Cross-cultural adaptation and psychometric evaluation of the Shinners Artificial Intelligence Perception Scale (SAIP): evidence from Turkish nursing students

Abstract

Background

Integrating artificial intelligence (AI) technologies into nursing practice is expanding rapidly, yet validated instruments to assess nursing students’ perceptions of AI remain scarce in non-English-speaking contexts. The Shinners Artificial Intelligence Perception (SAIP) Scale was developed to measure healthcare professionals’ perceptions of AI; however, its applicability in Turkish culture has not been tested.

Methods

This methodological study was conducted between February and July 2025 and involved 412 nursing students from three public nursing faculties in Ankara, Türkiye. The cross-cultural adaptation process followed forward–backward translation, expert review, and content validity assessment using the Davis technique. Construct validity was examined through exploratory factor analysis (EFA) and confirmatory factor analysis (CFA). Reliability was assessed using item–total correlations, Cronbach’s alpha, and test–retest analysis with a subgroup of 32 participants.

Results

Content validity index (CVI) was 0.88. The KMO sample adequacy was 0.802, and Bartlett’s test yielded χ²(36) = 651.198 (p < 0.001). EFA supported a two-factor structure, explaining 48.1% of the total variance. CFA supported the two-factor structure, with incremental fit indices (CFI = 0.968, TLI = 0.954, IFI = 0.969) and absolute fit indices (χ²/df = 1.608, GFI = 0.966, SRMR = 0.064) indicating generally acceptable fit. However, the RMSEA value of 0.090 (90% CI: 0.065–0.122) suggests borderline model fit, as the upper limit of the confidence interval slightly exceeds commonly recommended thresholds. Cronbach’s alpha values were 0.772 for the overall scale, 0.763 for Factor 1 (Perception of AI’s impact on professional role), and 0.723 for Factor 2 (Perceptions of preparedness for AI). Test–retest reliability demonstrated significant positive correlations (r = 0.584–0.752, p < 0.001) and no significant differences between the two measurement points.

Conclusions

In this study, psychometric evaluation yielded an 8-item Turkish version of the SAIP (SAIP-TR); thus, the adapted instrument should be considered a shortened form of the original scale. The SAIP-TR may provide a useful framework for assessing AI-related perceptions among nursing students and may support educators and researchers in understanding students’ readiness for AI integration into nursing practice and in informing future educational strategies. However, further validation studies in more diverse populations are recommended to strengthen the scale’s generalizability and conceptual stability.

Clinical trial number

Not applicable.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12912-026-04926-z.

Introduction

Artificial intelligence (AI) systems, designed using algorithms such as machine learning, deep learning, and cognitive computing, can analyze large datasets effectively and provide timely, reliable solutions to complex healthcare problems [1–3]. The use of AI in healthcare is expanding rapidly, supporting accurate and cost-effective delivery of diagnosis and treatment services [4–7]. Research across various specialties (e.g., pathology, cardiology, radiology, and endocrinology) has shown that AI applications can identify diseases and complications with accuracy comparable to that of healthcare professionals [8–11]. Moreover, AI has been reported to guide medication selection and dosage adjustment, contributing to safer care and higher patient satisfaction [12, 13].

The integration of AI into nursing practice has also gained momentum. By processing patient data, AI tools support nurses in identifying accurate nursing diagnoses, selecting appropriate interventions, and developing individualized care plans [14–16]. Furthermore, AI-based clinical decision support may reduce preventable adverse events by providing reminders and evidence-based recommendations, thereby improving patient safety [17–19]. On the other hand, nurses acknowledge ethical concerns regarding transparency in AI-driven decision-making, data security, and patient privacy, highlighting the dual opportunities and risks of AI adoption [20–26].

The concept of AI perception can be evaluated within established theoretical frameworks of technology adoption and technology readiness. The Technology Readiness Index (TRI) explains individuals’ propensity to adopt and use new technologies through psychological dimensions such as optimism, innovativeness, discomfort, and insecurity [27]. Similarly, the Technology Acceptance Model (TAM and TAM2) emphasizes that the primary determinants of attitudes toward and acceptance of technology—particularly in healthcare contexts—are perceived usefulness and perceived ease of use [28, 29]. However, more recent approaches consider artificial intelligence literacy not merely a technical competency but a multidimensional capability encompassing cognitive, ethical, and practical dimensions [30]. Accordingly, AI perception represents a profession-oriented construct that combines evaluations of AI-related role transformation with perceptions of preparedness to function in AI-enabled healthcare environments. Unlike general AI attitude measures, AI perception focuses on how individuals anticipate the impact of AI on their professional practice and their readiness to engage with AI in clinical settings.

In nursing, AI perception refers to how nurses evaluate the implications of artificial intelligence for professional practice, clinical decision-making, and ethical responsibilities [31]. Although AI perception is theoretically related to constructs such as perceived usefulness, role change, technology readiness, and AI literacy, it reflects a broader evaluation of how AI may influence professional practice and preparedness for future clinical environments. Therefore, assessing AI perception in nursing is critical to understanding the integration of digital health technologies into clinical practice.

As the future workforce, nursing students will work in technology-intensive healthcare environments in which AI-supported tools are increasingly integrated into clinical workflows. Therefore, understanding nursing students’ perceptions of AI is crucial for developing educational strategies and curricula that support the safe, ethical, and effective integration of AI technologies into nursing practice. Various instruments have been developed to measure healthcare professionals’ or students’ attitudes toward AI [32–34]. However, these instruments are not designed to directly assess nursing students’ perceptions of AI or how these perceptions relate to professional roles. In contrast, the Shinners Artificial Intelligence Perception (SAIP) Scale focuses on profession-specific anticipatory appraisals of AI [35], capturing perceived impact on the professional role and work processes, as well as perceived preparedness/self-efficacy to engage with AI in practice. Therefore, the present study aimed to adapt the SAIP Scale into Turkish and examine its psychometric properties in a sample of Turkish nursing students. Obtaining a valid and reliable Turkish version of the SAIP may contribute to the systematic assessment of perceptions of artificial intelligence and related readiness levels in nursing education. In turn, this may enable educators and researchers to better understand students’ attitudes toward AI-based healthcare technologies and to develop appropriate educational strategies.

Methods

Study design

This methodological study was conducted to adapt the SAIP Scale into Turkish and to evaluate its psychometric properties. The study adhered to internationally accepted recommendations for cross-cultural adaptation and instrument validation, including those outlined by the COSMIN initiative [36].

Setting and participants

The study was conducted in three public nursing faculties in Ankara, Türkiye. A multi-center design with non-probability convenience sampling was adopted to reduce potential bias associated with single-institution sampling and to enhance sample diversity. The target population consisted of undergraduate nursing students enrolled in these three faculties during the 2024–2025 academic year, excluding those in the preparatory class (N = 2450). Furthermore, since the sample was drawn from students at metropolitan public universities with potentially higher technological exposure, this context may have influenced AI-related perceptions and should be considered regarding the generalizability of the findings.

Sample size was determined a priori based on the recommendation of 20 participants per item to ensure robust exploratory and confirmatory factor analyses, yielding a minimum target of 400 participants for this 10-item scale [37–39]. In total, 412 students participated in the study, exceeding the recommended thresholds. For test–retest reliability, 32 students were re-assessed after two weeks. However, given the small scale of the test–retest subgroup (n = 32), temporal stability estimates should be interpreted cautiously and viewed as preliminary evidence.

Inclusion criteria

actively enrolled in nursing education and continuing/having participated in at least one clinical practice.

Exclusion criterion

actively employed in healthcare delivery (e.g., nurse aide, nurse assistant).

Instrument

Demographic information form

The researchers developed this form to collect information on nursing students’ age, gender, grade point average (GPA), training in artificial intelligence, and use of artificial intelligence.

Shinners artificial intelligence perception scale

The original SAIP Scale was developed to assess healthcare professionals’ perceptions of artificial intelligence [35] and has been used in the literature to determine student nurses’ perceptions of artificial intelligence [40]. It consists of 10 items across two subscales: Perception of AI’s impact on professional role and Perceptions of preparedness for AI. Responses are rated on a 5-point Likert scale from 1 (strongly disagree) to 5 (strongly agree). The original validation study reported Cronbach’s alpha as 0.82 for the overall scale [35].

Translation and cultural adaptation

The adaptation process followed a forward–backward translation procedure. Two bilingual experts translated the scale into Turkish. A Turkish language and literature specialist reviewed the translations for clarity and comprehensibility. Two independent bilingual experts then performed back-translation into English. The research team reviewed and compared the original and back-translated versions of the scale to confirm semantic equivalence, and subsequently produced the final version of the scale. In the final step, the finalized version of the scale was submitted to the original developer, and formal approval regarding its suitability was received via email.

The content validity of the scale was assessed using the Davis technique based on the opinions of six experts in nursing education and clinical practice. Experts were asked to rate each item on a 4-point scale (1 = not suitable, 2 = needs revision, 3 = suitable but requires minor revision, 4 = entirely appropriate). The item-level Content Validity Index (I-CVI) was calculated as the proportion of experts who rated the item 3 or 4, divided by the total number of experts. The scale-level Content Validity Index (S-CVI) was computed as the arithmetic mean of the I-CVI values across all items. In addition, feedback was obtained from experts who rated items as 2 or 3 to improve the item statements. The authors comprehensively evaluated expert opinions and made necessary revisions to the scale items in accordance with the recommendations.

Pilot study

A pilot study was conducted with 10 students to assess the clarity and comprehensibility of the scale items. No negative feedback or suggestions for improvement regarding the scale’s comprehensibility and applicability were reported by participants following the pilot study.

Data collection

Data were collected face-to-face in the classroom between February and July 2025, at times that did not interfere with courses or examinations. In coordination with the school administrations, participants were approached after the completion of their theoretical classes, taking their academic timetables into account. The study was conducted in accordance with the Declaration of Helsinki and approved by the Ethics Committee of Hacettepe University Health Sciences Ethics Committee, Ankara, Turkey (Decision No: 2025/03–05; Study ID: SBA 24/1128). Students were informed about the purpose of the study, the voluntary of participation, and the confidentiality of their responses. Students who volunteered provided written consent. For the test–retest analysis, 32 students completed the scale again two weeks later.

Data analysis

All analyses except CFA were performed using IBM SPSS Statistics 30. CFA was conducted using JASP 0.95.2 [41]. Descriptive statistics included mean, standard deviation, median, interquartile range, frequency, and percentage.

To avoid conducting EFA and CFA on the same sample, the dataset was randomly split into two homogeneous subgroups: one for EFA (n = 206) and one for CFA (n = 206), consistent with established recommendations [42]. Construct validity was examined using EFA with Principal Axis Factoring and Promax rotation, after confirming sampling adequacy with the Kaiser–Meyer–Olkin (KMO) measure and Bartlett’s test of sphericity. Factor retention was evaluated using multiple criteria, including the eigenvalue > 1 rule, visual inspection of the scree plot, theoretical interpretability, and consistency with the original scale structure. Items with communalities < 0.30 were excluded.

CFA was subsequently conducted to test the model fit of the extracted factor structure using the Weighted Least Squares Mean and Variance adjusted (WLSMV) estimator, which is appropriate for ordinal data. Model fit was evaluated using χ²/df, GFI, CFI, NFI, TLI, IFI, RMSEA, and SRMR. Acceptable thresholds were set as χ²/df ≤ 3, CFI/TLI/NFI/IFI/GFI ≥ 0.95, RMSEA ≤ 0.10, and SRMR ≤ 0.08 [43, 44]. Convergent validity was assessed using Average Variance Extracted (AVE) and Composite Reliability (CR) values. Discriminant validity was examined using the Fornell-Larcker criterion and the Heterotrait-Monotrait (HTMT) ratio.

Reliability was assessed through item–total correlations, Cronbach’s alpha, and McDonald’s omega coefficients for the overall scale and subscales, and test–retest reliability. Test–retest reliability was evaluated using Pearson correlation and the intraclass correlation coefficient (ICC; two-way mixed model, absolute agreement) [44]. A p-value < 0.05 was considered statistically significant.

Results

Participant characteristics

The mean age was 20.87 ± 1.63, and 84.7% were female. The distribution by academic year was as follows: first year (24,3%), second year (9,0%), third year (27,2%), and fourth year (39,6%). Most students reported daily internet use exceeding 3 h, while only a minority had prior experience with AI technologies (Table 1).

Table 1.

Demographic and technology-related characteristics of participants (n = 412)

Variable	n (%) / Mean
Age (years)	20.87 ± 1.6 (Min-Max:18–25)
Gender	Female: 360 (87.4%) / Male: 52 (12.6%)
Academic year	1st: 100 (24.3%)
	2nd: 37 (9.0%)
	3rd: 112 (27.2%)
	4th: 163 (39.6%)
Academic achievement score (GPA)*	< 2: 4 (1.0%)
	2–2.5:54 (13.1%)
	2.51–3: 223 (54.1%)
	3.01–3.5:123 (29.9%)
	> 3.51: 8 (1.9%)
Daily internet use	1–3 h: 74 (18.0%)
	4–6 h: 233 (56.6%)
	> 6 h: 105 (25.4%)
Prior AI education	Yes: 12 (2.9%)	No: 400 (97.1%)
Prior experience with AI	Yes: 329 (79.9%)	No: 83 (20.1%)
Willingness to receive AI training	Yes: 220 (53.4%)	No: 192 (46.6%)
Willingness to use AI in nursing practice	Not willing at all: 10 (2.4%)
	Slightly willing: 30(7.3%)
	Moderately willing: 165 (40.0%)
	Willing: 158 (38.3%)
	Very willing: 49 (11.9%)
AI application(s) used**	None: 83 (20.1%) ChatGPT: 252 (61.2%) Google Assistant: 40 (9.7%) Siri: 21 (5.1%) Other: 13 (3.2%) Cortana: 3 (0.7%)
Purpose(s) of AI use**	Research/Information seeking: 82 (19.9%) Course or assignment support: 66 (16.0%) Academic or educational use: 49 (11.9%) Question–answer/problem solving: 39 (9.5%) Daily life support: 33 (8.0%) Social or entertainment: 26 (6.3%) Language learning: 20 (4.9%) Creative or design purposes: 13 (3.2%)

*Self-reported GPA on a 4.0 scale

**Multiple responses allowed; percentages may exceed 100%

Content validity

Expert evaluations indicated satisfactory content validity. I-CVI values ranged between 0.83 and 1.00, and the overall S-CVI was 0.88. Following expert suggestions, minor wording changes were made to items 4 and 7.

Exploratory factor analysis

The suitability of the data for exploratory factor analysis was confirmed by a KMO value of 0.802 and a significant Bartlett’s test of sphericity (ꭓ²= 651.198, df = 36, p < 0.001). Principal Axis Factoring with Promax rotation was used to identify the factor structure. During the analysis, Item 10 was excluded as its communality (0.195) was below the threshold of 0.30. Visual inspection of the scree plot indicated a clear inflection after the second factor, supporting retention of a two-factor solution (Supplementary Figure. 1). The final 9-item scale yielded a two-factor structure, accounting for 48.071% of the total variance. While this level of explained variance is considered adequate for psychometric instruments in social science research, it reflects a modest factor structure. Factor 1 consisted of six items (Items 1, 2, 3, 4, 5, and 6) with factor loadings ranging from 0.566 to 0.797. Factor 2 consisted of three items (Items 7, 8, and 9) with factor loadings ranging from 0.482 to 0.810. Although Item 10 was removed for psychometric reasons, the resulting two-factor structure remained broadly aligned with the primary dimensions of the original SAIP. However, some theoretically relevant content regarding professional role displacement and ethical accountability was not retained in this version. This 9-item structure was subsequently utilized for confirmatory factor analysis (Table 2).

Table 2.

Exploratory factor analysis and sampling adequacy and sphericity tests results (n = 206)

Items	Factor loadings
	Factor 1	Factor 2
Item 2	0.797
Item 3	0.734
Item 1	0.676
Item 5	0.658
Item 4	0.621
Item 6	0.566
Item 8		0.810
Item 7		0.766
Item 9		0.482
Explained variance (%)	35.898	12.173
Cumulative variance (%)	35.898	48.071
Sampling adequacy and sphericity tests
Kaiser–Meyer–Olkin measure of sampling adequacy	0.802
Bartlett’s test of sphericity	ꭓ2=651.198
Degrees of freedom (df)	df = 36
Significance level (p)	p < 0.001

Confirmatory factor analysis

Confirmatory factor analysis (CFA) was conducted to validate the two-factor structure. In the initial assessment, Item 6 was excluded due to high error covariance and conceptual redundancy within its dimension. Furthermore, a covariance was specified between the error terms of Item 4 (“belief in AI changing the professional role”) and Item 5 (“belief in AI reducing job-related costs”) (r = 0.42) as shown in Fig. 1. This adjustment was theoretically justified as participants likely perceived the transformation of professional roles and economic efficiency as interconnected facets of AI’s future systemic impact. The final 8-item model comprised five items (Items 1, 2, 3, 4, 5) loading on Factor 1 and three items (Items 7, 8, 9) loading on Factor 2 (Fig. 1).

Fig. 1.

Path diagram of the confirmatory factor analysis (CFA) model of the SAIP-TR

The CFA results are presented in Table 3, which is divided into three sections: Part A (Factor Loadings), Part B (Model Fit Indices), and Part C (Discriminant Validity). In Part A, standardized estimates (ranging from 0.515 to 0.871), standard errors (SE), z-values, and 95% confidence intervals (CI) are reported, all of which were statistically significant (p < 0.001). In Part B, the results demonstrated generally acceptable/borderline model fit. The incremental fit indices indicated good fit (CFI = 0.976, TLI = 0.963, NFI = 0.962, IFI = 0.976, and GFI = 0.993), and the χ²/df ratio was within recommended limits (χ²/df = 2.67). The SRMR value (0.056) also indicated an acceptable fit. However, the RMSEA value (0.090) suggested borderline fit, indicating that some degree of model misspecification cannot be entirely ruled out. Construct validity was further supported by discriminant validity results presented in Part C of Table 3.

Table 3.

Confirmatory factor analysis results: factor loadings, model fit indices, and discriminant validity (n = 206)

Part A: Factor Loadings
Factors	Indicator	Std. Estimate	Std. Error	z-value	p	95% CI Lower	95% CI Upper
Factor 1	Item 1	0.772	0.033	23.134	< 0.001	0.707	0.838
	Item 2	0.871	0.027	32.527	< 0.001	0.818	0.923
	Item 3	0.796	0.031	26.082	< 0.001	0.736	0.856
	Item 4	0.518	0.054	9.539	< 0.001	0.411	0.624
	Item 5	0.515	0.053	9.809	< 0.001	0.412	0.618
Factor 2	Item 7	0.736	0.041	17.758	< 0.001	0.655	0.817
	Item 8	0.784	0.041	19.328	< 0.001	0.705	0.864
	Item 9	0.707	0.047	15.121	< 0.001	0.615	0.798
Part B: Model Fit Indices
χ²/df	CFI	TLI	NFI	IFI	RMSEA	SRMR	GFI
2.67 (< 3.00)	0.976 (≥ 0.95)	0.963 (≥ 0.95)	0.962 (≥ 0.95)	0.976 (≥ 0.95)	0.090 (≤ 0.10)	0.056 (≤ 0.08)	0.993 (≥ 0.95)
Part C: Discriminant Validity
Variables		Factor 1			Factor 2
Factor 1		(*0.504/0.710)			—
Factor 2		0.420 / 0.419			(*0.552/0.743)

χ²/df = Chi-square/degrees of freedom ratio (χ² = 48.002, df = 18, p < 0.001); CFI = Comparative Fit Index; TLI = Tucker-Lewis Index; NFI = Normed Fit Index; IFI = Incremental Fit Index; RMSEA = Root Mean Square Error of Approximation (90% CI [0.060, 0.122]); SRMR = Standardized Root Mean Square Residual; GFI = Goodness of Fit Index. Values in parentheses indicate threshold criteria. Diagonal values in Part C represent square root of *AVE. Below diagonal: Fornell-Larcker criterion / HTMT ratio (threshold < 0.85)

The Average Variance Extracted (AVE) values for Factor 1 (0.504) and Factor 2 (0.552) met the 0.50 threshold. Furthermore, the HTMT ratio (0.419) remained well below the 0.85 criterion, and the inter-factor correlation (0.420) was lower than the square root of the AVE for each factor (0.710 and 0.743), satisfying the Fornell-Larcker criterion (Table 4).

Table 4.

Item–total correlations and reliability analysis of the SAIP-TR

Items	Factor 1 item–total correlation	Factor 2 item–total correlation	Corrected item–total correlation	Cronbach’s alpha α (≥ 0.70)	McDonald’s Omega ω (≥ 0.70)	Composite Reliability CR (≥ 0.70)	Average Variance Extracted AVE (≥ 0.50)
Item 1	0.569		0.505
Item 2	0.646		0.574
Item 3	0.615		0.562	0.763	0.757	0.830	0.504
Item 4	0.410		0.432
Item 5	0.471		0.404
Item 7		0.571	0.430
Item 8		0.597	0.439	0.723	0.734	0.787	0.552
Item 9		0.468	0.486
			Total	0.772	0.751

Reliability analysis

The internal consistency and construct reliability of the SAIP-TR were rigorously evaluated using multiple indicators. As shown in Table 4, the Cronbach’s alpha coefficients were 0.772 for the total scale, 0.763 for Factor 1, and 0.723 for Factor 2. To address the limitations of alpha for short scales, McDonald’s omega (omega) and Composite Reliability (CR) were also calculated. For Factor 1, omega was 0.757, and CR was 0.830; for Factor 2, omega was 0.734, and CR was 0.787. Additionally, the Average Variance Extracted (AVE) values exceeded the 0.50 threshold (Factor 1 = 0.504; Factor 2 = 0.552), confirming convergent validity. Corrected item–total correlations ranged between 0.404 and 0.574, indicating that all items contributed significantly to their respective factors (Table 4).

Temporal stability was assessed via test–retest reliability with 32 participants over a two-week interval. As presented in Table 5, the Intraclass Correlation Coefficient (ICC) values confirmed strong stability for the total scale (ICC = 0.852; 95% CI = 0.697–0.927), Factor 1 (ICC = 0.819; 95% CI = 0.633–0.911), and Factor 2 (ICC = 0.734; 95% CI = 0.461–0.869). Complementary Pearson correlation coefficients also demonstrated significant positive relationships for the total scale (r = 0.752, p < 0.001), Factor 1 (r = 0.697, p < 0.001), and Factor 2 (r = 0.584, p < 0.001). Furthermore, paired t-tests showed no significant mean difference between test and retest scores (total scale t = 1.685, p = 0.102), supporting the instrument’s reliability over time.

Table 5.

Test–retest reliability results of the SAIP-TR and its subscales (n = 32)

Measure	Total Scale	Factor 1	Factor 2
Test (x̄±SD)	3.20 ± 0.520	3.74 ± 0.601	2.66 ± 0.639
Retest (x̄±SD)	3.09 ± 0.518	3.64 ± 0.594	2.54 ± 0.518
Paired t-test	1.685	1.215	1.235
p-value	0.102	0.234	0.226
ICC	0.852	0.819	0.734
95% CI	0.697–0.927	0.633–0.911	0.461–0.869
Pearson r	0.752	0.697	0.584
p-value	< 0.001	< 0.001	< 0.001
95% CI (r)	0.547–0.872	0.460–0.841	0.295–0.775

x̄ = mean; SD = standard deviation; ICC = Intraclass Correlation Coefficient; CI = Confidence Interval; r = Pearson correlation coefficient

Discussion

This study adapted the SAIP Scale into Turkish and evaluated its psychometric properties among nursing students. The findings indicate that the Turkish version (SAIP-TR), refined as an 8-item adapted scale, demonstrates adequate validity and reliability. While the final structure required the exclusion of items related to professional role displacement and ethical accountability to achieve acceptable model fit, the SAIP-TR remains a conceptually relevant instrument for assessing core dimensions of AI perception in the Turkish context.

Expert evaluation was conducted to determine the extent to which the scale items accurately reflect the construct under study. Based on the experts’ ratings, the Content Validity Index (CVI) was calculated, yielding a value exceeding 0.80. A CVI above this threshold is generally regarded as evidence of strong content validity [45, 46]. Overall, these findings indicate that the scale provides a valid and meaningful assessment of the construct under investigation.The Kaiser–Meyer–Olkin (KMO) measure of sampling adequacy and Bartlett’s test of sphericity confirmed that the dataset was suitable for factor analysis. The KMO value indicated good sampling adequacy [37–39].

Although most fit indices indicated acceptable-to-good fit, the RMSEA (0.090) with an upper confidence limit of 0.122 suggests that model fit may be borderline and that some degree of misspecification cannot be ruled out. In addition, the final model was obtained after post hoc steps (item deletion and a correlated residual), which improved fit but may also increase the risk of overfitting to the present CFA subsample. Therefore, the structural solution should be interpreted as a well-supported candidate model rather than a definitive representation of the SAIP structure in Türkiye.

Item 6 (“I believe that one day AI may take over part of my role as a healthcare professional”) and Item 10 (“I believe that should AI technology make an error, full responsibility lies with the healthcare professional”) were excluded due to inadequate psychometric performance. The original scale also reported relatively low factor loadings for item 10 [35]. This suggests that participants may have interpreted these items differently, both conceptually and contextually. Item 6 presents a forward-looking view of AI replacing healthcare professionals’ roles. However, because the integration of AI applications into nursing care processes in Türkiye is still limited, students may have found it difficult to relate this statement to concrete clinical experiences. As a result, responses to this item may have relied more on hypothetical judgments and may not have been adequately represented within the factor structure. Item 10 addresses the ethical and legal dimensions of who bears responsibility for errors made by AI technologies. Nursing students’ limited professional experience may make it difficult to evaluate such complex accountability issues. In addition, the lack of clearly established legal frameworks regulating liability related to AI use in the Turkish healthcare system may have hindered participants from providing consistent responses. This finding suggests that themes such as ethical accountability and role replacement may be more ambiguous and conceptually challenging across cultural contexts.

Theoretically, the two-factor structure of the SAIP-TR can be understood in relation to the TAM, TRI, and AI literacy frameworks. The first factor, which examines the impact of AI on professional roles, shares similarities with constructs such as perceived usefulness and technology-related attitudes. The second factor, measuring preparedness for AI, overlaps with dimensions of AI literacy and technology readiness, such as confidence and competence. However, since a direct model test was not conducted in this study, these relationships remain conceptual rather than empirical. Furthermore, unlike general technology acceptance models, the SAIP-TR possesses a more profession-specific structure. The scale evaluates nursing students’ perceptions regarding the impact of AI on professional roles and their perceived preparedness for AI integration into clinical practice. Finally, it should be noted that due to the two items removed during the adaptation process, the Turkish version may not fully reflect certain theoretical dimensions of the original scale, such as professional role displacement and ethical accountability. Accordingly, the SAIP-TR may be best viewed as an adapted 8-item version of the original instrument that preserves its core conceptual domains while providing a somewhat narrower representation of AI perception.

The factor structure of the SAIP-TR explained 48% of the total variance, representing an adequate yet improvable framework. The scale’s reliability coefficients and temporal stability were found to be satisfactory [45]. Notably, the internal consistency of the Turkish version was higher than that of the original scale [35], which may be attributed to the homogeneity of the participant group. However, considering the proportion of unexplained variance, it is evident that factors such as trust in artificial intelligence, ethical attitudes, and prior experience play a critical role in shaping students’ perceptions. This suggests that future research should focus on testing more comprehensive theoretical models.

Limitations

This study has several limitations. First, the research was conducted exclusively with nursing students and did not include healthcare professionals from other disciplines, which may limit the generalizability of the findings. Second, the study was conducted within a single time frame and in a single metropolitan area, thereby limiting the geographic diversity of the sample. The nursing schools and clinical practice settings involved in this study were characterized by frequent use of information and communication technologies, which may not be representative of all institutions. Because the sample was drawn from public universities in a large metropolitan area (Ankara) and consisted of students familiar with technology, participants’ attitudes toward and anxiety levels about artificial intelligence may have been influenced, which may limit the generalizability of the findings to the broader Turkish nursing student population. Therefore, further studies with larger, more diverse samples, including students from private universities, different geographic regions, and nurses working in clinical settings, are recommended to further support the validity and reliability evidence of the SAIP-TR. Another limitation concerns the relatively small subgroup used for the test–retest reliability analysis. Although the findings indicated temporal stability over a two-week interval, the test–retest sample consisted of only 32 participants. Future research, including larger samples, is recommended to further support evidence regarding the scale’s temporal stability. Another limitation is that the CFA model required post hoc modifications (item deletion and correlated residuals). While such steps are common in scale validation, they may reflect sample-specific optimization and can reduce structural stability. Replication in independent samples, preferably with pre-specified models and measurement invariance testing, is needed before treating this structure as stable.

The removal of two items from the original scale may also have affected the conceptual coverage of the Turkish version. Specifically, the excluded items addressed professional role displacement by AI and ethical/legal accountability for AI-related errors, which represent theoretically important aspects of AI perception. Although the retained two-factor structure suggests that the core domains of the original scale were preserved, the SAIP-TR may provide a somewhat narrower representation of the construct than the original instrument. Therefore, the Turkish version should be interpreted as an adapted 8-item form of the SAIP, and direct comparisons with findings obtained using the original 10-item scale should be made with caution. Finally, the explained variance is adequate but modest, indicating that additional dimensions of AI perception may remain unmeasured and that the results should be interpreted with caution regarding the scale’s overall construct representation.

Conclusion

In conclusion, the Turkish adaptation of the SAIP Scale (SAIP-TR) demonstrated adequate psychometric properties, with a two-factor structure that remains conceptually aligned with the original instrument. However, the interpretation of these findings should account for the modifications made during the adaptation process—specifically the exclusion of two items—and the borderline model fit observed in the confirmatory analysis. These factors, along with the use of post hoc modifications, suggest that the current structural solution may be sensitive to sample-specific characteristics. Therefore, while the 8-item SAIP-TR is a valuable tool for assessing core AI perceptions among nursing students in Türkiye, further validation across diverse, larger samples is essential. Such research is needed to provide stronger evidence regarding the scale’s structural stability, generalizability, and its conceptual equivalence to the original framework.

Supplementary Information

Below is the link to the electronic supplementary material.

Supplementary Material 1 ^{(24.7KB, docx)}

Abbreviations

AGFI

Adjusted goodness-of-fit index

AI

Artificial Intelligence

AMOS

Analysis of moment structures

CFA

Confirmatory factor analysis

CFI

Comparative fit index

CVI

Content validity index

EFA

Exploratory factor analysis

GFI

Goodness-of-fit index

ICC

Intraclass correlation coefficient

KMO

Kaiser-meyer-olkin

NFI

Normed fit index

SAIP

Shinners artificial intelligence perception

SPSS

Statistical package for the social sciences

SRMR

Standardised root mean square residual

RMSEA

Root mean square of error approximation

Author contributions

MÖK: Conceptualization, Methodology, Formal Analysis, Investigation, Resources, Data Curation, Writing-Original Draft.EK: Conceptualization, Methodology, Formal Analysis, Investigation, Resources, Data Curation, Writing-Original Draft.FK: Conceptualization, Methodology, Resources, Writing- Review & Editing, Supervision. All authors approved the final manuscript for submission.

Funding

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Data availability

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

Ethical approval was obtained from the Hacettepe University Health Sciences Ethics Committee, Ankara, Turkey (Decision No: 2025/03–05; Study ID: SBA 24/1128). The original developer granted permission to use and adapt the SAIP Scale via email. Institutional permissions were also obtained. The study was conducted in accordance with the principles of the Declaration of Helsinki, and all participants provided written informed consent.

Consent for publication

Not Applicable.

Competing interests

The authors declare no competing interests.