute clearly to the emerging Chinese factor structure. However, not all removed items showed weak standalone psychometric performance, and Q3 and Q18 require more nuanced interpretation. Q3 addressed privacy-related content and showed acceptable corrected item–total correlation and EFA loading values. Nevertheless, retaining a single privacy-related item would not have supported a stable and interpretable privacy/data-governance component in the abbreviated scale. Q18 also showed acceptable standalone performance, but its content overlapped with retained items related to professional accountability, clinical judgement, workflow appropriateness, and safety oversight, particularly items addressing clinician responsibility, verification of AI-generated recommendations, and reporting of AI-related safety problems. Therefore, the deletion of Q3 and Q18 should not be interpreted as evidence that these ethical topics are unimportant; rather, these decisions reflected short-form refinement at the scale level rather than item-level statistical weakness alone. Conversely, Q16 and Q17 were retained despite weaker loading patterns in the initial 21-item EFA, and Q17 also showed cross-loading. This decision was made because both items represented core nursing responsibilities in AI-supported practice. Q16 reflects the need to prioritise professional judgement when AI advice conflicts with clinical assessment, whereas Q17 reflects safety reporting and accountability when AI-related problems or adverse events occur. Both items also showed satisfactory standardised loadings in the final 15-item CFA model. Their retention therefore helped preserve clinically important content that might have been lost if item reduction had been based only on initial EFA thresholds. These findings highlight that item deletion in cross-cultural scale adaptation should not be understood as a purely statistical process. Some items may function differently after translation because of differences in clinical language, professional norms, regulatory expectations, hospital information-governance practices, and nurses’ familiarity with AI-related systems. In health AI governance, privacy, data governance, transparency, human oversight, and accountability are often positioned as system-level responsibilities involving both professionals and organisations [28]. Trustworthy-AI guidance similarly treats privacy and data governance, transparency, human agency, oversight, and accountability as requirements that need to be embedded across the implementation environment [29]. Chinese nursing ethics literature has framed AI-related ethical issues in clinical nursing as involving data, patients, healthcare professionals, technologies and equipment, and laws and regulations, rather than as isolated bedside communication issues [30]. This broader framing supports the interpretation that nurses may understand privacy, data governance, transparency, and accountability as responsibilities distributed across clinical practice, organisational systems, and regulatory structures. In the present sample, privacy- and data-governance-related items may therefore have been interpreted partly as institutional, regulatory, or technical responsibilities rather than as direct bedside nursing responsibilities, whereas items concerning patient communication, professional judgement, safety checking, accountability, and continuing ethical learning may have been more closely aligned with nurses’ daily clinical roles. Importantly, the removal of privacy- and data-governance-related items should not be interpreted as indicating that these issues are unimportant to nurses in Chinese clinical contexts or irrelevant to AI-supported nursing practice. Rather, these findings suggest that the translated items may not have captured privacy, confidentiality, and data-governance responsibilities in a way that was psychometrically stable or sufficiently aligned with nurses’ perceived clinical roles in this sample. Therefore, the retained 15-item version should be interpreted as assessing several practice-facing aspects of nurses’ AI ethical awareness, including patient rights, informed disclosure, professional accountability, bias awareness, safety oversight, responsible implementation, and continuing ethical learning, rather than as an exhaustive measure of all AI ethics domains. Further studies, including cognitive interviewing or qualitative work, should refine Chinese-language items addressing privacy and information governance and evaluate whether these items perform differently across hospitals, regions, and levels of digital health implementation [31, 32]. Reliability, convergent validity, and discriminant validity The adapted 15-item Chinese version demonstrated high internal consistency in the full sample and satisfactory convergent validity in the CFA subsample. The total Cronbach’s α was high, and the retained items showed strong standardised factor loadings, composite reliability values above the recommended threshold, and AVE values greater than 0.50. These findings indicate that the retained items coherently measured nurses’ awareness of ethical issues related to AI-supported nursing practice. However, very high internal consistency should not be interpreted as sufficient evidence of multidimensional independence, because Cronbach’s α is sensitive to item interrelatedness and does not by itself establish construct validity [33]. The reliability findings should therefore be interpreted together with dimensionality and discriminant-validity evidence. In this study, inter-factor correlations were substantial, the Fornell–Larcker criterion was not fully satisfied, and one HTMT value exceeded the stricter 0.85 criterion, although all HTMT values remained below the more lenient 0.90 threshold. These findings suggest that discriminant validity was only partially supported. This interpretation is consistent with updated recommendations that discriminant validity should be evaluated using dedicated validity criteria and multiple sources of evidence rather than inferred from reliability coefficients alone [34]. This pattern is consistent with the bifactor results and supports a total-score-first interpretation. The three domains appear to represent clinically meaningful but overlapping aspects of AI ethical awareness rather than clearly separable subscales. Bifactor modelling is useful for evaluating construct-relevant multidimensionality and for distinguishing whether scale-score variance is mainly attributable to a general factor or to independently interpretable specific factors [35]. Accordingly, the domain scores may be used cautiously for descriptive or educational purposes, whereas the total score should be prioritised for formal research interpretation until further evidence from multicentre validation, criterion-related testing, and measurement invariance analysis becomes available. Score interpretation, recommended use, and practical implications Based on the structural and validity evidence described above, the total score should be regarded as the primary scoring approach for the adapted 15-item Chinese version. This score provides the most stable summary of nurses’ overall awareness of ethical issues related to AI-supported nursing practice and is therefore most suitable for research, education, and organisational assessment. The three domain scores may still provide useful descriptive information, but they should not be interpreted as fully independent subscales at this stage. Domain-level results may help identify broad educational needs, such as strengthening patient communication and consent, reinforcing professional accountability and safety oversight, or promoting continuing ethical learning about AI use. However, these interpretations should remain exploratory and supplementary until further validation confirms the independent reliability, discriminant validity, and external relevance of the domain scores. In practice, the Chinese 15-item version may provide a concise tool for assessing practice-facing aspects of nurses’ AI ethical awareness, identifying training needs, evaluating ethics-focused education, and supporting organisational discussions about ethically governed AI implementation. AI ethics scholarship has emphasised the need to translate abstract principles into usable tools, methods, and implementation processes [36]. Nursing-specific guidance also highlights patient advocacy, professional accountability, human oversight, and safety as key responsibilities when AI is used in nursing practice [37]. The current findings do not support the use of rigid cut-off values or strong claims about independent subscale profiles. Future multicentre studies should further evaluate score stability, measurement invariance, criterion-related validity, known-groups validity, predictive validity, and applicability across different hospital levels, lower-tier cities, and rural nursing contexts. Strengths and limitations This study has several strengths. First, it followed a systematic translation and cross-cultural adaptation process, including expert review and pilot testing. Second, EFA and CFA were conducted in independent subsamples, strengthening the credibility of the structural evaluation. Third, item reduction was based on both statistical and conceptual criteria rather than a single model-fitting rule. Fourth, the use of alternative structural models and bifactor indices allowed a more nuanced interpretation of dimensionality and score use. Several limitations should be acknowledged. First, the cross-sectional design precluded assessment of test–retest reliability and temporal stability. Second, participants were recruited by convenience sampling from a single tertiary hospital in Shenzhen, which may limit generalisability to nurses in lower-tier cities, county-level hospitals, community health institutions, or rural settings. Third, discriminant validity was only partially supported; therefore, the total score should be regarded as the primary scoring approach, whereas subscale scores should be interpreted cautiously. Fourth, deletion of several privacy- and information-governance-related items may have narrowed content coverage, and these items require further refinement in Chinese nursing contexts. Finally, future studies should verify the scale using larger multicentre samples, robust ordinal estimators, measurement invariance testing, external criterion validation, and strategies to reduce potential social desirability bias. Conclusion The present study provides preliminary support for an adapted 15-item Chinese version of the Nurses’ AI Ethical Awareness Scale as a concise instrument for hospital nurses. The retained version demonstrated high internal consistency in the full sample and satisfactory convergent validity in the CFA subsample, and showed three closely related content domains and a dominant general factor. Accordingly, the total score appears to provide the most robust representation of nurses’ AI ethical awareness, whereas subscale scores should be interpreted with greater caution. Privacy, confidentiality, and data-governance dimensions require further item refinement and validation before they can be adequately represented in the Chinese version. This instrument may be useful for assessing practice-facing aspects of nurses’ ethical awareness regarding AI use, identifying educational needs, and supporting ethically oriented AI implementation in nursing research, education, and organisational practice. Further multicentre validation is needed to confirm its structure, score interpretation, and applicability across diverse nursing contexts. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1