A context-sensitive scale for assessing GenAI-contextualized acceptance in conservation-related architectural education: development and initial validation

Abstract

Introduction:

Generative artificial intelligence (GenAI) is increasingly shaping higher education in professional fields where AI-supported outputs must be evaluated against disciplinary, ethical, and evidentiary criteria. Conservation-related architectural education is one such context because technically convincing outputs still require assessment against authenticity, material knowledge, intervention ethics, and heritage standards. This study developed and initially validated the Scale for Generative AI Technology Acceptance in Conservation Education (SGATE-CE) to assess GenAI-contextualized acceptance among undergraduate architecture students receiving conservation- and restoration-related coursework.

Methods:

After expert review, a 35-item pool was reduced to 27 items and examined using principal axis factoring with direct oblimin rotation in an exploratory sample (n = 157). Following pre-confirmatory measurement-model finalization, a 17-item structure comprising Perceived Usefulness, Behavioral Intention, Applied Educational Engagement, and Perceived Conservation-Specific Utility was tested through confirmatory factor analysis in a separate validation sample (n = 201). Reliability, convergent validity, discriminant validity, and theoretically specified structural associations were also examined.

Results:

The unmodified four-factor model showed acceptable fit: CMIN/df = 2.044, GFI = .880, CFI = .942, TLI = .930, RMSEA = .072, SRMR = .058, and RMR = .037. Alpha ranged from .797 to .896, omega from .804 to .897, composite reliability from .843 to .897, and AVE from .574 to .636. Structural modeling showed positive associations overall, although the PU-to-BI association was weaker and less stable under bootstrap sensitivity analysis.

Discussion:

The findings provide initial psychometric support for the four perceived-acceptance dimensions in conservation-related architectural education. Because some retained items may evoke broader AI or digital technologies, SGATE-CE measures GenAI-contextualized acceptance rather than exclusive GenAI acceptance. It does not measure subject knowledge, ethical judgment, critical evaluation, responsible AI use, or disciplinary competence. The Turkish version was tested, whereas the English wording is an informational translation. Further validation in dedicated conservation and heritage programs is required.

1 Introduction

Generative artificial intelligence (GenAI) is increasingly reshaping higher education by changing how students search for information, generate alternatives, produce representations, and evaluate knowledge claims. Unlike earlier educational technologies, GenAI tools can produce fluent text, plausible visual outputs, design alternatives, summaries, and interpretive suggestions. This creates new pedagogical opportunities, but it also introduces important risks because AI-generated outputs may appear coherent, efficient, or visually convincing without necessarily being accurate, ethically defensible, or discipline-specific. In higher education, therefore, the central question is no longer only whether students use GenAI, but how they evaluate and integrate AI-supported outputs within the epistemic, ethical, and professional standards of their fields.

Recent studies have begun to examine students' perceptions of and responses to GenAI in higher education. Chan and Hu (2023), for example, showed that students recognize the potential benefits of GenAI for learning support, writing, brainstorming, research, and analysis, while also expressing concerns about accuracy, privacy, ethics, over-reliance, and the implications of AI use for learning and professional development. Chan and Zhou (2023) developed an Expectancy-Value Theory-based instrument to examine students’ GenAI perceptions through constructs such as knowledge of GenAI, perceived value, perceived cost, and intention to use. These studies indicate that students’ responses to GenAI extend beyond technical usability and are associated with perceived educational value, perceived risks, and informed judgments about AI-supported outputs.

This issue is particularly important in professional and practice-oriented education. In such fields, students are expected not only to adopt digital tools but also to justify decisions, compare alternatives, interpret evidence, and evaluate outputs against disciplinary criteria. Technology acceptance in these contexts cannot be reduced to general usefulness or ease of use alone. It is also associated with whether students perceive AI-supported outputs as meaningful, valid, and applicable within the specific knowledge structures of their discipline. For this reason, professional higher education requires context-sensitive measurement approaches that can capture both general technology-acceptance constructs and domain-specific perceptions of utility.

Conservation-restoration education provides a high-stakes context for examining this problem. Students in this field are required to evaluate decisions in relation to authenticity, historical evidence, intervention ethics, material knowledge, interdisciplinary collaboration, and international conservation principles (Ahunbay, 2009; ICOMOS, 1964; UNESCO/ICCROM/ICOMOS, 1994). In this context, AI-generated outputs may be useful for documentation, representation, analysis, and ideation; however, they may also produce historically misleading, visually overconfident, or ethically problematic results. Previous work on AI-generated architectural heritage representations has shown that visually persuasive outputs can contain hallucinations, stylistic inconsistencies, and cultural inaccuracies (Sukkar et al., 2024). Therefore, the educational value of GenAI in conservation is not determined only by efficiency or ease of use, but also by students’ capacity to interpret AI-supported outputs through disciplinary criteria.

The Technology Acceptance Model (TAM) remains one of the most influential frameworks for explaining technology adoption, particularly through perceived usefulness and perceived ease of use (Davis, 1989). Later extensions, including TAM2 and TAM3, incorporated additional social and cognitive determinants of acceptance (Venkatesh and Bala, 2008; Venkatesh and Davis, 2000). Broader models such as the Unified Theory of Acceptance and Use of Technology (UTAUT) and UTAUT2 further expanded technology acceptance research by considering constructs such as performance expectancy, effort expectancy, social influence, facilitating conditions, hedonic motivation, price value, and habit (Venkatesh et al., 2003, 2012). These models have contributed substantially to explaining technology adoption in educational contexts (Granić and Marangunić, 2019; Scherer et al., 2019). However, general acceptance models do not fully address the evaluative demands of professional education, where students must judge whether AI-supported outputs are compatible with disciplinary standards, ethical obligations, and evidence-based reasoning.

In parallel, AI literacy and GenAI literacy research has emphasized competencies such as understanding AI, using AI tools, evaluating AI outputs, recognizing limitations, and engaging with AI systems ethically and responsibly (Laupichler et al., 2022; Long and Magerko, 2020; Ng et al., 2021). More recent work has also begun to distinguish GenAI literacy from broader AI literacy by focusing on learners’ ability to understand, use, evaluate, and critically engage with generative systems (Jin et al., 2025). These literacy-oriented frameworks are highly relevant to responsible AI education, but they differ from technology acceptance instruments. They focus primarily on competencies, knowledge, or performance-based literacy rather than on students’ perceived acceptance of GenAI-supported learning within a specific professional field.

Taken together, existing TAM/UTAUT-based models, GenAI perception instruments, and AI literacy frameworks provide valuable foundations, but they do not directly capture the conservation-specific context in which AI-supported outputs are appraised against authenticity, material appropriateness, interdisciplinary restoration reasoning, and international conservation standards. This gap is important because a student may perceive GenAI as generally useful or easy to use while remaining uncertain about its conservation-specific utility. Conversely, students may report stronger acceptance when they perceive AI-supported tools as useful for tasks such as existing-condition analysis, digital modeling, material selection, interpretation of heritage standards, or conservation-oriented decision-making. These perceptions should not be interpreted as evidence of subject knowledge, ethical judgment, critical evaluation, or responsible AI use.

In the present study, acceptance is operationalized as a multidimensional self-report appraisal comprising perceived usefulness, behavioral intention, applied educational engagement, and perceived conservation-specific utility. The intended referent is GenAI because the survey introduction and every scale section instructed participants to answer by considering generative AI tools. Nevertheless, several retained items use the broader expressions “AI technologies” or “technological tools.” The resulting scores are therefore interpreted as GenAI-contextualized acceptance rather than as a technology-neutral measure or an instrument that empirically separates GenAI from CAD, BIM, GIS, photogrammetry, conventional AI, or other digital technologies. This boundary is central to the construct-validity interpretation of SGATE-CE.

To address this gap, the present study develops and initially validates the Scale for Generative AI Technology Acceptance in Conservation Education (SGATE-CE). The conceptual structure includes four perceived-acceptance constructs: Perceived Usefulness (PU), Behavioral Intention (BI), Applied Educational Engagement (AEE), and Perceived Conservation-Specific Utility (PCSU). PU refers to students’ perceptions of the usefulness of GenAI-contextualized tools for conservation-related learning and practice. BI refers to students’ intention to use, recommend, and further learn about AI tools in conservation-related architectural education. AEE refers to students’ self-reported applied educational engagement with these tools. PCSU refers to students’ perceptions of conservation-specific utility, including perceived usefulness for material selection, interdisciplinary restoration work, heritage-identity-related decision support, and alignment with international conservation standards. These dimensions measure perceived usefulness, intention, engagement, and conservation-specific utility; they do not measure subject knowledge, ethical judgment, critical evaluation, responsible AI use, or disciplinary competence.

PCSU is specified as a first-order reflective latent construct rather than a formative index. The proposed direction of measurement is from an underlying general perception of conservation-specific utility to responses across several task contexts. Heritage identity, material selection, interdisciplinary collaboration, and standards-aligned decision support are treated as contextually varied manifestations of a common appraisal of whether GenAI-supported tools are useful for conservation-related architectural reasoning, not as independent components whose weighted combination creates PCSU (Jarvis et al., 2003). Its conceptual boundary is perceived utility: PCSU does not claim to exhaust conservation practice or measure objective knowledge, competence, responsible use, or the quality of decisions.

The dimension originally coded as Applied Use (AU) is reported as Applied Educational Engagement (AEE), a label that more accurately reflects the retained self-report items. The original AU prefix is retained only when referring to dataset variable names or earlier item-pool coding.

The objective of this study is to develop and initially validate SGATE-CE as a context-sensitive instrument for assessing GenAI-contextualized acceptance in conservation-related architectural education. In line with TAM-based acceptance research, the study examines the theoretically specified associations among PCSU, PU, BI, and AEE while recognizing that the cross-sectional design does not permit causal inference. The following relational hypotheses were formulated:

H1: Perceived Usefulness (PU) will be positively associated with Behavioral Intention (BI).

H2: Behavioral Intention (BI) will be positively associated with Applied Educational Engagement (AEE).

H3: Perceived Conservation-Specific Utility (PCSU) will be positively associated with Perceived Usefulness (PU).

H4: Perceived Conservation-Specific Utility (PCSU) will be positively associated with Behavioral Intention (BI).

These hypothesized relationships are summarized in Figure 1.

Figure 1

In addition to these relational hypotheses, the study pursued two psychometric objectives. First, it aimed to explore the factor structure of the draft scale and to test the resulting 17-item structure in a separate validation sample after pre-confirmatory measurement-model finalization. Because the final validation form did not include the intermediate SMK2 item, this test is presented as initial validation evidence rather than as a fully independent confirmation of every item-refinement decision. Second, the study aimed to examine the internal consistency, composite reliability, convergent validity, and discriminant validity of the retained dimensions. Accordingly, the study proceeded through item generation, expert review, exploratory factor analysis, pre-confirmatory measurement-model finalization, confirmatory factor analysis, reliability and validity assessment, and structural model testing.

2 Materials and methods

2.1 Research design

This study was designed as a quantitative, non-experimental, cross-sectional, instrumental psychometric study aimed at developing and initially validating a context-sensitive scale for assessing GenAI-contextualized acceptance among undergraduate architecture students receiving conservation- and restoration-related coursework. The study followed an exploratory-confirmatory validation logic with an important qualification regarding item finalization. First, an initial item pool was generated based on technology acceptance literature, GenAI-related higher education research, and conservation-related architectural education considerations. Second, the item pool was reviewed by experts to assess content relevance, conceptual fit, clarity, measurability, and suitability for the target group. Third, the revised draft form was administered to an exploratory sample and examined using exploratory factor analysis (EFA). Fourth, the final 17-item form specified after pre-confirmatory measurement-model finalization was administered to a separate validation sample and tested using confirmatory factor analysis (CFA). Finally, reliability, convergent validity, discriminant validity, and theoretically specified structural associations among the retained constructs were examined. The final structure is therefore presented as initial validation evidence requiring external replication rather than as a fully independent confirmation of all item-refinement decisions.

The research design was cross-sectional because data were collected at a single time point from each participant. It was non-experimental because no intervention, treatment condition, or random assignment was used in the scale validation process. The study also included a correlational component because the relationships among Perceived Conservation-Specific Utility (PCSU), Perceived Usefulness (PU), Behavioral Intention (BI), and Applied Educational Engagement (AEE) were examined. These relationships were interpreted as theoretically specified directional associations rather than causal effects, given the cross-sectional nature of the data.

2.2 Population and sample

The target population for this initial validation consisted of undergraduate architecture students enrolled in architecture departments with conservation- and restoration-related curricular content in Türkiye. The accessible population consisted of architecture students from three departments during the spring semester of the 2024–2025 academic year. This population should not be equated with students enrolled in dedicated conservation-restoration, heritage conservation, architectural restoration, or restoration-focused degree programs; those specialist groups were outside the present validation sample. Participants were selected because they represented an accessible group receiving conservation-related architectural education in which AI-supported tools may be used for documentation, analysis, representation, interpretation, and conservation-oriented decision-making.

Data were collected after ethics approval through a structured online questionnaire using purposive sampling. Both samples were drawn from three architecture departments located in different cities in Türkiye, which reduced the risk of relying on a single institutional source but did not constitute probability-based sampling across all conservation, restoration, architectural restoration, or cultural heritage programs. The questionnaire was distributed through instructor-mediated and institution-related online communication channels. Because the invitation was circulated through these channels rather than through a centrally controlled mailing list, the exact number of students who received the invitation and the response rate could not be calculated. For this reason, the study is reported as a purposive online survey based on complete voluntary responses rather than as a probability-based sample with a known response rate. The questionnaire was administered in Turkish. The introductory page identified the survey as part of a doctoral research project on generative AI-supported learning in conservation education, and each scale section instructed participants to answer the items by considering generative AI tools. Participation was voluntary, no personally identifying information was collected, and responses were used only for scientific purposes. The estimated completion time was 10–15 min. The accessible recruitment sources consisted of architecture students from three departments: the Department of Architecture, Faculty of Architecture, Erciyes University; the Department of Architecture, Faculty of Architecture, Abdullah Gül University; and the Department of Architecture, Faculty of Architecture, Bozok University.

The study was approved by the Social and Human Sciences Ethics Committee of Erciyes University (Application No. 231; decision date: 27 May 2025). All participants were informed about the purpose of the study, voluntary participation, anonymity, and the scientific use of the data before completing the questionnaire.

A total of 358 participants were included across two separate samples in a two-phase validation design. The exploratory sample consisted of 157 students and was used for item reduction and exploratory factor analysis (EFA). The validation sample consisted of 201 students and was used for confirmatory factor analysis (CFA), reliability analysis, convergent validity, discriminant validity, and structural model testing. The two samples were not produced by randomly splitting a single dataset; rather, they were treated as separate samples corresponding to the exploratory and confirmatory phases of the scale-development process. This procedure was used to reduce capitalization on chance and to provide a more rigorous initial test of the factor structure.

In the exploratory sample, 70.7% of participants were female (n = 111) and 29.3% were male (n = 46). The mean age was 21.06 years (SD = 2.89; range = 18–47). In the validation sample, 71.1% of participants were female (n = 143) and 28.9% were male (n = 58). The mean age was 21.39 years (SD = 2.71; range = 18–47). The demographic characteristics of the exploratory and validation samples are summarized in Table 1.

Although the participating institutional sources were identified at the recruitment level, the analytical datasets did not retain case-level institutional counts, program-level subgroup identifiers, academic year or level, prior GenAI experience, frequency of GenAI use, specific GenAI tools used, or previous GenAI training in a form suitable for subgroup analysis. Consequently, institutional and program-level comparisons could not be conducted, and the samples are described at the available aggregate level. This limitation is acknowledged in the limitations section.

Participants were included if they were undergraduate architecture students enrolled in one of the accessible architecture departments with conservation- and restoration-related curricular content; were able to respond to the Turkish questionnaire; and voluntarily agreed to participate. Responses were excluded if they were incomplete, contained missing data that prevented scale scoring, showed clear signs of careless responding, or failed to meet the data-screening criteria described below.

The adequacy of the sample sizes was evaluated in relation to common recommendations for exploratory and confirmatory factor analysis. For the EFA phase, the draft form contained 27 items after expert review, yielding approximately 5.8 participants per item in the exploratory sample. This ratio was considered acceptable for preliminary scale development when interpreted together with the adequacy of communalities, factor loadings, and the Kaiser-Meyer-Olkin measure. For the CFA phase, the validation sample of 201 participants was considered sufficient for testing the 17-item four-factor measurement model, as it exceeded the commonly cited minimum threshold of 200 cases for confirmatory factor analysis and provided more than 10 participants per retained item. These sample sizes were therefore considered appropriate for an initial psychometric validation study.

2.3 Instrument development and content validation

The initial item pool was developed to represent technology acceptance constructs and conservation-specific considerations relevant to GenAI-supported conservation-related architectural education. In the first stage, 35 items were generated under five theoretically informed dimensions: Perceived Usefulness (PU), Perceived Ease of Use (PEOU), Behavioral Intention (BI), Applied Use (AU; subsequently labeled Applied Educational Engagement, AEE), and a conservation-specific dimension initially labeled Subject Matter Knowledge and subsequently labeled Perceived Conservation-Specific Utility (PCSU).

The item-writing process was informed by TAM-based technology acceptance literature and by conservation education concepts related to authenticity, material knowledge, intervention ethics, international conservation principles, and field-based decision-making. The conservation-specific items were initially developed to represent perceptions of utility across conservation-restoration task contexts, including material selection, interdisciplinary restoration processes, heritage-identity considerations, and alignment with conservation standards. The final retained construct is interpreted only as perceived conservation-specific utility and not as a measure of subject knowledge, ethical judgment, critical evaluation, or responsible AI use.

The 35-item pool was evaluated through a two-round expert-based qualitative content review. The expert panel consisted of seven academics selected purposively to represent both disciplinary and methodological expertise: two professors in conservation-restoration, one assistant professor in architectural history, two assistant professors in educational sciences, and two assistant professors in measurement and evaluation. Their approximate professional experience ranged from 10 to 30 years. In the first review round, experts provided item-level written comments on content relevance, conceptual fit with the intended dimension, clarity of wording, measurability, suitability for undergraduate students, and conceptual redundancy. These comments were organized into a decision matrix and used to identify items that should be retained, revised, removed before EFA, or retained for empirical testing. In the second review round, the revised 27-item draft form was re-examined for clarity, redundancy, target-group suitability, and consistency with the intended construct definitions. Because the expert form collected qualitative written comments rather than numerical relevance or essentiality ratings, formal quantitative indices such as CVI or CVR (Lawshe, 1975) were not calculated. Therefore, this stage is interpreted as preliminary expert-based qualitative evidence of content relevance and item suitability, not as comprehensive quantitative content validation. Detailed expert-panel characteristics and the qualitative review procedure are reported in Supplementary Table S17.

Following the expert review, the initial 35-item pool was reduced to a 27-item draft form. Specifically, two items were removed from Perceived Usefulness, two from Perceived Ease of Use, two from Behavioral Intention, and two from Applied Educational Engagement, while the seven items in the original conservation-specific dimension were retained at this stage. The removed items were excluded because they were judged to create conceptual overlap or to measure constructs that were not central to the intended dimension. For example, some removed items were considered to reflect motivation, aesthetic presentation, technical self-efficacy, career orientation, payment willingness, fixed frequency of use, or direct project use rather than the intended acceptance construct. The full item pool, expert-review decisions, and item-level removal rationales are provided in the Supplementary Material.

The 27-item draft form included five items for Perceived Usefulness, five items for Perceived Ease of Use, five items for Behavioral Intention, five items for Applied Use (reported as Applied Educational Engagement), and seven items for the conservation-specific dimension. The questionnaire was administered in Turkish. The survey introduction and each scale section included the instruction “Üretken yapay zekâ araçlarını dikkate alarak cevaplayınız” (“Please answer by considering generative AI tools”). This provided a common GenAI referent for items that sometimes used broader AI-technology or technological-tool wording. The item wording was not reformulated after data collection because replacing these terms retrospectively would create a mismatch between the instrument administered and the psychometric evidence reported.

The Turkish version was the administered and psychometrically tested form. The English item wording provided in the Supplementary Material was prepared after the Turkish form had been finalized in order to support transparency and international readability. The initial Turkish-to-English translation was prepared by the first author and reviewed with the second author for semantic consistency, conservation-related terminology, and disciplinary appropriateness. In addition, an architect colleague with advanced English proficiency reviewed the English wording for readability, clarity, and architectural terminology. Any wording issues or semantic ambiguities identified during this process were resolved through discussion between the authors. No formal back-translation, independent bilingual expert panel, or cross-cultural adaptation procedure was conducted. Accordingly, the English version should be regarded as an informational translation rather than as a formally adapted and independently validated English-language version of the scale.

Accordingly, the conservation-specific dimension was relabeled as Perceived Conservation-Specific Utility (PCSU) rather than Subject Matter Knowledge. This change was made because the retained items assess perceptions of conservation-specific utility rather than objective disciplinary knowledge. The items address perceived usefulness for material selection, interdisciplinary restoration reasoning, heritage-identity-related decision support, and alignment with international conservation standards. PCSU therefore reflects the empirical content of the retained items without implying subject knowledge, ethical judgment, critical evaluation, responsible AI use, or professional competence. The construct was specified reflectively: the retained items were intended to express a shared perception of conservation-specific utility across different task contexts rather than to form an exhaustive checklist of conservation capabilities.

No separate pilot test or cognitive interview was conducted with students before the main validation study. Instead, the initial comprehensibility and content suitability of the items were evaluated through the expert review process. This decision is acknowledged as a limitation because student-level cognitive interviewing could have provided additional evidence on how participants interpreted terms related to GenAI, authenticity, conservation standards, ethical responsibility, and AI-supported decision-making.

2.4 Data collection procedure and data screening

Data were collected after ethics approval through a structured online questionnaire administered in Turkish via Google Forms. The questionnaire was distributed online to undergraduate architecture students in the three participating architecture departments with conservation- and restoration-related curricular content during the spring semester of the 2024–2025 academic year. Data collection was conducted between 1 June and 15 June 2025. Participants first viewed an introductory information page explaining the purpose of the study, voluntary participation, anonymity, scientific use of the data, and estimated completion time. The questionnaire did not collect personally identifying information.

The scale sections instructed participants to answer the items by considering generative AI tools. This contextual instruction was included to provide a common GenAI referent even when some items used broader wording such as “AI technologies” or “technological tools.” It supports interpretation of the resulting scores as GenAI-contextualized acceptance, but it does not completely eliminate ambiguity because respondents could associate the broader expressions with non-generative systems such as CAD, BIM, GIS, photogrammetry, or conventional AI. Consequently, SGATE-CE should not be treated as a pure item-level discriminator of exclusive GenAI acceptance or as a technology-neutral measure of general digital-tool acceptance. Responses were recorded on a five-point Likert-type scale ranging from 1 = Strongly disagree to 5 = Strongly agree.

All scale items were set as required fields in Google Forms. Therefore, submitted questionnaires did not contain missing item-level data. Because no personally identifying information was collected, e-mail-, account-, or IP-based duplicate control was not applied in order to preserve participant anonymity. Completion-time screening could not be applied because response-duration data were not retained in the analytical files. Before analysis, the datasets were screened for completeness, straight-lining and long-string patterns, person-level response variability, Mahalanobis distance, and multicollinearity. These procedures were used to identify possible response-quality concerns for transparent reporting rather than to remove cases automatically.

The exploratory analysis was conducted on 157 complete responses, and no listwise exclusions were required in the EFA dataset. The validation analysis was conducted on a separate sample of 201 complete responses. The final analytical dataset therefore consisted of 358 participants across the two separate samples.

The data were also examined for assumptions relevant to exploratory and confirmatory factor analysis. Item distributions were inspected using means, standard deviations, skewness, and kurtosis. Sampling adequacy for EFA was evaluated using the Kaiser-Meyer-Olkin (KMO) measure, and the factorability of the correlation matrix was assessed using Bartlett's test of sphericity. In the validation sample, formal multivariate normality was examined using Mardia's multivariate skewness and kurtosis (Mardia, 1970). Because multivariate normality was not supported, CFA and SEM findings were interpreted cautiously together with available robustness and sensitivity evidence reported in the results and Supplementary Material.

2.5 Data analysis

Data were analyzed using IBM SPSS Statistics 28 and IBM SPSS AMOS 24. The analysis followed a sequential exploratory-confirmatory procedure. First, descriptive statistics were examined for all items, including means, standard deviations, skewness, and kurtosis. The suitability of the exploratory dataset for factor analysis was assessed using the Kaiser-Meyer-Olkin (KMO) measure of sampling adequacy and Bartlett's test of sphericity.

Exploratory factor analysis (EFA) was conducted on the first sample (n = 157) using principal axis factoring (PAF). PAF was selected because the purpose of the analysis was to identify the underlying latent factor structure of the draft instrument rather than to reduce observed variables into components. Because the dimensions of technology acceptance were theoretically expected to be correlated, direct oblimin rotation with Kaiser normalization was applied. Factor retention was evaluated using eigenvalues greater than 1, scree plot inspection, interpretability of the factor solution, and theoretical coherence with the construct definitions.

Parallel analysis (Horn, 1965) was additionally conducted on the 27-item exploratory dataset to provide an empirical factor-retention check beyond the eigenvalue > 1.00 rule and scree plot inspection. Both the mean random eigenvalue criterion and the more conservative 95th-percentile criterion were examined. Where the parallel analysis criteria diverged, factor retention was interpreted in relation to theoretical coherence, item content, the final rotated solution, and the subsequent CFA and alternative-model comparisons.

Item retention and elimination decisions were based on both statistical and conceptual criteria. As practical guidelines, items were expected to show a primary loading of approximately |.40| or higher, sufficient communality, and clear conceptual alignment with the intended construct. No fixed absolute communality threshold was preregistered; however, extraction communalities below approximately .30 were treated as low and were considered for removal, particularly when accompanied by weak primary loading, problematic cross-loading, or conceptual misfit. Cross-loadings of |.30| or higher were inspected, and a primary-secondary loading difference of approximately |.20| or greater was preferred when evaluating whether an item clearly belonged to one factor. Factors were expected to retain at least three interpretable items. These criteria were applied iteratively rather than as preregistered strict decision rules, and borderline decisions were evaluated by considering both empirical behavior and conceptual fit. The Perceived Ease of Use (PEOU) dimension was removed during the EFA process because its items did not form a clearly interpretable and psychometrically stable factor. The retained EFA records documented an intermediate 18-item, four-factor solution that included SMK2 (“Digital tools reduce human errors in conservation processes”). Available records indicated that SMK2 showed acceptable EFA performance and was not removed because it failed the stated loading, communality, or cross-loading guidelines. After the exploratory solution had been obtained, the conservation-specific dimension was conceptually re-examined while the final validation form was being specified. SMK2 was removed at this pre-confirmatory, theory-driven stage because its generic digital-tool and human-error-reduction content could apply equally to CAD, BIM, GIS, photogrammetry, or other non-generative technologies and did not represent the bounded construct of GenAI-contextualized conservation-specific utility. The transition was therefore: 27-item draft form, iterative EFA refinement, intermediate 18-item four-factor solution, theory-driven exclusion of SMK2, and final 17-item validation form. Because the 201-participant validation sample received only the 17-item form, an 18-item CFA including SMK2 could not be estimated in that sample. The final reported 17-item solution had primary factor loadings ranging from |.461| to |.914|.

Confirmatory factor analysis (CFA) was conducted on the second sample (n = 201) using AMOS 24 to test the final 17-item four-factor measurement model administered to the validation sample. Maximum likelihood estimation was used. Although the items were measured on a five-point Likert-type scale, they were treated as ordered-continuous indicators and evaluated with multiple fit indices. Formal multivariate normality was assessed using Mardia's multivariate skewness and kurtosis. Because the validation data did not satisfy multivariate normality, model results were interpreted cautiously. The unmodified four-factor measurement model was additionally examined using a Bollen-Stine bootstrap sensitivity check, and the measurement structure was further evaluated through alternative CFA model comparisons. The structural paths were complemented with estimator/objective sensitivity checks and bootstrap-based robustness evidence where available.

To further evaluate the distinctiveness of the retained four-factor measurement structure, additional CFA model comparisons were conducted in the validation sample. The theoretically specified four-factor model was compared with more parsimonious alternative models, including a one-factor model, three-factor models combining conceptually close dimensions, and a two-factor model separating general acceptance from perceived conservation-specific utility.

Reliability was assessed using Cronbach's alpha and McDonald's omega total for each retained factor. Cronbach's alpha values of .70 or higher were interpreted as acceptable for internal consistency. McDonald's omega was additionally reported because it provides a model-based reliability estimate that does not require the strict assumption of tau-equivalence (Hayes and Coutts, 2020). Composite reliability (CR) and average variance extracted (AVE) were calculated to evaluate construct reliability and convergent validity (Fornell and Larcker, 1981). CR values of .70 or higher and AVE values of .50 or higher were interpreted as evidence of acceptable construct reliability and convergent validity. Omega total, CR, AVE, Fornell-Larcker comparisons, and HTMT values were computed using standard formula-based procedures from the retained-item correlation matrices and standardized loading information; HTMT was based on Pearson rather than polychoric correlations.

Discriminant validity was examined using inter-factor correlations, the Fornell-Larcker criterion (Fornell and Larcker, 1981), and the heterotrait-monotrait ratio of correlations (HTMT; Henseler et al., 2015). HTMT values were calculated from Pearson correlations, and bootstrap confidence intervals were estimated as an additional sensitivity check. According to the Fornell-Larcker criterion, the square root of the AVE for each construct should exceed its correlations with other constructs. HTMT values below .85 were interpreted as conservative evidence of discriminant validity, while values below .90 were considered acceptable for conceptually related constructs.

Finally, structural models were specified to examine the theoretically proposed directional associations among the retained constructs. Because the Introduction specified H1–H4, the four-path model including PCSU → PU, PU → BI, PCSU → BI, and BI → AEE was treated as the primary hypothesized structural model. A more parsimonious sequential model including PCSU → PU, PU → BI, and BI → AEE was additionally examined as a comparison model, not as the primary hypothesized model. Structural model fit was evaluated using the same fit indices adopted for the CFA model. Standardized path coefficients, unstandardized estimates, standard errors, critical ratios, p-values, approximate 95% confidence intervals for unstandardized estimates, and explained variance values (R²) were examined. Given the cross-sectional and non-experimental design, these paths were interpreted as theoretically specified directional associations rather than causal effects. Additional sensitivity checks were used to examine whether the direction of the structural paths remained stable across alternative estimation objectives and bootstrap resampling.

2.6 Ethics statement

All procedures involving human participants were reviewed and approved by the Social and Human Sciences Ethics Committee of Erciyes University (Application No. 231; decision date: 27 May 2025). Data were collected after ethics approval between 1 June and 15 June 2025. Participation was voluntary, no personally identifying information was collected, and all participants provided informed consent electronically before accessing the questionnaire.

3 Results

3.1 Data screening and preliminary analysis

Before conducting the factor analyses, the datasets were examined for completeness, response consistency, and suitability for psychometric analysis. Because all scale items were set as required fields in Google Forms, submitted questionnaires did not contain missing item-level data. The exploratory dataset consisted of 157 complete responses, and no listwise exclusions were required. The validation dataset consisted of 201 complete responses. Thus, the final analytical dataset included 358 participants across the two separate samples.

Response patterns were inspected to identify possible signs of careless responding. Long-string and person-level variability checks identified two zero-variance response profiles in the EFA dataset and 16 zero-variance response profiles in the validation dataset. Because the retained items were positively worded acceptance items, uniform high agreement could reflect substantive endorsement rather than unequivocal careless responding. Therefore, these cases were flagged for transparency rather than removed automatically. Mahalanobis-distance screening also identified seven candidate multivariate outliers in the EFA dataset and 10 in the validation dataset at p < .001. No listwise exclusions were applied solely on the basis of these screening indicators. Additional screening details retained in the analytical records are reported in Supplementary Table S14.

Item-level descriptive statistics were examined separately for the exploratory and validation samples. In the exploratory sample, all 27 draft items used the full response range from 1 to 5. Skewness values ranged from −1.360 to 0.164, and kurtosis values ranged from −1.221 to 3.272. In the validation sample, all 17 retained items also used the full response range from 1 to 5. Skewness values ranged from −1.415 to −0.489, and kurtosis values ranged from 0.163 to 3.529. These values indicated negatively skewed but not extremely non-normal item distributions, supporting the use of the planned factor-analytic procedures with cautious interpretation through multiple fit and validity indices.

Formal multivariate normality was examined in the validation sample using Mardia's coefficients (Mardia, 1970). The results did not support multivariate normality: Mardia's multivariate skewness was 74.906, χ²(969) = 2509.360, p < .001, and Mardia's multivariate kurtosis was 428.939, z = 29.547, p < .001. Therefore, the CFA and SEM results were interpreted with caution. The unmodified four-factor CFA measurement model was additionally examined using Bollen-Stine bootstrap sensitivity analysis. The four-factor measurement structure was also evaluated through alternative CFA model comparisons, and the structural paths were supplemented by estimator/objective and bootstrap sensitivity checks.

The suitability of the exploratory sample for factor analysis was assessed using the Kaiser-Meyer-Olkin (KMO) measure of sampling adequacy and Bartlett's test of sphericity. The KMO value was .887, indicating very good sampling adequacy according to Kaiser's criterion (Kaiser, 1974). Bartlett's test of sphericity was statistically significant, χ²(351) = 2088.299, p < .001, confirming that the correlation matrix was factorable. These results supported proceeding with exploratory factor analysis on the 27-item draft form. The KMO measure and Bartlett's test results are presented in Table 2.

3.2 Exploratory factor analysis

Exploratory factor analysis was conducted on the 27-item draft form using principal axis factoring with direct oblimin rotation and Kaiser normalization in the exploratory sample (n = 157). The analysis followed an iterative item-elimination procedure based on factor loadings, cross-loadings, communalities, and conceptual fit with the intended construct definitions.

In the initial EFA solution, the Perceived Ease of Use (PEOU) items did not form a clearly interpretable and psychometrically stable factor. Several PEOU items showed weak clustering, low factor loadings, or cross-loading patterns. Therefore, the PEOU dimension was removed from the emerging measurement structure. This result indicated that ease of use did not operate as a distinct empirical dimension in this conservation-related architectural education sample. The removal of PEOU should not be interpreted as evidence that ease of use is irrelevant to GenAI-supported learning; rather, it suggests that the PEOU items developed for this draft instrument did not separate adequately from the other acceptance-related dimensions in the present dataset.

After the removal of PEOU, additional items were removed iteratively when they showed weak primary loadings, problematic cross-loadings, low communalities, or conceptual inconsistency with the emerging factor structure. This data-driven EFA refinement produced an intermediate 18-item, four-factor solution that retained SMK2. Available EFA records indicated acceptable empirical performance for SMK2 (h² = .429, MSA = .810, primary loading λ = .658, with no problematic secondary loading); its removal was therefore not based on a failed loading, communality, or cross-loading criterion. During the subsequent pre-confirmatory conceptual review, SMK2 was excluded because the statement “Digital tools reduce human errors in conservation processes” expressed generic digital-tool efficiency and could refer to CAD, BIM, GIS, photogrammetry, or other non-generative systems rather than GenAI-contextualized conservation-specific utility. This was a theory-driven refinement undertaken after EFA and before administration of the validation form. The transition was thus from the 27-item draft form to an intermediate 18-item EFA structure and then to the final 17-item validation form. Because SMK2 was not administered to the 201-participant validation sample, an 18-item CFA including this item could not be estimated. The final retained factors were Perceived Usefulness (PU), Behavioral Intention (BI), Applied Educational Engagement (AEE), and Perceived Conservation-Specific Utility (PCSU). The fourth factor was relabeled as PCSU because the retained items captured perceived conservation-specific utility rather than objective subject matter knowledge.

The final four-factor solution explained 67.37% of the total variance based on the eigenvalue > 1.00 criterion and scree plot inspection. Factor loadings in the final reported PAF solution ranged from |.461| to |.914|. Because direct oblimin rotation was used, negative loadings in the Applied Educational Engagement factor reflected factor orientation rather than negative substantive relationships; therefore, interpretation was based on the magnitude of the absolute loadings. Overall, the retained structure showed a theoretically interpretable pattern in which the items clustered around perceived usefulness, behavioral intention, applied educational engagement, and perceived conservation-specific utility. The eigenvalues and percentages of explained variance are presented in Table 3, and the final pattern matrix is presented in Table 4.

Parallel analysis provided partial empirical support for the retained structure. Under the mean random eigenvalue criterion, the first four observed eigenvalues exceeded the corresponding random mean eigenvalues, supporting a four-factor solution. Under the more conservative 95th-percentile criterion, only the first two observed eigenvalues exceeded the random 95th-percentile eigenvalues. The four-factor solution was retained because it was theoretically interpretable, corresponded to the intended acceptance and conservation-specific constructs, produced a coherent final rotated solution, and was subsequently supported by CFA and alternative measurement-model comparisons. The full parallel-analysis results are reported in Supplementary Table S11.

To improve transparency, the Supplementary Material provides a comprehensive removed-item decision table reporting the removal stage, reason, extraction communality, primary pattern coefficient, largest secondary pattern coefficient, and primary-secondary loading difference for each empirically evaluated removed item. It also reports the complete, unthresholded pattern and structure matrices for the retained 17-item solution.

3.3 Confirmatory factor analysis

The retained 17-item, four-factor structure was tested through confirmatory factor analysis using the separate validation sample (n = 201), which completed the final 17-item validation form. The tested measurement model included four correlated latent factors: Perceived Usefulness (PU), Behavioral Intention (BI), Applied Educational Engagement (AEE), and Perceived Conservation-Specific Utility (PCSU). Because SMK2 was not administered in this validation form, the final CFA should be interpreted as an initial test of the pre-specified 17-item structure in a separate sample rather than as a fully independent confirmation of all item-refinement decisions.

The CFA was conducted in AMOS 24 using maximum likelihood estimation. Modification indices were examined; however, the suggested model relaxations largely involved cross-loadings or theoretically weak error-covariance proposals. Because the unmodified four-factor model already showed acceptable fit, no post hoc error covariances or cross-loadings were added. The results are therefore reported for the theoretically specified unmodified measurement model. The final 17-item form was not modified after administration to the validation sample; however, external replication is still required to confirm the stability of the retained structure in new samples.

The unmodified four-factor model demonstrated acceptable-to-good fit: CMIN/df = 2.044, GFI = .880, CFI = .942, TLI = .930, RMSEA = .072, 90% CI [.059, .086], SRMR = .058, and RMR = .037. These values indicate that the proposed four-factor model provided an adequate representation of the validation data. The CFA model fit indices are summarized in Table 5. Although PCLOSE = .004 did not support a close-fit interpretation under the strict RMSEA <.05 criterion, the combined evidence from CMIN/df, CFI, TLI, RMSEA, SRMR, GFI, and RMR supported retention of the unmodified measurement model.

Because multivariate normality was not supported, the same unmodified CFA model was additionally examined using Bollen-Stine bootstrap sensitivity analysis with 1,000 bootstrap samples. The Bollen-Stine bootstrap result was significant (p = .010), indicating that exact model fit was not supported; therefore, the CFA evidence was interpreted as acceptable approximate fit rather than close or exact fit. The standardized loading pattern remained adequate across the retained indicators (PU = .652–.879, BI = .724–.843, AEE = .781–.827, PCSU = .705–.852), as reported in Supplementary Table S13A.

Additional measurement-model comparisons were conducted to examine whether the retained four-factor structure provided a better representation of the validation data than more parsimonious alternatives. The four-factor model showed better fit than the one-factor model, the two-factor model, and the three-factor models in which conceptually close constructs were combined. In particular, combining Behavioral Intention with Applied Educational Engagement, Applied Educational Engagement with Perceived Conservation-Specific Utility, or Behavioral Intention with Perceived Conservation-Specific Utility resulted in weaker fit than the retained four-factor model. These comparisons supported the distinctiveness of the four retained SGATE-CE dimensions despite their theoretically expected correlations. The alternative model fit indices are reported in Supplementary Table S10.

The four-factor model also showed the lowest chi-square-based AIC among the compared models (AIC = 310.953). The one-factor model, the two-factor model, and the three-factor alternatives all showed larger chi-square values, weaker incremental fit, poorer RMSEA values, and substantially higher AIC values. These results supported retaining the four correlated dimensions rather than collapsing them into broader factors. Supplementary Table S10 reports χ², df, CFI, TLI, RMSEA, SRMR, AIC, Δχ², Δdf, ΔCFI, and ΔAIC for the compared models.

Standardized factor loadings were positive and statistically meaningful across the retained items, supporting the adequacy of the four-factor measurement structure. The CFA results therefore provided initial support for the retained 17-item SGATE-CE structure. The standardized factor loadings and error terms are presented in Figure 2, and the item-level loadings are reported in the final scale table.

Figure 2

3.4 Reliability and construct validity

Reliability and construct validity were examined for the four retained dimensions: Perceived Usefulness (PU), Behavioral Intention (BI), Applied Educational Engagement (AEE), and Perceived Conservation-Specific Utility (PCSU). Internal consistency was evaluated using Cronbach's alpha and McDonald's omega in both the exploratory and validation samples.

In the exploratory sample (n = 157), Cronbach's alpha values were .797 for PU, .886 for BI, .814 for AEE, and .826 for PCSU. McDonald's omega values were .804 for PU, .888 for BI, .815 for AEE, and .835 for PCSU. In the validation sample (n = 201), Cronbach's alpha values were .842 for PU, .896 for BI, .871 for AEE, and .839 for PCSU. McDonald's omega values were .852 for PU, .897 for BI, .872 for AEE, and .839 for PCSU. These results indicate satisfactory internal consistency across all retained subdimensions in both samples.

Composite reliability (CR) and average variance extracted (AVE) were calculated for the validation sample to evaluate construct reliability and convergent validity. CR values were .849 for PU, .897 for BI, .872 for AEE, and .843 for PCSU, exceeding the recommended .70 threshold. AVE values were .587 for PU, .636 for BI, .629 for AEE, and .574 for PCSU, exceeding the recommended .50 threshold. These findings supported the convergent validity of the four retained constructs.

Discriminant validity was examined using inter-factor correlations, the Fornell-Larcker criterion, and the heterotrait-monotrait ratio of correlations (HTMT). The square root of AVE for each construct exceeded its correlations with the other constructs: PU (.766), BI (.797), AEE (.793), and PCSU (.758). Inter-factor correlations were positive and ranged from .531 to .745, indicating moderate to moderately high relationships among the constructs. The highest correlations were observed between BI and PCSU (r = .745) and between AEE and PCSU (r = .722), suggesting conceptual closeness between students’ intention, applied educational engagement, and perceived conservation-specific utility. However, the Fornell-Larcker criterion was met for all constructs, and all HTMT coefficients were below the .85 threshold, with the highest HTMT value observed between BI and PCSU (.737). These results supported adequate discriminant validity among the four subdimensions. Reliability, composite reliability, AVE, and interfactor correlations are summarized in Table 6.

Bootstrap confidence intervals for HTMT provided additional discriminant-validity evidence. All HTMT point estimates remained below .85. The upper bootstrap interval for BI-PCSU slightly exceeded the conservative .85 threshold, whereas the values remained below or close to the .90 threshold used for conceptually related constructs. Accordingly, discriminant validity was interpreted as adequate but not definitive, with particular conceptual closeness between Behavioral Intention and Perceived Conservation-Specific Utility. The full HTMT bootstrap results are provided in Supplementary Table S15.

Because SGATE-CE is a multidimensional instrument, reliability evidence was interpreted at the subscale level rather than as evidence for a single global score. Accordingly, interpretation emphasizes the reliability and validity of the four dimensions separately.

3.5 Structural model testing

Structural model testing was conducted to examine the theoretically specified directional associations among Perceived Conservation-Specific Utility (PCSU), Perceived Usefulness (PU), Behavioral Intention (BI), and Applied Educational Engagement (AEE). Because H4 specified a direct association between PCSU and BI, the four-path model including PCSU → PU, PU → BI, PCSU → BI, and BI → AEE was treated as the primary hypothesized structural model. A more parsimonious sequential model excluding the direct PCSU → BI path was additionally examined as a comparison model. The comparison model showed weaker fit, whereas the hypothesized four-path model provided a better representation of the proposed associations.

The hypothesized four-path structural model showed better fit than the parsimonious comparison model. The CMIN/df value decreased from 2.616 to 2.190, CFI increased from .907 to .932, TLI increased from .892 to .920, RMSEA decreased from .090 to.077, SRMR decreased from .119 to .073, RMR decreased from .076 to .049, and AIC decreased from 377.489 to 327.808. Although PCLOSE = .000 did not indicate close fit, the combined fit indices supported the adequacy of the hypothesized four-path model for examining the proposed associations. The fit indices for both structural models are presented in Table 7.

All structural paths in the hypothesized four-path model were statistically significant under maximum likelihood estimation. PCSU was positively associated with PU [B = .494, SE = .077, 95% CI (.343, .645), CR = 6.406, p < .001, β = .635], supporting H3. PU was positively associated with BI under ML estimation [B = .303, SE = .118, 95% CI (.072, .534), CR = 2.575, p = .010, β = .211], providing partial and tentative support for H1 because this association was weaker and less stable under bootstrap sensitivity analysis. PCSU also had a direct positive association with BI [B = .708, SE = .108, 95% CI (.496, .920), CR = 6.548, p < .001, β = .636], supporting H4. Finally, BI was positively associated with AEE [B = .742, SE = .083, 95% CI (.579, .905), CR = 8.984, p < .001, β = .711], supporting H2. The structural path estimates for the hypothesized four-path model are presented in Table 8.

Estimator/objective sensitivity checks indicated that the structural paths remained positive across MLW, DWLS, and ULS objectives. A preliminary nonparametric bootstrap sensitivity check yielded positive intervals for PCSU → PU, PCSU → BI, and BI → AEE. The PU → BI interval crossed zero in this bootstrap check, so H1 was interpreted as significant under ML estimation but comparatively weaker and less stable than the other structural associations. These results are reported in Supplementary Table S13B.

The hypothesized four-path structural model explained 40.3% of the variance in PU, 62.0% of the variance in BI, and 50.6% of the variance in AEE. These findings indicate that students’ perceived conservation-specific utility of GenAI-supported tools is strongly associated with both general usefulness perceptions and behavioral intention. In addition, behavioral intention showed a strong positive association with applied educational engagement. The PU → BI association was positive and significant under ML estimation but should be interpreted cautiously because its bootstrap sensitivity interval crossed zero. Because the study used a cross-sectional, non-experimental design, these findings should be interpreted as theoretically specified directional associations rather than evidence of causal effects. A standardized diagram of the hypothesized four-path structural model is provided in Supplementary Figure S1.

3.6 Final SGATE-CE scale

The final version of SGATE-CE consisted of 17 items across four dimensions: Perceived Usefulness (PU), Behavioral Intention (BI), Applied Educational Engagement (AEE), and Perceived Conservation-Specific Utility (PCSU). The former Subject Matter Knowledge label was revised to PCSU because the retained items assess students’ perceived conservation-specific utility of GenAI-supported tools rather than objective disciplinary knowledge. The final scale uses a five-point Likert-type response format ranging from 1 = Strongly disagree to 5 = Strongly agree. Higher scores on each subscale indicate stronger perceived usefulness, stronger behavioral intention, higher applied educational engagement, or stronger perceived conservation-specific utility, respectively.

The final item structure and standardized CFA factor loadings are presented in Table 9. The original Turkish item wording and English translations are provided in the Supplementary Material to support transparency and reproducibility.

The final item wording in Table 9 reproduces the administered Turkish form and its informational English translation. Broader terms were not retrospectively replaced with “generative AI” because the reported EFA, CFA, reliability, and validity evidence applies to the wording participants actually completed. Any revised version that uniformly names GenAI should be treated as a new form requiring cognitive testing and psychometric revalidation.

4 Discussion

4.1 Psychometric quality, dimensional structure, and construct scope of SGATE-CE

This study developed and initially validated SGATE-CE as a context-sensitive instrument for assessing GenAI-contextualized acceptance among undergraduate architecture students in departments with conservation- and restoration-related curricular content. Based on an exploratory sample and a separate validation sample, the final 17-item structure retained four dimensions: Perceived Usefulness, Behavioral Intention, Applied Educational Engagement, and Perceived Conservation-Specific Utility. The findings provided initial evidence for the internal structure, reliability, convergent validity, and discriminant validity of the scale, while also indicating the need for external replication of the theory-refined 17-item structure.

The EFA results produced an interpretable four-factor structure after the removal of Perceived Ease of Use (PEOU) and additional items that showed weak loadings, cross-loadings, low communalities, or conceptual misfit. The intermediate EFA solution retained SMK2 statistically, but this item was removed before administration of the final validation form because its broad digital-tool and human-error-reduction wording did not sufficiently represent the revised PCSU construct. The retained 17-item four-factor structure was then tested in a separate validation sample and showed acceptable-to-good fit without adding post hoc error covariances or cross-loadings. This provides initial support for the internal structure of the final form, but it should not be interpreted as a fully independent confirmation of every item-refinement decision. External replication is therefore needed.

Reliability and construct validity evidence also supported the four retained subdimensions. Cronbach's alpha and McDonald's omega values were satisfactory across both samples, while CR and AVE values supported construct reliability and convergent validity in the validation sample. Discriminant validity was supported by the Fornell-Larcker criterion and HTMT values, although the moderate-to-high correlations among BI, AEE, and PCSU indicate that these constructs are conceptually close. This closeness is theoretically understandable because students who perceive GenAI-supported tools as useful for conservation-specific tasks may also be more willing to use them and more likely to report applied educational engagement.

The findings should be interpreted at the subscale level rather than as evidence for a single global score. SGATE-CE is a multidimensional instrument, and its four dimensions capture related but distinct aspects of GenAI-contextualized acceptance in conservation-related architectural education. Therefore, the four subscales should be interpreted separately instead of treating the instrument as a unidimensional total score.

From a psychometric perspective, these findings are consistent with scale-development recommendations that emphasize internal structure, reliability, convergent validity, discriminant validity, and validation in separate samples as core sources of early validation evidence (Boateng et al., 2018; DeVellis, 2016). At the same time, SGATE-CE should be positioned as an initial context-specific instrument rather than as a fully established universal measure. The present evidence supports the internal structure and reliability of the four subscales, but further external validation is still needed before the scale can be used for broad comparative, predictive, or cross-cultural purposes.

SGATE-CE measures GenAI-contextualized perceived acceptance within conservation-related architectural education. Its scores reflect students’ self-reported usefulness judgments, behavioral intention, applied educational engagement, and conservation-specific utility under an explicit instruction to consider generative AI tools. The scale does not establish that respondents distinguished generative systems from every adjacent technology, nor does it measure acceptance of all digital technologies in a technology-neutral manner.

The broader wording retained in several items introduces a risk of construct contamination because terms such as “AI technologies” and “technological tools” may evoke CAD, BIM, GIS, photogrammetry, conventional AI, or other digital systems. The common GenAI instruction provides a methodological anchor but cannot eliminate this risk. post hoc reformulation was avoided because changing the wording after data collection would disconnect the validated scores from the administered instrument. Future versions should directly compare explicit GenAI wording with broader AI and digital-tool wording, use cognitive interviews to examine participants’ referents, and test whether GenAI-contextualized acceptance is empirically distinguishable from broader AI-technology and digital-tool acceptance.

4.2 Removal of perceived ease of use from the final structure

The removal of PEOU is one of the most important findings of the scale-development process. In the initial item pool and 27-item draft form, PEOU was included in line with the classical TAM framework. However, the EFA results showed that the PEOU items did not form a clearly interpretable and psychometrically stable factor in the exploratory sample. For this reason, PEOU was removed from the final SGATE-CE structure.

This result should be interpreted cautiously. The removal of PEOU does not mean that ease of use is irrelevant to GenAI-contextualized acceptance. Rather, it indicates that the PEOU items developed for this draft instrument did not operate as a distinct empirical dimension in the present conservation-related architectural education sample. This may be related to item wording, overlap with other acceptance constructs, or the specific educational context in which the scale was administered. Because the present study did not directly measure students’ digital familiarity, prior GenAI experience, or technical self-efficacy in sufficient detail, the removal of PEOU should not be explained as a consequence of prior familiarity with digital tools.

The finding nevertheless has theoretical relevance. In the present conservation-related architectural education context, students may place greater emphasis on perceived usefulness, intention, applied educational engagement, and perceived conservation-specific utility than on the draft PEOU indicators. This interpretation is limited to the retained measurement structure and does not imply that SGATE-CE assesses students’ ability to judge whether outputs are ethically acceptable or disciplinarily valid. Future studies should develop and test alternative PEOU items that more precisely capture the usability of GenAI tools in conservation-related tasks and examine whether PEOU re-emerges in larger or more diverse samples. This cautious interpretation is also consistent with broader educational technology-acceptance research showing that the role of ease of use may vary by user group, technology type, and learning context (Scherer et al., 2019; Şumak et al., 2011).

4.3 Perceived conservation-specific utility: conceptual boundaries and reflective specification

The most important context-sensitive contribution of the present study concerns the former Subject Matter Knowledge dimension, which was relabeled as Perceived Conservation-Specific Utility (PCSU). The retained indicators reference heritage identity, material selection, interdisciplinary restoration work, and standards-aligned decision support. Although these task contexts are substantively different, the construct is bounded not by the tasks themselves but by the shared evaluative judgment they elicit: whether GenAI-supported tools are perceived as useful for conservation-related architectural reasoning. PCSU remains a perceived utility construct, not a measure of objective knowledge, professional competence, or the actual quality of conservation decisions.

PCSU was modeled as a reflective rather than formative construct. Under this specification, a stronger underlying perception of conservation-specific utility is expected to produce higher endorsement across several task contexts; the indicators are manifestations of that general appraisal. They are expected to covary because they express the same underlying judgment, and removing one task example would narrow content coverage without redefining the construct. A formative specification would instead treat heritage identity, material selection, interdisciplinary collaboration, and standards alignment as non-redundant components that jointly create the construct, which was not the intended conceptualization (Jarvis et al., 2003). PCSU therefore does not represent a checklist or aggregate index of conservation capabilities, and omission of a particular practice does not imply that the practice is unimportant.

The validation results provide initial, not definitive, support for this reflective interpretation. PCSU item loadings ranged from .705 to .852, composite reliability was .843, and AVE was .574, indicating shared variance consistent with a coherent latent perception. However, four items cannot rule out narrower subdomains, and the content breadth may conceal distinctions among technical, heritage-identity-related, collaborative, and standards-related utility. Future research should develop larger PCSU item pools and compare a single reflective factor with correlated-subdomain, higher-order, and theoretically justified formative alternatives.

The structural model results further characterize the associations involving PCSU. PCSU was positively associated with PU, indicating that students who reported stronger perceived conservation-specific utility also tended to report stronger general usefulness. PCSU was also positively associated with BI when PU was included in the model. In the present conservation-related architectural education sample, intention to use was therefore associated with both general usefulness and perceived conservation-specific utility. These associations do not establish that PCSU causes intention or functions as a mechanism.

These findings are consistent with previous technology-acceptance research. The PU → BI path was significant under maximum likelihood estimation and directionally consistent with the TAM proposition that perceived usefulness is positively associated with intention to use technology, but it should be interpreted cautiously because it was weaker under bootstrap sensitivity analysis (Davis, 1989). Similarly, the BI → AEE association is compatible with TAM- and UTAUT-based research in which intention is related to subsequent use or self-reported use behavior (Venkatesh et al., 2003, 2012; Venkatesh and Davis, 2000). In the present study, the PCSU → PU and PCSU → BI paths indicate that perceived conservation-specific utility was positively associated with both general usefulness perceptions and behavioral intention. These results identify context-specific associations within this architectural education sample; they do not establish that PCSU causes intention, produces engagement, or functions as a causal mechanism.

4.4 Epistemological, pedagogical, and ethical implications

The following discussion concerns the broader educational context of GenAI-supported conservation-related architectural education and should not be interpreted as evidence that SGATE-CE measures epistemic competence, subject knowledge, ethical judgment, critical evaluation, or responsible AI use. Its relevance to dedicated conservation-restoration and heritage programs requires further validation. In conservation-restoration education, knowledge is not validated only by the fluency, visual quality, or apparent coherence of an output, but through established conservation principles concerning authenticity, material evidence, intervention ethics, and heritage significance (Ahunbay, 2009; ICOMOS, 1964; UNESCO/ICCROM/ICOMOS, 1994). A conservation-related claim or design proposal must be evaluated in relation to historical evidence, material authenticity, intervention principles, documentation standards, and professional judgment. These requirements describe the educational setting in which acceptance is examined; they are not outcomes measured by SGATE-CE.

From an epistemological perspective, AI-generated outputs should be treated as provisional artifacts that require disciplinary validation. GenAI systems can generate plausible descriptions, interpretations, visualizations, or design alternatives, but plausibility does not guarantee historical accuracy, ethical appropriateness, or conservation validity. This distinction is especially important in heritage-related education, where visually persuasive outputs may reproduce stylistic assumptions, omit material evidence, or present speculative reconstructions as if they were documented knowledge. Therefore, students need to learn not only how to produce AI-supported outputs but also how to question, verify, justify, and revise them through conservation-specific evidence and standards.

The PCSU construct is relevant because it records whether students perceive AI-supported tools as useful for conservation-specific tasks. A high PCSU score indicates stronger perceived conservation-specific utility only. It does not indicate subject knowledge, ethical judgment, critical evaluation, responsible AI use, historical or material accuracy, or the quality of decisions. This distinction is central to the interpretation of SGATE-CE.

Pedagogically, the results may inform the design of structured learning tasks involving comparison, justification, and critique, but the present study did not test the effectiveness of any instructional intervention. Instructors may ask students to compare AI-supported outputs with archival evidence, measured drawings, conservation charters, material analyses, site observations, and expert feedback. Such activities are proposed as contextual applications rather than as competencies or learning outcomes measured by SGATE-CE.

Ethically, the use of GenAI in heritage education requires attention to authenticity, cultural representation, authorship, accountability, and professional responsibility. In conservation, an AI-generated proposal may be technically impressive but still incompatible with the cultural values, material conditions, or intervention ethics of a heritage site. Therefore, responsible use of GenAI requires students to understand the limits of AI-generated outputs and to avoid treating them as neutral or authoritative. The teacher's role remains essential in mediating the relationship between AI-supported production and disciplinary validation.

Accordingly, SGATE-CE should be understood as an initial instrument for assessing four dimensions of GenAI-contextualized acceptance among undergraduate architecture students receiving conservation- and restoration-related teaching: perceived usefulness, behavioral intention, applied educational engagement, and perceived conservation-specific utility. It does not measure subject knowledge, ethical judgment, critical evaluation, responsible AI use, conservation competence, or actual decision quality.

4.5 Limitations and future research

This study has several limitations. First, both samples were drawn from undergraduate architecture students in three architecture departments with conservation- and restoration-related curricular content in Türkiye using purposive sampling. Architecture students encounter conservation within a broader architectural curriculum and should not be treated as equivalent to students in dedicated conservation-restoration, architectural restoration, heritage conservation, or restoration-focused degree programs. Accordingly, the current evidence primarily supports SGATE-CE within conservation-related architectural education. Although participants were recruited from multiple institutions and cities, the findings should be interpreted as an initial validation in a specific national, disciplinary, and programmatic context rather than as evidence of broad statistical generalizability. The results should not be generalized directly to specialist conservation and heritage programs without validation in those settings. In addition, because the questionnaire was circulated through instructor-mediated and institution-related online communication channels, the exact number of students who received the invitation and the response rate could not be calculated. Future studies should test SGATE-CE with larger and more diverse samples, including students in dedicated conservation-restoration, architectural restoration, heritage conservation, and related programs, preferably using sampling procedures that allow response rates to be documented.

In addition, although the recruitment sources were identified as architecture departments at Erciyes University, Abdullah Gül University, and Bozok University, the analytical datasets did not retain case-level institutional counts, program-level subgroup identifiers, academic year or level, prior GenAI experience, frequency of GenAI use, specific GenAI tools used, or previous GenAI training in a form suitable for subgroup analysis. Therefore, the study could not examine whether SGATE-CE scores differed across institutions, academic levels, or prior GenAI-experience profiles.

Second, the present evidence is limited to internal structure, internal consistency, convergent validity, discriminant validity, and theoretically specified structural associations. The study did not examine test-retest reliability, temporal stability, predictive validity, concurrent validity, or measurement invariance across groups. The EFA produced an intermediate 18-item solution, whereas the CFA tested the theory-refined 17-item form administered to the validation sample. Thus, the use of separate samples reduced direct capitalization on chance, but the CFA did not confirm the exact intermediate EFA-derived structure. Because SMK2 was removed after EFA for conceptual reasons and was not administered to the 201-participant validation sample, an 18-item CFA including SMK2 could not be estimated. The final 17-item structure should therefore be interpreted as an initially tested, theory-refined structure requiring external replication rather than as a fully independent confirmation of all item-refinement decisions. Future research should assess whether the factor structure remains stable over time and whether the scale operates equivalently across gender, academic level, institution type, prior GenAI experience, or national context. Future validation studies should also replicate factor-retention and measurement-model evaluation in larger and more diverse samples, including parallel analysis, formal checks of multivariate normality, and robustness checks using alternative estimators where appropriate.

Relatedly, factor-retention evidence should be interpreted cautiously. Parallel analysis supported four factors under the mean random eigenvalue criterion but was more conservative under the 95th-percentile criterion. The retained four-factor structure is therefore justified by combining factor-retention evidence with theoretical interpretability, final EFA patterning, CFA fit, and alternative model comparisons rather than by relying on a single statistical rule.

Third, the study relied on self-report data. As a result, the findings may be influenced by social desirability, subjective evaluation bias, or students’ perceived access to AI tools. The Applied Educational Engagement dimension should therefore be interpreted as self-reported applied educational engagement rather than as direct behavioral observation or frequency-based actual use. Accordingly, the AEE label is used in this restricted sense and should not be read as observed actual use. Future studies should complement SGATE-CE with log data, performance-based tasks, studio rubrics, or instructor evaluations to examine how students actually use GenAI tools in conservation-related learning activities.

AEE is defined as a self-reported educational engagement construct rather than as directly observed behavior or a simple frequency measure. The retained items are interpreted as complementary indicators of applied engagement in GenAI-supported conservation-related architectural education: students’ reported use of AI-supported tools, the priority they assign to such tools in conservation-related tasks, their perceived opportunity to use them in hands-on educational activities, and their use for connecting conservation work with international charters and regulations. For this reason, the more cautious label Applied Educational Engagement was adopted in place of Applied Use, while the original dataset variable prefix AU was retained only as a technical coding label.

Fourth, although the questionnaire was administered anonymously to protect participants’ privacy, this also limited the possibility of e-mail-, account-, or IP-based duplicate screening. The dataset was checked for incomplete submissions and straight-lining response patterns. Response-pattern indicators were flagged for transparency but were not treated as definitive evidence of invalid responding or used as automatic exclusion criteria; however, more advanced careless-response indices or duplicate-response controls were not applied. Future studies may consider privacy-preserving methods for response verification.

Data-screening indicators should also be interpreted cautiously. Some zero-variance and long-string response patterns were identified, but these were not treated as automatic exclusion criteria because the items were positively worded and uniform endorsement could reflect substantive agreement. Future studies should retain response-duration metadata and include attention checks or more advanced response-quality indicators to support stronger screening decisions.

Fifth, the content review was expert-based and qualitative rather than a formal quantitative CVI- or CVR-based content validity procedure. Although seven experts with disciplinary and methodological expertise evaluated the item pool across two rounds and provided item-level written comments, the review form did not collect numerical relevance or essentiality ratings. Therefore, CVI and CVR values could not be calculated. The expert-review stage should be interpreted as preliminary qualitative evidence of content relevance and item suitability rather than as comprehensive quantitative content validation.

Sixth, no separate pilot test, cognitive interview, or student comprehension assessment was conducted before the main validation study. Although the item pool was reviewed by experts and the questionnaire instructions framed all scale sections within a GenAI-supported conservation-related architectural education context, future research should conduct cognitive interviews with students to examine how they interpret terms such as generative AI, authenticity, conservation standards, ethical responsibility, and AI-supported decision-making.

The English item wording also represents an informational translation of the Turkish administered form rather than an independently validated English-language version. The English wording was prepared by the first author, reviewed with the second author for semantic and disciplinary consistency, and additionally checked by an architect colleague with advanced English proficiency for readability and architectural terminology. However, no formal back-translation, independent bilingual panel review, cognitive interviewing, cross-cultural adaptation, or separate English-language validation was conducted. Future studies should complete these procedures before using SGATE-CE as an administered English-language instrument.

Seventh, the former Subject Matter Knowledge dimension was relabeled as Perceived Conservation-Specific Utility (PCSU) to align the construct label with the retained item content. PCSU records perceived conservation-specific utility only and should not be interpreted as subject knowledge, ethical judgment, critical evaluation, responsible AI use, conservation competence, or actual decision quality. Relationships between PCSU and independent measures of these distinct constructs should be examined in future studies.

Relatedly, because some retained items refer to broader “AI technologies” or “technological tools,” the present version carries a risk of construct contamination. The common instruction to consider generative AI tools contextualized the responses but cannot establish that participants excluded CAD, BIM, GIS, photogrammetry, conventional AI, or other digital technologies from their interpretations. SGATE-CE should therefore be interpreted as a GenAI-contextualized acceptance instrument rather than as a measure of exclusive GenAI acceptance. Future versions should compare explicit GenAI-specific wording with broader AI and digital-tool wording, conduct cognitive interviews, and test discriminant validity between GenAI-contextualized acceptance, broader AI-technology acceptance, and digital-tool acceptance.

Finally, GenAI technologies evolve rapidly, and students’ familiarity with these tools may change over time. Some items may therefore require periodic review, refinement, or revalidation as new tools and educational practices emerge. Separate instruments or independently validated complementary measures are needed to assess constructs such as critical evaluation, trust, transparency, authorship, ethical judgment, and responsible AI use; these constructs should not be inferred from SGATE-CE scores.

5 Conclusion

This study developed and initially validated SGATE-CE as a context-sensitive instrument for assessing GenAI-contextualized acceptance among undergraduate architecture students in departments with conservation- and restoration-related curricular content. Based on an exploratory sample and a separate validation sample, the final 17-item structure retained four dimensions: Perceived Usefulness, Behavioral Intention, Applied Educational Engagement, and Perceived Conservation-Specific Utility. The findings provided initial evidence for the internal structure, reliability, convergent validity, and discriminant validity of the scale, while also indicating the need for external replication of the final 17-item structure.

A central contribution of the retained model is the inclusion of Perceived Conservation-Specific Utility. PCSU is conceptualized as a reflective perceived-utility construct in which heritage identity, material selection, interdisciplinary restoration reasoning, and standards-based decision support are treated as manifestations of a broader appraisal of conservation-specific utility. The current four-item evidence supports this interpretation provisionally, but it does not rule out narrower subdomains or higher-order alternatives in future expanded item pools. PCSU does not measure subject knowledge, ethical judgment, critical evaluation, responsible AI use, conservation competence, or actual decision quality.

The structural model indicated positive associations of PCSU with Perceived Usefulness and Behavioral Intention, and a positive association of Behavioral Intention with Applied Educational Engagement. The PU → BI association was significant under maximum likelihood estimation but weaker and less stable under bootstrap sensitivity analysis. In this conservation-related architectural education sample, GenAI-contextualized acceptance was associated with both general usefulness perceptions and perceived conservation-specific utility. Because the study was cross-sectional and non-experimental, these relationships are interpreted as associations rather than causal effects or mechanisms.

SGATE-CE is therefore proposed as an initial, context-specific instrument for examining four dimensions of GenAI-contextualized acceptance in conservation-related architectural education. It does not exclusively isolate GenAI from broader AI or digital technologies, and it does not measure subject knowledge, ethical judgment, critical evaluation, responsible AI use, disciplinary competence, or actual decision quality. The Turkish version was the administered and psychometrically tested instrument; the English version is an informational translation only. Future studies should extend the present validation by examining revised GenAI-specific wording, construct separation from broader AI and digital-tool acceptance, temporal stability, predictive validity, concurrent validity, measurement invariance, and validation in dedicated conservation and heritage programs.

References

• 1

  AhunbayZ. (2009). Tarihi Çevre Koruma ve Restorasyon. 5th Edn. Istanbul: YEM Yayın.

  • Google Scholar
  • View reference in article

• 2

  BoatengG. O.NeilandsT. B.FrongilloE. A.Melgar-QuiñonezH. R.YoungS. L. (2018). Best practices for developing and validating scales for health, social, and behavioral research. Front. Public Health6, 149. 10.3389/fpubh.2018.00149

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 3

  ChanC. K. Y.HuW. (2023). Students’ voices on generative AI: perceptions, benefits, and challenges in higher education. Int. J. Educ. Technol. High. Educ20, 43. 10.1186/s41239-023-00411-8

  • CrossRef
  • Google Scholar
  • View reference in article

• 4

  ChanC. K. Y.ZhouW. (2023). An expectancy value theory (EVT) based instrument for measuring student perceptions of generative AI. Smart Learn. Environ10, 64. 10.1186/s40561-023-00284-4

  • CrossRef
  • Google Scholar
  • View reference in article

• 5

  DavisF. D. (1989). Perceived usefulness, perceived ease of use, and user acceptance of information technology. MIS. Q.13, 319–340. 10.2307/249008

  • CrossRef
  • Google Scholar
  • View reference in article

• 6

  DeVellisR. F. (2016). Scale Development: Theory and Applications. 4th edn. Thousand Oaks, CA: SAGE Publications.

  • Google Scholar
  • View reference in article

• 7

  FornellC.LarckerD. F. (1981). Evaluating structural equation models with unobservable variables and measurement error. J. Mark. Res18, 39–50. 10.1177/002224378101800104

  • CrossRef
  • Google Scholar
  • View reference in article

• 8

  GranićA.MarangunićN. (2019). Technology acceptance model in educational context: a systematic literature review. Br. J. Educ. Technol50, 2572–2593. 10.1111/bjet.12864

  • CrossRef
  • Google Scholar
  • View reference in article

• 9

  HayesA. F.CouttsJ. J. (2020). Use omega rather than Cronbach's Alpha for estimating reliability. Commun. Methods Meas14, 1–24. 10.1080/19312458.2020.1718629

  • CrossRef
  • Google Scholar
  • View reference in article

• 10

  HenselerJ.RingleC. M.SarstedtM. (2015). A new criterion for assessing discriminant validity in variance-based structural equation modeling. J. Acad. Mark. Sci43, 115–135. 10.1007/s11747-014-0403-8

  • CrossRef
  • Google Scholar
  • View reference in article

• 11

  HornJ. L. (1965). A rationale and test for the number of factors in factor analysis. Psychometrika30, 179–185. 10.1007/BF02289447

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 12

  HuL. T.BentlerP. M. (1999). Cutoff criteria for fit indexes in covariance structure analysis. Struct. Equ. Model6, 1–55. 10.1080/10705519909540118

  • CrossRef
  • Google Scholar
  • View reference in article

• 13

  ICOMOS (1964). International Charter for the Conservation and Restoration of Monuments and Sites (Venice Charter). Venice: ICOMOS.

  • Google Scholar
  • View reference in article

• 14

  JarvisC. B.MacKenzieS. B.PodsakoffP. M. (2003). A critical review of construct indicators and measurement model misspecification in marketing and consumer research. J. Consum. Res30, 199–218. 10.1086/376806

  • CrossRef
  • Google Scholar
  • View reference in article

• 15

  JinY.Martinez-MaldonadoR.GaševićD.YanL. (2025). GLAT: the generative AI literacy assessment test. Comput. Educ. Artif. Intell9, 100436. 10.1016/j.caeai.2025.100436

  • CrossRef
  • Google Scholar
  • View reference in article

• 16

  KaiserH. F. (1974). An index of factorial simplicity. Psychometrika39, 31–36. 10.1007/BF02291575

  • CrossRef
  • Google Scholar
  • View reference in article

• 17

  KlineR. B. (2011). Principles and Practice of Structural Equation Modeling. 3rd edn. New York, NY: Guilford Press.

  • Google Scholar
  • View reference in article

• 18

  LaupichlerM. C.AsterA.SchirchJ.RaupachT. (2022). Artificial intelligence literacy in higher and adult education: a scoping literature review. Comput. Educ. Artif. Intell3, 100101. 10.1016/j.caeai.2022.100101

  • CrossRef
  • Google Scholar
  • View reference in article

• 19

  LawsheC. H. (1975). A quantitative approach to content validity. Pers. Psychol28, 563–575. 10.1111/j.1744-6570.1975.tb01393.x

  • CrossRef
  • Google Scholar
  • View reference in article

• 20

  LongD.MagerkoB. (2020). “What is AI literacy? Competencies and design considerations,” in Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems, eds. BernhauptR.MuellerF.VerweijD.AndresJ. (New York, NY: Association for Computing Machinery), 1–16. 10.1145/3313831.3376727

  • CrossRef
  • Google Scholar
  • View reference in article

• 21

  MardiaK. V. (1970). Measures of multivariate skewness and kurtosis with applications. Biometrika57, 519–530. 10.1093/biomet/57.3.519

  • CrossRef
  • Google Scholar
  • View reference in article

• 22

  NgD. T. K.LeungJ. K. L.ChuS. K. W.QiaoM. S. (2021). Conceptualizing AI literacy: an exploratory review. Comput. Educ. Artif. Intell2, 100041. 10.1016/j.caeai.2021.100041

  • CrossRef
  • Google Scholar
  • View reference in article

• 23

  SchererR.SiddiqF.TondeurJ. (2019). The technology acceptance model (TAM): a meta-analytic structural equation modeling approach to explaining teachers’ adoption of digital technology in education. Comput. Educ128, 13–35. 10.1016/j.compedu.2018.09.009

  • CrossRef
  • Google Scholar
  • View reference in article

• 24

  SukkarA. W.FareedM. W.YahiaM. W.AbdallaS. B.IbrahimI.SenjabK. A. K. (2024). Analytical evaluation of Midjourney architectural virtual lab: defining major current limits in AI-generated representations of Islamic architectural heritage. Buildings14, 786. 10.3390/buildings14030786

  • CrossRef
  • Google Scholar
  • View reference in article

• 25

  ŞumakB.HeričkoM.PušnikM. (2011). A meta-analysis of e-learning technology acceptance: the role of user types and e-learning technology types. Comput. Hum. Behav27, 2067–2077. 10.1016/j.chb.2011.08.005

  • CrossRef
  • Google Scholar
  • View reference in article

• 26

  UNESCO/ICCROM/ICOMOS (1994). Nara Document on Authenticity. Nara: UNESCO/ICCROM/ICOMOS.

  • Google Scholar
  • View reference in article

• 27

  VenkateshV.BalaH. (2008). Technology acceptance model 3 and a research agenda on interventions. Decis. Sci39, 273–315. 10.1111/j.1540-5915.2008.00192.x

  • CrossRef
  • Google Scholar
  • View reference in article

• 28

  VenkateshV.DavisF. D. (2000). A theoretical extension of the technology acceptance model: four longitudinal field studies. Manage. Sci46, 186–204. 10.1287/mnsc.46.2.186.11926

  • CrossRef
  • Google Scholar
  • View reference in article

• 29

  VenkateshV.MorrisM. G.DavisG. B.DavisF. D. (2003). User acceptance of information technology: toward a unified view. MIS. Q.27, 425–478. 10.2307/30036540

  • CrossRef
  • Google Scholar
  • View reference in article

• 30

  VenkateshV.ThongJ. Y. L.XuX. (2012). Consumer acceptance and use of information technology: extending the unified theory of acceptance and use of technology. MIS. Q.36, 157–178. 10.2307/41410412

  • CrossRef
  • Google Scholar
  • View reference in article