Factors influencing large language model adoption among dental students: a cross-sectional study.

Factors influencing large language model adoption among dental students: a cross-sectional study - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Sci Rep . 2026 Apr 10;16:17113. doi: 10.1038/s41598-026-47512-8 Factors influencing large language model adoption among dental students: a cross-sectional study Angel M Joseph 1, ✉ , Rahaf E Almutairi 2 , Wejdan A Alrashidi 2 , Reem M Almutairi 1 , Yahya Theeban 1 , Sami Aldhuwayhi 1 , Ahmed Altuwalah 1 , Mohammed Z Mustafa 1 , Amar A Thakare 1 , Saikarthik Jayakumar 3 , Ashwin Joseph 4 , Sreekanth K Mallineni 5, 6 Author information Article notes Copyright and License information 1 Department of Restorative Dentistry and Prosthodontics, College of Dentistry, Majmaah University, 11952 Al-Majmaah, Saudi Arabia 2 College of Dentistry, Majmaah University, 11952 Al-Majmaah, Saudi Arabia 3 Department of Maxillofacial Surgery and Diagnostic Sciences, College of Dentistry, Majmaah University, 11952 Al-Majmaah, Saudi Arabia 4 ECHS Polyclinic, Manarcadu, Kottayam, Kerala 686019 India 5 Department of Pediatric Dentistry, Dr. Sulaiman Al Habib Medical Group, Ar Rayyan, Riyadh Saudi Arabia 6 Division for Globalization Initiative, Liaison Center for Innovative Dentistry, Graduate School of Dentistry, Tohoku University, Sendai, Miyagi Japan ✉ Corresponding author. Received 2026 Feb 15; Accepted 2026 Mar 31; Collection date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/ . PMC Copyright notice PMCID: PMC13230515  PMID: 41963506 Abstract This research evaluates the factors influencing the behavioural intention (BI) to adopt large language models (LLMs) among dental students in education, clinical decision support (CDS), and research, using the original unified theory of acceptance and use of technology (UTAUT) model, representing the first application of this model in this specific context. LLM adoption among Saudi dental students is unstructured and unregulated, making empirical evidence on adoption factors an educational and governance priority. A cross-sectional, web-based survey using convenience sampling was conducted among 700 students from six universities in central Saudi Arabia (60.71% aged 18–22 years; 58.71% female; 74% undergraduate). The UTAUT constructs: performance expectancy (PE), effort expectancy (EE), social influence (SI), and facilitating conditions (FC), were assessed using a five-point Likert scale. Measurement equivalence across English and Arabic language questionnaire versions were confirmed. All UTAUT constructs showed borderline acceptable to excellent internal consistency ( α = 0.678–0.907). Structural equation modelling (SEM) with bootstrapped 95% confidence intervals (CI) was used for statistical analysis. Hierarchical multiple regression was used for post-hoc sensitivity analysis to examine the moderating effects of demographic variables. The model accounted for 44.1% of the variance in BI (SEM-derived R 2 = 0.441, p < 0.001). EE was the strongest significant factor influencing BI ( β = 0.351, p = 0.017, 95% CI [0.056, 0.518]), followed by FC ( β = 0.136, p = 0.024, 95% CI [0.019, 0.183]). The PE ( β = 0.158, p = 0.162) was non-significant, due to its high latent correlation with EE, and SI ( β = 0.090, p = 0.131) was also non-significant, consistent with absent institutional mandates. The UTAUT–BI relationships were not significantly moderated by age, gender, or academic level. The results indicate strong expressed BI to adopt LLMs among dental students; however, BI is a theoretical proxy and should not be equated with actual adoption behaviour, clinical competence, or ethical and effective LLM application. These findings offer preliminary evidence which may inform curriculum design and institutional governance for responsible LLM integration in dentistry. The tentatively proposed LLM competency framework represents theoretical concepts which require empirical validation before implementation. Supplementary Information The online version contains supplementary material available at 10.1038/s41598-026-47512-8. Keywords: Artificial intelligence, Machine learning, Natural language processing, Clinical decision support, Dental education Subject terms: Health care, Medical research, Psychology, Psychology Introduction The Fédération Dentaire Internationale has implemented strong governance frameworks to integrate artificial intelligence (AI) in dental education, clinical practice, and research 1 . AI-driven diagnostic, treatment planning, and clinical decision support (CDS) tools are essential components in digital dentistry. Growing evidence supports its integration into dental education can improve evidence-based practice, clinical reasoning, and communication 2 – 4 . Using deep learning and natural language processing, large language models (LLMs) can recognize, create, and analyse human language 5 . The LLM landscape contains general-purpose foundational models and dental-specific multimodal vision language models 6 – 9 . LLM deployment across dental education, CDS, and research has fundamentally diverse performance profiles, risk levels, and governance requirements that has not been extensively investigated. Dental students confront LLMs in all three areas, even without domain-specific supervision. Instead of three discrete decisions, their adoption is best viewed as a common motivational disposition influenced by cross-domain exposure. Governance strategies must address LLM adoption within a single framework identifying the common motivational and contextual factors comprehensively. Although performance profiles and governance needs vary by domain, individual motivational variables of adoption, such as perceived usefulness, ease of use, social norms, and institutional support, are not domain specific. Thus, a single adoption assessment across all three domains is theoretically sound 10 . In this study, the education domain encompasses the use of LLMs for information acquisition, examination preparation, adaptive learning, and personalised feedback in the formal dental curriculum. Exam performance has dominated in this research context. A systematic evaluation and meta-analysis found that newer models perform better on international dental license exams but fall short of clinical applicability standards 11 . Discipline-specific analyses further reveal that performance is uneven across specialities, with accuracy ranging from 72.4 to 85.5% on National Board Dental Examination questions, alongside statistically significant inter-model and inter-disciplinary variation 12 . In prosthodontics, LLM performance on specialisation examination questions has been shown to depend on querying method and model version 13 , with chatbot accuracy on prosthodontic education content ranging from 54.8% to 73% across platforms 14 . In endodontics, performance on patient-facing clinical questions varies markedly across models, with scores ranging from 4.81/5.00 for the highest-performing model to substantially lower ratings for others. Systematic misinformation tendencies have been documented regardless of overall performance level 15 . Beyond examination contexts, LLMs have shown potential as personalised feedback tools. AI-assisted feedback on radiographic diagnostic performance has been shown to significantly improve students’ diagnostic accuracy compared to conventional feedback under expert supervision 16 . Collectively, these findings suggest that while LLMs offer meaningful educational support, their performance remains variable, model-dependent, and insufficiently reliable for unsupervised use across dental specialities 6 , 17 , 18 . This study defines the CDS domain as LLM use in in supporting diagnostic reasoning, treatment planning, and patient communication within supervised clinical training, where students apply LLM-generated information under faculty oversight rather than as independent decision-makers. LLM potential has been documented across multiple dental specialities, including implantology, orofacial pain treatment, and clinical anamnesis 8 , 19 – 23 . Evidence shows that LLMs may quickly synthesize and disseminate clinical information but cannot duplicate expert clinical thinking 6 , 11 , 24 . LLMs diagnosed oral lesions with 87.4% accuracy but declined on challenging clinical issues, suggesting a complementing role under clinician supervision 25 . High response consistency and a mean clinical accuracy of 57.33% with significant difficulty-dependent variation in endodontics suggest that LLMs cannot replace clinical decision-making 26 . Variable performance across diagnostic categorization dimensions in periodontology showed unreliable independent diagnostic support 27 . Expert oversight in implantology is strengthened by context-dependent dependability across complicated contexts 8 . LLMs supporting prosthetic treatments are also insufficient for unsupervised clinical application in prosthodontics 28 . Image-dependent applications further compound this performance ceiling. LLM reliability decreases significantly in visual data interpretation tasks including radiographic analysis and orthodontic aesthetic evaluation due to visual input restrictions of text-trained models 29 , 30 . Thus, text-based versus image-based activities are a clinically important LLM deployment barrier. Multi-model evaluations show that no single LLM optimises accuracy, comprehensiveness, and readability across clinical circumstances 31 , reinforcing the need for expert monitoring in clinical dental settings. Speciality, job complexity, and model affect the gap between LLM information retrieval and clinical decision-making skill. This variation highlights the need for research in empirical adoption factors to facilitate contextually appropriate, governance-informed LLM integration across dentistry specialities. LLM support for literature searching, evidence synthesis, academic writing, and hypothesis development in undergraduate and postgraduate dentistry research is the research domain in this study. LLMs are useful for literature screening, data synthesis, and scientific writing 32 – 34 . They pose a unique and underestimated risk in research. LLMs may fabricate citations, misattribute findings, or produce statistically consistent but factually wrong summaries, worsening evidence quality and scientific integrity 24 , 34 . Their use in dental research is still unregulated 34 , and their inclusion into systematic review and literature screening workflows is beginning to be formally examined 33 . Despite rapid LLM integration in dentistry, formal incorporation into dental curricula in Saudi Arabia remains limited, and institution-specific governance frameworks are lacking. In a countrywide survey of 1,370 Saudi dental students, 58.1% were aware of LLM capabilities, but only 15% reported institutional guidelines, with ethical concerns about plagiarism and academic dishonesty 35 . The Saudi Data and Artificial Intelligence Authority (SDAIA) has issued generative AI guidelines for government employees and the public 36 , 37 and announced a national cross-disciplinary AI curriculum for university students 38 . However, there are no national LLM-specific adoption policy or governance framework for health professions or dental education at present. Therefore, Saudi dental students are already adopting LLMs in an unstructured and unregulated manner, making empirical evidence on adoption factors an important educational and governance priority. LLMs have documented “hallucination” potential 24 , 34 , which raises concerns about dependability and patient safety in high-stakes dental education and clinical practice 11 , 24 . Overall, the evidence across these three domains indicates that LLM performance in dentistry is variable, context-dependent, and poorly controlled. Performance studies cannot address the factors that motivate or deter the dental students from adopting this technology. Understanding adoption factors is necessary to translate performance information into effective educational and governance frameworks, and this constitutes the specific research problem addressed in this study. Saudi Vision 2030 aims for healthcare digitalization are driving institutional momentum to integrate AI and digital technology into undergraduate dental curricula. A countrywide assessment of Saudi dental schools indicated that just 64.4% had formally embraced digital dental technology 39 , and dedicated AI education including LLM-specific proficiency is inconsistent. Updated dental curricula for AI and LLM-specific literacy are identified as a global educational need 17 , 18 . This adoption-focused evidence is lacking in Saudi Arabia which can aid curriculum designers and institutional policymakers to understand the students’ motivational willingness to embrace these technologies, making this study’s findings relevant to Saudi dentistry education policy. The original unified theory of acceptance and use of technology (UTAUT) and its successor UTAUT2 have been validated for assessing technology adoption in medical education and healthcare 40 – 43 , but their application in dental education is underexplored. UTAUT2 includes consumer-focused constructs, but the original UTAUT model was chosen for this study because it was developed and validated for organizational and institutional technology adoption contexts 10 , making it the most theoretically appropriate framework for structured, professionalized educational settings like dental education. Prior UTAUT studies have examined general AI and chatbot adoption in medical education 40 – 43 , but none have investigated LLM-specific adoption in dental education. Informal student use, inter-model variation in task-specific performance, and the lack of institutional governance indicate that performance benchmarking alone does not sufficiently represent the dynamics of LLM adoption. Consequently, theoretical frameworks that examine motivational and contextual factors influencing technology acceptance, such as UTAUT, are necessary to explain adoption trends in formal educational environments. The original UTAUT model evaluates four constructs predicting behavioural intention (BI) to adopt LLM technology. Performance expectancy (PE) is the assumption that LLMs improve academic and professional performance in education, CDS, and research. Effort expectancy (EE) measures perceived ease of learning and using LLMs. Social influence (SI) refers to the influence of faculty, peers, and institutional norms on students’ willingness to adopt LLMs. Facilitating conditions (FC) includes available institutional support, infrastructure, and training to support LLM use. Figure 1 shows that all four constructs directly influence BI 44 . According to UTAUT, larger PE, EE, SI, and FC indicate stronger BI. These relationships are additive rather than hierarchical, and the model does not assume that one construct subsumes another 10 . BI captures students expressed willingness and likelihood of using LLMs across educational, CDS, and research contexts. BI is recognized as a theoretically proven proxy for technology adoption behavior 10 and should not be equated with actual adoption behaviour, clinical competence, or ethical and effective LLM application. The UTAUT model also identifies four moderating variables (age, gender, experience, and voluntariness of use) that affect construct-BI correlations 10 . Age and gender were explored as demographic factors in this study. Since LLM-specific competency is not formally integrated into the AI literacy component of dental curriculum in Saudi Arabia, experience and voluntariness of use were excluded from the model. The documented LLM-specific dangers are theoretically relevant to these constructs. Hallucination and fluctuating precision may lower PE because inconsistent outputs lower students’ LLM utility perceptions 10 , 24 , 34 . Text-based engagement may exaggerate EE regardless of clinical reliability, generating a gap between apparent ease-of-use and real performance 10 . Institutional oversight deficits in FC may limit adoption preparedness and increase patient safety hazards 10 , 11 , 24 . SI may influence risk awareness since faculty endorsement and peer norms may encourage cautious LLM usage or normalise uncritical use due to the lack of guidelines 10 , 35 , 39 . Collectively, examining the constructs influencing BI holds significant implications for both educational practice and institutional governance of LLM adoption. Fig. 1. Open in a new tab UTAUT model (Copyright: Sage Publications 2024) 44 . Therefore, the aim of this study was to evaluate the factors influencing the BI to adopt LLMs among dental students across educational, CDS, and research, using the UTAUT framework. This study addressed the following research questions (RQ). RQ1: Which UTAUT constructs are significant factors influencing dental students’ BI to adopt LLMs across educational, CDS, and research contexts? RQ2: Does the UTAUT model, as a whole, significantly explain the variance in BI among dental students in this context? RQ3: Do demographic variables, including age, gender, and academic level, moderate the relationships between UTAUT constructs and BI? The UTAUT model provided the foundation for the following five hypotheses: H1: PE is positively associated with BI to adopt LLMs among dental students. H2: EE is positively associated with BI to adopt LLMs among dental students. H3: SI is positively associated with BI to adopt LLMs among dental students. H4: FC is positively associated with BI to adopt LLMs among dental students. H5: The UTAUT model significantly predicts BI to adopt LLMs among dental students. Theoretically, this study is among the first to apply UTAUT framework to LLM-specific adoption in dental education, CDS, and research. Practically, since Saudi dental students are already adopting LLM in an unstructured and ungoverned manner, the findings provide preliminary empirical evidence to inform curriculum design and institutional governance for LLM integration within Saudi Arabia’s Vision 2030 healthcare digitalisation mission. The gap between expressed intention and actual behaviour is well recognised within technology adoption research 10 , and findings should be interpreted within this boundary. Methods Ethical considerations This research was conducted after approval from the Majmaah University Institutional Review Board (IRB No. MUREC-Dec.09/COM-2025/193) and was conducted in accordance with the World Medical Association Declaration of Helsinki and its revision in 2024. Informed consent was obtained from all students prior to participation in the study. Sample size calculation An a priori sample size calculation was performed in G*Power (version 3.1) using multiple linear regression parameters (fixed model, R 2 deviation from zero), with a two-tailed significance level of 0.05, 80% power, and four UTAUT constructs as independent variables. Based on prior UTAUT-based research within the same regional context 44 reporting a model R 2 of 0.32 for the Saudi dental community, a conservative R 2 of 0.15 was adopted for the present study to account for the LLM-specific context and student-specific population, where adoption may be influenced by factors beyond the classical UTAUT constructs, thereby reducing the expected explanatory power. The corresponding Cohen’s effect size was derived using the formula f 2 = R 2 /(1 − R 2 ) = 0.15/0.85 ≈ 0.18 45 , representing a small-to-medium effect, and yielding a minimum required sample size of 85 participants. The final sample of 700 participants substantially exceeded both the regression-based minimum of 85 and the structural equation modelling (SEM) minimum of 200 46 , yielding a participant-to-observed-variable ratio of 46.7:1, well above the recommended 10:1 threshold 47 . This large sample was chosen to maximise statistical power, parameter estimate stability, and institution-wide generalisability. Based on f 2 = 0.18 with four constructs at a two-tailed significance level of 0.05, the achieved statistical power exceeded 99%, confirming adequacy for all primary SEM and post-hoc regression analyses. Study instrument and survey design The questionnaire (Supplementary file online) was developed by Author 1 and adapted from previous research 44 with minor modifications to align with the study aim (Copyright: Sage Publications 2024). The first section included demographic data, namely age, gender, and academic level, identified as the most contextually relevant variables for a student-specific institutional sample 42 . Among the four UTAUT moderating variables, age and gender were collected and examined as direct moderators. Academic level was retained as a contextual demographic variable, though it does not directly correspond to any UTAUT moderator. The second section operationalised the UTAUT constructs, adapted with minimal modification from the original model 10 by replacing the word “system” with “LLMs”, and recorded participants’ perceptions using five-point Likert-scale items ranging from 1 (strongly disagree) to 5 (strongly agree). The questionnaire comprised 17 items distributed across five constructs: PE (3 items), EE (4 items), SI (3 items), FC (4 items), and BI (3 items). This study measures BI to adopt LLMs, reflecting students’ motivational readiness, and should not be equated with actual adoption behaviour, clinical competence, or ethical and effective LLM application. The questionnaire was administered in both English and Arabic. In Saudi Arabian health professions education, English serves as the primary medium of instruction across dental and medical programmes 43 , and the Arabic version was offered as a language preference option for respondents more comfortable completing the survey in Arabic. The Arabic version was developed through a forward-backward translation procedure. It was first translated into Arabic by a native Arabic-speaking team member with professional English proficiency (Author 2), then back-translated into English by a second independent team member with similar proficiency (Author 3) to verify conceptual and linguistic equivalence. Discrepancies were resolved through team consensus (Authors 4–7) before finalizing the Arabic version. Content validity of both language versions was jointly evaluated by a panel of four subject-matter experts with demonstrated proficiency in AI applications and survey methodology. Each expert independently rated every item on a four-point relevance scale (1 = not relevant, 2 = somewhat relevant, 3 = quite relevant, 4 = highly relevant), consistent with the Content Validity Index (CVI) framework recommended by Lynn 48 and Polit and Beck 49 . The item-level CVI (I-CVI) was calculated as the proportion of experts assigning a rating of 3 or 4, with an acceptable threshold of ≥ 0.75 for a four-expert panel 48 , 49 ; the scale-level CVI (S-CVI/Ave) was computed as the mean of all item-level CVIs, with a threshold of ≥ 0.90 indicative of excellent content validity 48 , 49 . All items achieved an I-CVI of 1.00 and an overall S-CVI/Ave of 1.00, which were well above the acceptable thresholds, confirming excellent content validity across both language versions, and no revisions to item wording were required prior to instrument finalisation. The instrument was subsequently piloted among twenty dental students per language version to assess clarity and internal consistency prior to full-scale administration. All pilot participants completed the survey without requesting clarification or reporting difficulty, providing informal evidence of item clarity and comprehensibility. The pilot study demonstrated acceptable internal consistency across all UTAUT constructs in both language versions, with Cronbach’s alpha ( α ) values ranging from 0.606 to 0.882 for the Arabic version ( n = 20) and from 0.657 to 0.873 for the English version ( n = 20). The marginally lower α values for the FC construct in both versions (α = 0.606 and 0.660, respectively) were acceptable for a pilot sample, consistent with the widely accepted threshold of α ≥ 0.60 for exploratory and pilot studies 50 . These findings confirmed the overall suitability of the instrument for full-scale administration without modification. Study design and sampling This cross-sectional study is reported following the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) recommendations 51 . A cross-sectional, web-based survey was conducted among dental students aged 18 or older enrolled in universities in central Saudi Arabia during December 2025 and January 2026. Convenience sampling was utilized to acquire data via WhatsApp, Snapchat, X (previously Twitter), and Telegram, which represent the most widely used social media platforms among Saudi university students 52 . No formal sampling frame was defined; hence the number of respondents, coverage, and response rates could not be calculated prospectively. This pragmatic sampling approach is used in web-based dental and medical education research, especially for geographically scattered student populations 53 . This sampling approach may introduce selection bias, as students with greater digital literacy or prior LLM exposure may have been more likely to participate, potentially overrepresenting positive BI. The potential risk was reduced by several strategies. To maximize reach and institutional diversity, the survey was disseminated across multiple platforms, a large sample of participants was targeted to improve estimate stability, and the demographic distribution of respondents across age, gender, academic level, and university was examined post-hoc to partially characterize sample representativeness without a formal response rate 54 . These measures decrease but cannot eliminate sampling bias and thus findings should be interpreted with caution regarding generalisability beyond this population. Participants were informed that their responses were anonymous, that there were no right or incorrect answers, and that their participation would not affect their academic standing to reduce social desirability bias. Several measures were implemented to ensure data quality and integrity. The survey was administered via Google Forms with account-based sign-in and one-response-per-account which prevented duplicate entries in the survey. All items were set as mandatory, ensuring only fully completed responses were accepted and eliminating incomplete submissions. The survey was structured sequentially, starting with demographic screening questions. Conditional branching logic automatically redirected ineligible respondents to a disqualification page, preventing further progression. UTAUT model access was limited to eligible respondents. As a result, all submitted responses were complete, met inclusion criteria, and were included in the study. Statistical analysis Multi-group confirmatory factor analysis (MGCFA) was conducted using IBM AMOS (version 29; IBM Corp., Chicago, USA), to confirm that the choice of questionnaire language version did not introduce measurement artefacts. Since language preference was recorded as an explicit variable at the outset of the survey, participants were classified into English ( n = 500) and Arabic ( n = 200) language groups for this analysis. Both groups independently met the minimum sample size requirements for reliable MGCFA 46 , 47 . Three sequential nested models were evaluated: a configural model confirming the same factor structure across both language groups, a metric invariance model constraining factor loadings to equality across groups, and a scalar invariance model additionally constraining item intercepts to equality. Measurement equivalence was evaluated using changes in the comparative fit index (ΔCFI ≤ 0.010) and root mean square error of approximation (ΔRMSEA ≤ 0.015) between nested models 55 . To assess the potential threat of common method bias (CMB), which may arise from collecting all variables through a single self-administered questionnaire at a single time point, two complementary approaches were employed. Harman’s single-factor test was performed by entering all study variables into an exploratory factor analysis; if the first unrotated factor accounted for less than 50% of the total variance, CMB was considered unlikely to severely distort the observed associations 56 . A common method factor model within the CFA framework was additionally attempted as a more rigorous statistical assessment of CMB, though its estimability was contingent on model convergence and the absence of inadmissible solutions. Descriptive statistics were used to describe demographic variables and the UTAUT constructs. Internal consistency of the Likert scale items was assessed using α . To evaluate the measurement model, a confirmatory factor analysis (CFA) was conducted to assess construct validity, convergent validity via average variance extracted (AVE ≥ 0.50) and composite reliability (CR ≥ 0.70), and discriminant validity using the heterotrait-monotrait (HTMT ≤ 0.90) ratio criterion 57 , 58 . SEM was selected as the primary analytical approach as it simultaneously estimates the direct effects of UTAUT constructs (PE, EE, SI, and FC) on BI while accounting for measurement error and evaluating both measurement and structural components within a single model. To obtain confidence intervals (CI) for the structural path coefficients and enhance the robustness of parameter estimates, bias-corrected bootstrapping with 5,000 resamples was performed in AMOS, generating 95% CI for all standardised path coefficients. Model fit was evaluated using the following indices: chi-square/ df ratio (≤ 3.0), CFI ≥ 0.90, Tucker-Lewis index (TLI ≥ 0.90), RMSEA ≤ 0.08, and standardised root mean square residual (SRMR ≤ 0.08). Demographic variables were not included as covariates in the primary SEM analysis, as the study aimed to evaluate the direct effects of UTAUT constructs on BI. They were subsequently incorporated as covariates in a hierarchical multiple regression model as a post-hoc sensitivity analysis to examine their potential influence on BI beyond the UTAUT constructs. Prior to conducting hierarchical multiple regression, all standard analytical assumptions were systematically examined using IBM SPSS (version 31; SPSS Inc., Chicago, USA). Normality of residuals was assessed using the Shapiro–Wilk test in conjunction with visual inspection of normal quantile–quantile (Q–Q) plots and frequency histograms of standardised residuals. Visual assessment of plots of standardised residuals against standardised predicted values and the Breusch-Pagan test assessed homoscedasticity. Linearity was examined through partial regression plots. Independence of residuals was verified using the Durbin-Watson statistic, with values between 1.5 and 2.5 considered indicative of acceptable independence 54 . Outliers were identified through standardised residuals exceeding ± 3.0, leverage values, and Cook’s distance, with values exceeding 1.0 indicating sensitivity analysis. Multicollinearity among UTAUT constructs was assessed using variance inflation factors (VIF) and tolerance values, given the theoretical intercorrelations anticipated among PE, EE, SI, and FC. Conservative thresholds of VIF < 5.0 and tolerance > 0.20 were adopted as primary criteria 47 , 57 , with VIF > 10.0 and tolerance < 0.10 considered indicative of severe multicollinearity requiring remediation 47 , 54 , 57 . Condition indices and variance decomposition proportions were used to detect collinearity patterns not recorded by VIF, with indices exceeding 30 considered problematic 57 . All assumptions were satisfactorily met prior to interpreting the regression results. The F -statistic and p -value were used to evaluate model fit at each hierarchical step. Demographic factors were quantified by measuring the change in R 2 (Δ R 2 ) and the corresponding F -change statistic (Δ F ) and its significance 54 , 57 . Explanatory power was assessed using R 2 and adjusted R 2 , with the latter being the favoured measure due to its compensation for covariates and sample size 54 . Effect size for each regression model was quantified using Cohen’s f 2 , calculated as f 2 = R 2 /(1 − R 2 ), and interpreted according to established benchmarks: small ( f 2 ≥ 0.02), medium ( f 2 ≥ 0.15), and large ( f 2 ≥ 0.35) 45 . The 95% CI was reported for all regression coefficients to characterise the precision of parameter estimates. All analyses were conducted at a two-tailed significance level of p < 0.05. Results Sample characteristics Table 1 presents the sample characteristics. Out of 700 dental students, 500 (71.43%) completed the English version and 200 (28.57%) completed the Arabic version. The majority were aged 18 to 22 years ( n = 425, 60.71%), followed by those aged 23 to 27 years ( n = 243, 34.71%), and 28 years or older ( n = 32, 4.57%), comparable with the expected age profile of dental students. Most participants were female ( n = 411, 58.71%), reflecting the dental student demographic composition in Saudi Arabia 59 . The majority were undergraduates in preclinical ( n = 200, 28.57%) and clinical ( n = 318, 45.43%) training phases, collectively representing 74% of the sample. Internship students comprised 16.86% ( n = 118) and postgraduate students 9.14% ( n = 64). Six Saudi universities offering dentistry programs in the central region responded to the survey, comprising Majmaah University ( n = 255, 36.43%), Qassim University ( n = 158, 22.57%), King Saud University ( n = 102, 14.57%), Princess Nourah Bint Abdulrahman University ( n = 87, 12.43%), Riyadh Elm University ( n = 66, 9.43%), and Vision Colleges ( n = 32, 4.57%). Table 1. Sample characteristics of dental students ( n = 700). Characteristics Category n % Language Arabic 200 28.57 English 500 71.43 Age (in years) 18–22 425 60.71 23–27 243 34.71 ≥ 28 32 4.57 Gender Male 289 41.29 Female 411 58.71 Academic level Preclinical (First and Second year) 200 28.57 Clinical (Third, Fourth and Fifth year) 318 45.43 Internship 118 16.86 Postgraduate 64 9.14 University Majmaah University 255 36.43 Qassim University 158 22.57 King Saud University 102 14.57 Princess Nourah Bint Abdulrahman University 87 12.43 Riyadh Elm University 66 9.43 Vision Colleges 32 4.57 Total – 700 100 Open in a new tab n , Number of participants; %, Percentage of total sample. Measurement equivalence across language versions Both the English ( n = 500) and Arabic ( n = 200) groups independently met the minimum sample size requirements for reliable MGCFA 46 , 47 . The configural model confirmed that the same five-factor UTAUT structure was supported across both language groups (CFI = 0.948, TLI = 0.932, RMSEA = 0.086, SRMR = 0.043). The metric invariance model demonstrated that factor loadings were equivalent across groups (ΔCFI = 0.001, ΔRMSEA =  − 0.004, Δ χ 2 (10) = 2.885, p = 0.984), supporting metric invariance. The scalar invariance model additionally confirmed that item intercepts were equivalent across groups (ΔCFI = 0.000, ΔRMSEA =  − 0.003, Δ χ 2 (10) = 7.455, p = 0.682), supporting full scalar invariance. All ΔCFI values were well within the recommended criterion of ≤ 0.010 55 and all ΔRMSEA values within ≤ 0.015 60 . These findings confirm that the UTAUT constructs were measured equivalently across both language versions, fully justifying the pooling of data for subsequent analyses. CMB assessment To assess the potential threat of CMB, Harman’s single-factor test was performed by entering all 17 original study items into an unrotated principal components analysis, and the first unrotated factor accounted for 47.744% of the total variance, below the 50% threshold 56 , suggesting that CMB is unlikely to severely distort the observed associations among constructs. The common method factor model within the CFA framework was not estimable due to negative residual variances (Heywood case), likely reflecting the high inter-construct correlations and the resulting inability to distinguish the method factor from the substantive UTAUT constructs. Descriptive statistics and internal consistency Table 2 presents the descriptive statistics and internal consistency values for all constructs. Mean scores ranged from 3.47 (FC, 95% CI [3.42, 3.53]) to 4.18 (PE, 95% CI [4.11, 4.24]), with the mean BI score of 4.10 (95% CI [4.03, 4.16]) reflecting a strong expressed intention to adopt LLMs in education, CDS, and research. The full response scale was utilised across the sample (range: 1.00–5.00) for all constructs. Median values were closely aligned with means across all constructs (PE: 4.33, EE: 4.00, SI: 4.00, FC: 3.50, BI: 4.00), consistent with the generally negative skewness observed across constructs. Internal consistency was satisfactory across all constructs, with PE ( α = 0.901) and SI ( α = 0.907) demonstrating excellent reliability, EE ( α = 0.884) and BI ( α = 0.861) demonstrating good reliability, and FC ( α = 0.678) demonstrating borderline acceptable reliability 50 . The borderline α for the original four-item FC scale is attributable to the weak item compatibility of FC3 and FC4, which demonstrated low corrected item-total correlations ( r = 0.439 and r = 0.359 respectively). These items were subsequently removed based on unacceptably low standardised factor loadings in the measurement model (FC3 = 0.340; FC4 = 0.244), as reported below. The retained two-item FC construct demonstrated strong inter-item correlation ( r = 0.709), α = 0.829, AVE = 0.714, and CR = 0.833, confirming satisfactory construct validity 50 , 57 . All construct scores demonstrated acceptable distributional characteristics, with skewness values ranging from − 1.46 (PE) to − 0.26 (FC) and kurtosis values ranging from − 0.55 (SI) to 2.87 (PE), all within the acceptable thresholds of |skewness|< 2 and |kurtosis|< 7 for large samples 54 . Table 2. Descriptive statistics and internal consistency of UTAUT constructs ( n = 700). Construct Items Mean SD 95% CI Range Median Skewness Kurtosis α PE 3 4.18 0.85 [4.11, 4.24] 1.00–5.00 4.33 − 1.46 2.87 0.901 EE 4 4.11 0.87 [4.04, 4.17] 1.00–5.00 4.00 − 1.30 2.18 0.884 SI 3 3.76 1.08 [3.68, 3.84] 1.00–5.00 4.00 − 0.55 − 0.55 0.907 FC 4 3.47 0.74 [3.42, 3.53] 1.00–5.00 3.50 − 0.26 0.62 0.678 BI 3 4.10 0.85 [4.03, 4.16] 1.00–5.00 4.00 − 0.96 0.98 0.861 Open in a new tab PE, Performance expectancy; EE, Effort expectancy; SI, Social influence; FC, Facilitating conditions; BI, Behavioural intention; SD, Standard deviation; CI, 95% Confidence interval ( t -based, df = 699); α , Cronbach’s alpha. All items were rated on a five-point Likert scale (1 = strongly disagree, 5 = strongly agree). Range reflects that all participants utilised the full response scale for each construct. All skewness and kurtosis values are within acceptable ranges (|skewness|< 2, |kurtosis|< 7) 54 . Cronbach’s α interpreted as: ≥ 0.90 excellent, ≥ 0.80 good, ≥ 0.70 acceptable, ≥ 0.60 marginal; values falling between 0.60 and 0.70 are considered borderline acceptable 50 . Measurement model An initial five-factor CFA was conducted on all 17 items. FC3 and FC4 demonstrated unacceptably low standardised factor loadings (0.340 and 0.244 respectively, below the recommended threshold of ≥ 0.50 47 , 57 ) and were therefore removed from the measurement model. A revised five-factor CFA was subsequently conducted on the 15-item model, which demonstrated satisfactory fit: χ 2 (80) = 371.443, χ 2 / df = 4.643, Robust CFI = 0.932, Robust TLI = 0.911, RMSEA = 0.072 [90% CI: 0.067, 0.078], SRMR = 0.046. Although the chi-square/ df ratio exceeded the conventional threshold of 3.0, this is widely recognised as an expected outcome with large samples ( n = 700) and was therefore not considered indicative of poor fit 46 , 47 . All standardised factor loadings were statistically significant ( p < 0.001) and ranged from 0.653 to 0.952, with BI1 (0.653) being marginally acceptable. Convergent validity was supported for all constructs, with AVE values exceeding the 0.50 threshold 57 (PE = 0.762, EE = 0.664, SI = 0.769, FC = 0.714, BI = 0.699) and CR values exceeding 0.70 57 (PE = 0.905, EE = 0.887, SI = 0.909, FC = 0.833, BI = 0.872). Discriminant validity was assessed using the HTMT criterion 58 ; all construct pairs demonstrated HTMT values below the threshold of 0.90, with the exception of the PE-EE pair (HTMT = 0.886), which approached but did not exceed this threshold. This borderline value is theoretically consistent with the established conceptual proximity between PE and EE within the UTAUT framework 10 . Structural model assessment and hypothesis testing The structural model was estimated on the basis of the satisfactory measurement model fit confirmed in the CFA (Robust CFI = 0.932, RMSEA = 0.072) 46 , 47 . Bias-corrected bootstrapped 95% CI based on 5,000 resamples were generated for all structural path coefficients. The UTAUT constructs collectively explained 44.1% of the variance in BI ( R 2 = 0.441, 95% CI [0.387, 0.495], p < 0.001, Cohen’s f 2 = 0.789, large effect 45 ), supporting H5. EE emerged as the strongest significant factor influencing BI ( β = 0.351, z = 2.378, p = 0.017, 95% CI [0.056, 0.518], excluding zero), supporting H2; FC was also a significant factor ( β = 0.136, z = 2.260, p = 0.024, 95% CI [0.019, 0.183], excluding zero), supporting H4. PE ( β = 0.158, z = 1.398, p = 0.162, 95% CI [− 0.048, 0.312], crossing zero) and SI ( β = 0.090, z = 1.509, p = 0.131, 95% CI [− 0.012, 0.134], crossing zero) were not significant factors influencing BI, and H1 and H3 were not supported. The non-significance of PE, despite its significant association in the post-hoc regression ( β = 0.160, p = 0.002), is likely attributable to the high latent inter-construct correlation between PE and EE ( r = 0.870), which reduces the unique structural contribution of PE when both constructs are modelled simultaneously. Latent inter-construct correlations ranged from 0.478 (SI-BI) to 0.870 (PE-EE) 58 , yet both constructs were retained as theoretically distinct UTAUT constructs. Table 3 presents a summary of the hypothesis testing results. Table 3. Structural equation modelling hypothesis testing results ( n = 700). Hypothesis Construct β z p 95% CI Result H1 PE 0.158 1.398 0.162 [− 0.048, 0.312] Not supported H2 EE 0.351 2.378 0.017* [0.056, 0.518] Supported H3 SI 0.090 1.509 0.131 [− 0.012, 0.134] Not supported H4 FC 0.136 2.260 0.024* [0.019, 0.183] Supported H5 UTAUT model R 2 = 0.441 – < 0.001* [0.387, 0.495] Supported Open in a new tab PE, Performance expectancy; EE, Effort expectancy; SI, Social influence; FC, Facilitating conditions; β , Standardised path coefficient; z , z-Statistic; p , Two-tailed significance value (* p < 0.05 is considered statistically significant); CI, Bias-corrected bootstrapped 95% confidence interval based on 5000 resamples; R 2 , Proportion of variance in behavioural intention explained by the UTAUT model; 95% CI for R 2 calculated using standard approximation formula 45 . Adjusted R 2 is not applicable to the SEM-derived R 2 reported for H5, as SEM computes the squared multiple correlation for the endogenous variable directly; this differs conceptually from the regression-based adjusted R 2 , which corrects for the number of constructs and sample size 47 . Post-hoc sensitivity analysis Prior to hierarchical multiple regression, all standard analytical assumptions were examined. Residual normality was assessed using the Shapiro–Wilk test ( W = 0.970, p < 0.001); the significant result is expected with n = 700 and is not considered indicative of meaningful non-normality 54 . Independence of residuals was supported (Durbin–Watson = 1.713). Linearity was supported by significant positive Pearson correlations between all UTAUT constructs and BI ( r = 0.416–0.572, all p < 0.001), confirmed through partial regression plots. The Breusch-Pagan test indicated statistically significant heteroscedasticity (Lagrange Multiplier, LM = 58.803, p < 0.001), likely attributable to the bounded nature of Likert-based mean scores and the negatively skewed BI distribution 54 . As the regression served as a post-hoc sensitivity analysis, and coefficient estimates remain unbiased under heteroscedasticity in large samples 54 , this violation does not affect the interpretation of findings. No multicollinearity concerns were identified, with all VIF values below 5.0 (range: 1.541–3.352) and tolerance values above 0.20 57 . Twelve cases (1.7%) showed standardised residuals exceeding ± 3.0; however, Cook’s distance exceeded 1.0 for no cases, indicating no influential outliers. The maximum condition index was 21.110, below the threshold of 30 57 . With the exception of homoscedasticity, all remaining assumptions were satisfactorily met. In Block 1, the UTAUT constructs significantly predicted BI ( R 2 = 0.359, Adjusted R 2 = 0.355, F(4,695) = 97.315, p < 0.001, RMSE = 0.681, Cohen’s f 2 = 0.560, large effect 45 ). EE demonstrated the strongest association ( β = 0.331, t = 5.951, p < 0.001, 95% CI [0.217, 0.430]), followed by PE ( β = 0.160, t = 3.163, p = 0.002, 95% CI [0.061, 0.259]) and FC ( β = 0.129, t = 3.432, p < 0.001, 95% CI [0.063, 0.232]). SI was not significant ( β = 0.075, t = 1.882, p = 0.060, 95% CI [− 0.003, 0.120]). To examine the potential influence of demographic variables on BI beyond the UTAUT constructs, they were incorporated as covariates in Block 2. The overall Block 2 model remained significant ( R 2 = 0.365, Adjusted R 2 = 0.356, F(10,689) = 39.570, p < 0.001, RMSE = 0.681, Cohen’s f 2 = 0.574, large effect 45 ). However, the addition of demographic variables did not significantly improve model fit over Block 1 (Δ R 2 = 0.006, Δ F (6,689) = 1.044, p = 0.396). None of the demographic variables significantly predicted BI. Age comparisons of 23 to 27 vs 18 to 22 years ( β = 0.046, t = 0.844, p = 0.399, 95% CI [− 0.062, 0.155]) and 28 years or older vs 18 to 22 years ( β =  − 0.044, t =  − 0.350, p = 0.726, 95% CI [− 0.290, 0.202]), gender ( β = 0.011, t = 0.210, p = 0.833, 95% CI [− 0.093, 0.115]), and academic level comparisons of clinical vs preclinical ( β =  − 0.018, t =  − 0.279, p = 0.781, 95% CI [− 0.142, 0.107]), internship vs preclinical ( β = 0.133, t = 1.641, p = 0.101, 95% CI [− 0.026, 0.293]), and postgraduate vs preclinical ( β = 0.113, t = 1.106, p = 0.269, 95% CI [− 0.087, 0.312]) were all non-significant. These findings indicate that no demographic variable significantly predicted BI beyond the four UTAUT constructs, suggesting that the UTAUT-BI relationships were not substantively confounded by age, gender, or academic level in this sample. Two R 2 estimates are reported in this study. The SEM-derived R 2 of 0.441 indicates the squared multiple correlation for the latent endogenous variable and is the primary indicator of explanatory power. Post-hoc hierarchical regression showed a lower adjusted R 2 of 0.355, as it does not compensate for measurement error and acts on observed composite scores. These two estimates are therefore not comparable. Discussion The UTAUT model implies that all constructs independently and positively predict BI, with these relationships moderated by age, gender, experience, and voluntariness of use. The findings of this study partially support these assumptions. EE and FC constructs influenced BI, multicollinearity with EE weakened PE, and SI was non-significant without institutional mandates. Age and gender did not moderate any of these relationships. High self-directed adoption preparedness is indicated by the mean BI score of 4.10 (95% CI [4.03, 4.16]) and negatively skewed distribution (skewness =  − 0.96). Since LLM-specific proficiency are not formally integrated into Saudi dental curricula and no institutional mandates exist 39 , this pattern supports the UTAUT proposition that perceived personal value drives voluntary technology adoption 10 and the intrinsic motivation constructs of UTAUT2 40 , 43 . EE emerged as the strongest significant factor influencing BI ( β = 0.351), consistent with findings from medical education, healthcare, and university settings where ease of use is reliably associated with stronger adoption intentions 43 , 61 – 65 . EE dominance in voluntary pre-adoption contexts is theoretically expected within UTAUT 10 , as performance beliefs may be more uniformly held due to informal prior exposure while effort perceptions vary more substantially between individuals, generating greater discriminating power for EE. This dominance reflects a specific characteristic of LLMs. LLMs use natural language online interfaces, unlike previous dental AI systems that required expert installation or domain-specific training 29 . This accessibility is supported by meaningful educational value, with LLM accuracy on standardized endodontic patient education questions ranging from 3.83 to 4.81 out of 5.00 15 , prosthodontic exam performance reaching 73–92% 14 , and AI-assisted feedback improving dental students’ radiographic diagnostic performance 16 . However, in image-dependent tasks such as bone age prediction, even the best-performing model achieved only a moderate correlation with reference standards ( r = 0.546) 29 , suggesting that EE dominance may reflect accessibility in straightforward tasks rather than a generalized perception of low effort across all dental use contexts. Maintaining accessible LLM integration pathways may represent a priority within institutional governance frameworks, provided such access does not compromise critical appraisal and clinical accuracy 13 , 33 , 66 . Unlike EE, PE did not reach statistical significance in the primary SEM analysis ( p = 0.162), though its significance in the post-hoc regression ( p = 0.002) indicates that perceived utility remains theoretically meaningful. Non-significance in the structural model is most parsimoniously explained by the high latent inter-construct correlation between PE and EE ( r = 0.870), consistent with their established conceptual proximity within UTAUT 10 and further indicated by the borderline HTMT value (0.886). This high correlation may reflect an LLM-specific characteristic whereby the natural language interface that makes LLMs easy to use is also the interface through which students assess their performance, conflating effort and performance cognitions at the interface level. LLM performance variability between models and dental specialties 8 , 11 , 12 , 14 , 28 may increase between-person PE variance not reflected by construct items. These findings are broadly consistent with the research by Alqaisi et al. 43 who suggested that EE and PE were the most influential factors influencing BI in Jordanian medical students using multiple regression. This SEM method utilized in this study adds methodological subtlety. The high latent inter-construct correlation suppresses the independent contribution of PE in the structural model, which regression-based designs cannot detect. Accordingly, institutional communication focused primarily on LLM performance metrics is unlikely to drive BI in this population, suggesting that motivational and contextual factors warrant greater emphasis in adoption strategies. While EE and PE reflect individual LLM interaction evaluations, FC captures the broader institutional context. FC was a significant factor influencing BI despite having the lowest mean score (mean = 3.47), showing that adoption intentions are partly linked to insufficient institutional infrastructure. FC is more predictive in UTAUT when institutional support is limited 10 , and in the LLM context, FC includes faculty oversight, validated prompting protocols, and governance frameworks, which are lacking in Saudi dental education 39 . Scoping reviews have consistently identified the lack of defined curriculum frameworks and faculty-endorsed guidelines as barriers to appropriate LLM implementation 17 , 18 , while resource-constrained e-learning and dentistry AI adoption have shown similar FC effects 67 , 68 . Converging speciality-specific evidence supports FC. Expert oversight is essential regardless of model performance tier in endodontic patient education 15 , radiographic content generation 16 , clinical anamnesis 23 , orofacial pain management 19 , and prosthodontics 28 , where LLM responses were too low for clinical use without expert validation. No single LLM optimises accuracy, comprehensiveness, and readability across specialities 8 , 31 . These findings suggest that the influence of FC reflects identified governance needs rather than discretionary institutional practice, underscoring the importance of structured support frameworks for responsible LLM integration. In contrast to EE and FC, SI was a non-significant factor influencing BI ( β = 0.090, p = 0.131). The mean SI score of 3.76 indicates moderate perceived social influence, however SI did not influence BI. Within UTAUT, SI predominantly drives adoption in mandatory contexts 10 , and its non-significance here is logically coherent given the lack of institutional LLM mandates or formal policies in Saudi dental curricula 35 , 39 . This finding is noteworthy in Saudi Arabia’s collectivist culture, where SI would theoretically be expected to play a stronger role in BI. One plausible explanation is that LLM use is conducted privately and individually, reducing the social visibility that typically activates peer and authority influence mechanisms in collectivist environments. The lack of faculty awareness and institutional endorsement of LLMs 66 suggests that the primary SI channel within the institutional context remains insufficiently operationalised. Evidence from mandatory adoption contexts documents significant SI effects among medical and university students 62 , 65 , and Liu et al. 62 proposed SI to be a significant factor influencing BI among clinical medicine students in a structured institutional context, contrasting directly with the present findings and supporting a context-specific interpretation. However, SI may emerge as a significant predictor after formal LLM integration into the dental curriculum, although peer-led and authority-driven adoption strategies are unlikely to serve as effective drivers within the prevailing voluntary adoption context. The SEM-derived R 2 of 0.441 surpasses the pre-specified planning estimate ( R 2 = 0.15) but remains below the ideal UTAUT specification benchmark of 69% 10 . Prior LLM experience is a significant unmeasured covariate, given documented inter-model performance variation across dental specialities 31 . Students using high-performing models may have different PE and EE beliefs than those who are using lower-performing ones. Additional constructs absent from the original UTAUT model, including, hedonic motivation, habit, digital literacy, ethical concerns 66 , 69 , or openness to technology 70 , may further explain BI variance in this cohort. The hypothesis support for H2 and H4, in the absence of significant effects for all four constructs, partially diverges from the original UTAUT model, which predicts independent positive contributions from PE, EE, SI, and FC to BI. In this voluntary adoption context, EE and FC were effective, while PE was attenuated by multicollinearity and SI was ineffective without institutional mandates. The findings have practical implications across three areas, namely curriculum design informed by the dominance of EE, institutional governance informed by the significance of FC despite its low mean score, and national policy situated within Saudi Arabia’s healthcare digitalisation mission. The FC findings may have direct governance implications. FC was a significant factor influencing BI, but its low mean score (3.47) suggests that institutional circumstances for LLM adoption are lacking. This gap warrants attention at multiple levels. For institutional leaders, the FC deficit suggests that foundational infrastructure, such as institutional access to validated LLM platforms, faculty training programs for LLM oversight, and clear usage policies, may be a governance priority, though cross-sectional data cannot establish causal direction 71 . For faculty, the dominance of EE suggests that structured, low-effort LLM access points like faculty-endorsed prompting guides and supervised comparison exercises between LLM outputs and authoritative reference materials may be the best reference point for faculty to integrate LLMs into teaching practice 13 , 31 , 66 . For curriculum designers, the lack of a correlation between academic level and BI suggests that LLM proficiency activities may be introduced from early curriculum stages rather than being reserved for advanced clinical training, but confirmation in larger samples is required before informing curriculum policy. Institutional policy development may benefit from alignment with the Saudi Commission for Health Specialties and the National Commission for Academic Accreditation and Assessment, which neither specifies LLM-specific proficiency for dental programs currently. This represents a regulatory gap that institutional policymakers may consider addressing proactively. How these considerations translate into curriculum policy, institutional governance, or national policy alignment remains a question for implementation research rather than adoption surveys. Translating these institutional recommendations into an educational structure requires a coherent competency framework. Based on these findings, we tentatively propose a conceptual LLM competency framework with three domains, namely knowledge, skills, and values. This framework needs empirical validation before implementation because it is theoretically informed and claims no originality, as it is based on broader literature. Li and Qin 42 proposed curriculum recommendations for AI integration in medical education comprising of knowledge, attitude, and skill domains, while Rincón et al. 72 mapped AI competency requirements in medical education through a scoping review encompassing similar dimensions. In dental education, scoping and systematic reviews have identified knowledge, application, and ethical awareness as recurring competency themes 17 , 18 , and faculty-level studies have highlighted awareness, knowledge, and attitudes as the main development areas 66 . However, existing frameworks have not adequately addressed LLM-specific characteristics, including hallucination risks, inter-model performance variability, and the accuracy-readability trade-off, or the institutional governance gap identified in the present study. Accordingly, the proposed LLM competency framework is grounded in the knowledge, skills, and values domain structure, adapted to address the competency requirements of responsible LLM integration within dental education. In the knowledge domain, the FC gap identified in this sample suggests that students currently lack sufficient institutional support for LLM engagement. Addressing this gap may begin with developing students’ understanding of LLM capabilities and limitations, including hallucination risks 24 , 34 , alongside structured prompting proficiency, which has been shown to improve LLM accuracy in specialised contexts 13 . In the skills domain, the dominance of EE suggests that activities reducing perceived effort while maintaining clinical rigour are likely to support adoption. Case-based learning requiring critical appraisal of LLM-generated clinical inferences against evidence-based dental guidelines may enhance clinical judgement, without conflating LLM-generated inferences with clinical competence 73 – 75 . LLM performance varies across selected dental specialities, from passing-level accuracy in dental licensing examinations 11 , 12 to limited and variable accuracy across prosthodontic examination questions, particularly in removable partial dentures 14 , 28 . Structured exposure to this performance range may help students develop realistic expectations of LLM capabilities. In the values domain, ethical guidelines emphasising the distinction between informational support and clinical decision authority, alongside reflective engagement with the governance implications identified in this sample, may cultivate responsible professional LLM utilization 73 , 76 . Unlike the knowledge and skills domains, the values domain is proposed based on professional ethics principles and governance literature rather than empirical evidence and represents the most speculative of the three domains. The LLM competency framework represent conceptual directions for future empirical validation through longitudinal and intervention-based research before translation into institutional practice. This study may have implications within the context of Vision 2030, particularly the Health Sector Transformation Program’s commitment to leveraging AI in healthcare and the Human Capability Development Program’s priority of aligning higher education outcomes with future digital workforce needs 77 , 78 . These implications are grounded in adoption intention data and should be interpreted as conceptual directions rather than policy recommendations. Several limitations constrain the interpretation and generalisability of these findings. The cross-sectional design precludes causal inference between UTAUT constructs and BI, as all associations are concurrent and bidirectional causation cannot be excluded. A further limitation concerns the construct validity of BI as an outcome measure, being a theoretical proxy for technology adoption and measures motivational readiness at a single time point 10 . Convenience sampling and open-access social media distribution may have introduced selection bias by over-representing digitally engaged students 53 , potentially limiting the generalisability of the observed BI ratings and inter-construct associations. Thus, the observed BI scores and construct connections may not fully reflect the opinions of less digitally engaged students. As the survey was openly distributed via social media over 2 months without a defined sampling frame, the total number of potential respondents could not be ascertained, precluding formal non-response bias assessment. The sample coverage does not represent all dental institutions in central Saudi Arabia, and conclusions may not apply to unsampled institutions, other regions or international dental education contexts. Prior LLM experience and voluntariness of use were not recorded, preventing a comprehensive study of all theoretically required UTAUT moderators. Voluntariness moderates the SI-BI relationship in non-mandatory adoption, making it important. Through familiarity and perceived competence effects, prior experience with specific LLM tools may also influence BI, although its relative contribution to unexplained variance in this population is unknown. Cross-sectional single-source designs still have CMB issues. Although procedural anonymity restrictions were adopted and Harman’s single-factor test showed no severe CMB, the common method factor model was non-estimable, preventing statistical ruling out of CMB. Temporal separation and marker variable techniques were design-level limitations that could not be remediated within the constraints of this study. Despite anonymity assurances and procedural measures to minimise social desirability bias, the intention-behaviour gap inherent in self-reported BI measurement may have inflated BI ratings relative to actual LLM adoption behaviour. Future research should prioritise several methodological improvements. Longitudinal designs tracking both BI and actual LLM adoption behaviour over time would enable the first direct examination of the intention-to-behaviour gap in this population, addressing a recognised empirical void in LLM adoption research. Probability-based or stratified sampling across all dental institutions in Saudi Arabia and in international contexts would improve representativeness and permit formal assessment of non-response bias. Future studies should systematically collect prior LLM experience data, including which models were used, for which tasks, and with what perceived accuracy outcomes, to more fully operationalise the UTAUT moderators and to test whether differential experience with specific LLM tools and their perceived task accuracy accounts for unexplained variance in BI. To strengthen causal inference and permit more rigorous CMB evaluation, future studies should consider temporal separation of construct and outcome measurements across two survey administrations, alongside prospective inclusion of marker variables. Incorporation of validated social desirability scales, in future survey instruments would permit statistical control of response bias and strengthen the interpretability of self-reported BI scores. Extended UTAUT2 or hybrid technology acceptance models incorporating digital literacy, perceived risk, ethical concerns, and intrinsic motivation may achieve greater explanatory power in LLM-specific adoption contexts 63 , 68 , 69 . Comparative studies across dental specialities, including prosthodontics, orthodontics, and endodontics, where LLM performance profiles differ substantially 14 , 23 , 28 , 30 , would allow assessment of whether the observed construct relationships generalise across clinical training contexts. Finally, intervention studies testing structured LLM integration strategies, such as expert-supervised prompting laboratories, case-based learning modules, and faculty-endorsed model selection frameworks, would provide empirically grounded evidence for translating expressed intention into structured educational practice, moving the field beyond BI as a proxy measure toward confirmed behavioural and learning outcomes. Conclusion The UTAUT model was supported overall, accounting for 44.1% of BI variance in this sample. Among the four constructs, EE was the strongest significant factor influencing BI, indicating that perceived ease of use was the most influential factor affecting LLM adoption in this voluntary, unstructured and unregulated context. FC was also significant despite its low mean score, indicating that adoption is linked to perceived institutional support even when it is insufficient. The PE was non-significant, due to its high latent correlation with EE, and SI was also non-significant, consistent with absent institutional mandates. The results indicate strong expressed BI to adopt LLMs among dental students; however, BI is a theoretical proxy and should not be equated with actual adoption behaviour, clinical competence, or ethical and effective LLM application. These findings offer preliminary evidence which may inform curriculum design and institutional governance for responsible LLM integration in dentistry. The tentatively proposed LLM competency framework represents theoretical concepts which require empirical validation before implementation. The implications are grounded in adoption intention data and should be interpreted as conceptual directions rather than policy recommendations. Within the limitations of this study, dental schools in central Saudi Arabia and similar contexts may consider these findings to create responsible, critically informed, and ethically governed LLM integration guidelines for dental education, CDS, and research within Saudi Arabia’s Vision 2030 healthcare digitalisation mission. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 (26.6KB, docx) Acknowledgements The authors would like to thank the participants for their valuable support in this study. A.M.J. has drawn Fig. 1 in this manuscript. Author contributions A.M.J. conceptualized the study, conducted the formal analysis and investigation, developed the methodology, and wrote the original draft of the manuscript. R.E.A. and W.A.A. performed data curation and investigation, contributed to the methodology, and reviewed and edited the manuscript. R.M.A., Y.D., S.A., and A.A. were responsible for validation and visualization. M.Z.M., A.A.T., S.J., and A.J. contributed software support and reviewed and edited the manuscript. S.K.M. provided supervision and resources. Funding The authors extend the appreciation to the Deanship of Postgraduate Studies and Scientific Research at Majmaah University, Al-Majmaah 11952, Saudi Arabia for funding this research work through the Project Number (R-2026-140). Data availability The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1. Tuygunov, N. et al. The transformative role of artificial intelligence in dentistry: A comprehensive overview Part 2: The promise and perils, and the International Dental Federation communique. Int. Dent. J. 75 , 397–404. 10.1016/j.identj.2025.02.006 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 2. Bonny, T. et al. Contemporary role and applications of artificial intelligence in dentistry. F1000Res 12 , 1179. 10.12688/f1000research.140204.1 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 3. Najeeb, M. & Islam, S. Artificial intelligence (AI) in restorative dentistry: Current trends and future prospects. BMC Oral Health 25 , 592. 10.1186/s12903-025-05989-1 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 4. Gao, S. et al. Artificial intelligence in dentistry: A narrative review of diagnostic and therapeutic applications. Med. Sci. Monit. 31 , e946676. 10.12659/MSM.946676 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. Schwendicke, F., Samek, W. & Krois, J. Artificial intelligence in dentistry: Chances and challenges. J. Dent. Res. 99 , 769–774. 10.1177/0022034520915714 (2020). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 6. Umer, F., Batool, I. & Naved, N. Innovation and application of large language models (LLMs) in dentistry: A scoping review. BDJ Open 10 , 90. 10.1038/s41405-024-00277-6 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 7. Puleio, F., Lo Giudice, G., Bellocchio, A. M., Boschetti, C. E. & Lo Giudice, R. Clinical, research, and educational applications of ChatGPT in dentistry: A narrative review. Appl. Sci. 14 , 10802. 10.3390/app142310802 (2024). [ Google Scholar ] 8. Wu, X. et al. A multi-dimensional performance evaluation of large language models in dental implantology: Comparison of ChatGPT, DeepSeek, Grok, Gemini and Qwen across diverse clinical scenarios. BMC Oral Health 25 , 1272. 10.1186/s12903-025-06619-6 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Farhadi Nia, M., Ahmadi, M. & Irankhah, E. Transforming dental diagnostics with artificial intelligence: Advanced integration of ChatGPT and large language models for patient care. Front. Dent. Med. 5 , 1456208. 10.3389/fdmed.2024.1456208 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 10. Venkatesh, V., Morris, M. G., Davis, G. B. & Davis, F. D. User acceptance of information technology: Toward a unified view. MIS Q. 27 , 425–478 (2003). [ Google Scholar ] 11. Liu, M. et al. Large language models in dental licensing examinations: Systematic review and meta-analysis. Int. Dent. J. 75 , 213–222. 10.1016/j.identj.2024.10.014 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 12. Chau, R. C. W. et al. Performance of generative artificial intelligence in dental licensing examinations. Int. Dent. J. 74 , 616–621. 10.1016/j.identj.2023.12.007 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. Dashti, M. et al. Assessing ChatGPT-4’s performance on the US prosthodontic exam: Impact of fine-tuning and contextual prompting vs. base knowledge, a cross-sectional study. BMC Med. Educ. 25 , 761. 10.1186/s12909-025-07371-9 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 14. Eraslan, R., Ayata, M., Yagci, F. & Albayrak, H. Exploring the potential of artificial intelligence chatbots in prosthodontics education. BMC Med. Educ. 25 , 321. 10.1186/s12909-025-06849-w (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 15. Demir Cicek, B. & Cicek, O. Evaluating the response of AI-based large language models to common patient concerns about endodontic root canal treatment: A comparative performance analysis. J. Clin. Med. 14 , 7482. 10.3390/jcm14217482 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 16. Gokkurt Yilmaz, B. N., Ozbey, F. & Yilmaz, B. E. Effect of artificial intelligence-assisted personalized feedback on radiographic diagnostic performance of dental students: A controlled study. BMC Med. Educ. 25 , 1403. 10.1186/s12909-025-07875-4 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. El-Hakim, M., Anthonappa, R. & Fawzy, A. Artificial intelligence in dental education: A scoping review of applications, challenges, and gaps. Dent. J. 13 , 384. 10.3390/dj13090384 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 18. Aura-Tormos, J. I., Llacer-Martinez, M. & Torres-Osca, I. Educational applications of ChatGPT in university-based dental education: A systematic review. Eur. J. Dent. Educ. 29 , 1–17. 10.1111/eje.70011 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 19. Vueghs, C., Shakeri, H., Renton, T. & Van der Cruyssen, F. Development and evaluation of a GPT-4-based orofacial pain clinical decision support system. Diagnostics 14 , 2835. 10.3390/diagnostics14242835 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 20. Samaranayake, L. et al. The transformative role of artificial intelligence in dentistry: A comprehensive overview. Part 1: Fundamentals of AI, and its contemporary applications in dentistry. Int. Dent. J. 75 , 383–396. 10.1016/j.identj.2025.02.005 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 21. Karuppan Perumal, M. K., Rajan Renuka, R., Kumar Subbiah, S. & Manickam Natarajan, P. Artificial intelligence-driven clinical decision support systems for early detection and precision therapy in oral cancer: A mini review. Front. Oral Health 6 , 1592428. 10.3389/froh.2025.1592428 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 22. Salvi, S., Vu, G., Gurupur, V. & King, C. Digital convergence in dental informatics: A structured narrative review of artificial intelligence, Internet of Things, digital twins, and large language models with security, privacy, and ethical perspectives. Electronics 14 , 3278. 10.3390/electronics14163278 (2025). [ Google Scholar ] 23. Yilmaz, B. E., Ozbey, F., Gokkurt Yilmaz, B. N. & Akpinar, H. Can large language models perform clinical anamnesis? Comparative evaluation of ChatGPT, Claude, and Gemini in diagnostic reasoning through case-based questioning in oral and maxillofacial disorders. J. Stomatol. Oral Maxillofac. Surg. 127 , 102644. 10.1016/j.jormas.2025.102644 (2026). [ DOI ] [ PubMed ] [ Google Scholar ] 24. Kong, M., Fok, E. H. W. & Yiu, C. K. Y. A scoping review of large language models in dental education: Applications, challenges, and prospects. Int. Dent. J. 75 , 103854. 10.1016/j.identj.2025.103854 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 25. Hajibagheri, P., Khosousi Sani, S., Samami, M., Tabari-Khomeiran, R. & Azadpeyma, K. ChatGPT’s accuracy in the diagnosis of oral lesions. BMC Oral Health 25 , 1229. 10.1186/s12903-025-06582-2 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 26. Suárez, A. et al. Unveiling the ChatGPT phenomenon: Evaluating the consistency and accuracy of endodontic question answers. Int. Endod. J. 57 , 108–113. 10.1111/iej.13985 (2024). [ DOI ] [ PubMed ] [ Google Scholar ] 27. Tastan Eroglu, Z., Babayigit, O., Ozkan Sen, D. & Ucan Yarkac, F. Performance of ChatGPT in classifying periodontitis according to the 2018 classification of periodontal diseases. Clin. Oral Investig. 28 , 407. 10.1007/s00784-024-05799-9 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 28. Freire, Y. et al. ChatGPT performance in prosthodontics: Assessment of accuracy and repeatability in answer generation. J. Prosthet. Dent. 131 , 659.e1-659.e6. 10.1016/j.prosdent.2024.01.018 (2024). [ DOI ] [ PubMed ] [ Google Scholar ] 29. Yıldırım, A., Cicek, O. & Genç, Y. S. Can AI-based ChatGPT models accurately analyze hand-wrist radiographs? A comparative study. Diagnostics 15 , 1513. 10.3390/diagnostics15121513 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 30. Yıldırım, A. & Cicek, O. Assessment of AI-driven large language models for orthodontic aesthetic scoring using the IOTN-AC. Diagnostics 15 , 3048. 10.3390/diagnostics15233048 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 31. Cicek, O. & Yıldırım, A. Evaluating the reliability of artificial intelligence in clinical decisions on early maxillary expansion: A comparative study of eight large language models. BMC Oral Health. 10.1186/s12903-026-08111-1 (2026). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 32. Zhou, H. et al. Large language models and machine learning framework for predicting dental ceramics performance. Int. Dent. J. 76 , 109358. 10.1016/j.identj.2025.109358 (2026). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 33. Colangelo, M. T., Guizzardi, S., Meleti, M., Calciolari, E. & Galli, C. Performance comparison of large language models for efficient literature screening. BioMedInformatics 5 , 25. 10.3390/biomedinformatics5020025 (2025). [ Google Scholar ] 34. Schwendicke, F., Sidhu, S. K., Ferracane, J. L., Tichy, A. & Jakubovics, N. S. Generative AI: Opportunities, risks, and responsibilities for oral sciences. J. Dent. Res. 104 , 1429–1431. 10.1177/00220345251356408 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 35. Al-Khalifa, K. S. et al. Evaluating dental students’ perspectives on artificial intelligence (AI)-driven large language models in education in Saudi Arabia. Eur. J. Dent. Educ. 10.1111/eje.70042 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 36. Saudi Data and Artificial Intelligence Authority. Generative artificial intelligence guidelines for government. SDAIA. https://sdaia.gov.sa/en/SDAIA/about/Files/GenAIGuidelinesForGovernmentENCompressed.pdf (2024). 37. Saudi Data and Artificial Intelligence Authority. Generative artificial intelligence guidelines for the public. SDAIA. https://sdaia.gov.sa/en/SDAIA/about/Files/GenerativeAIPublicEN.pdf (2024). 38. Saudi Data and Artificial Intelligence Authority. SDAIA launches national cross-disciplinary curriculum for university students in data and AI. Saudi Press Agency. https://spa.gov.sa/en/N2501273 (2026). 39. Alfallaj, H. A. et al. The status of digital dental technology implementation in the Saudi dental schools curriculum: a national cross-sectional survey for healthcare digitisation. Int. J. Environ. Res. Public Health 20 , 321. 10.3390/ijerph20010321 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 40. Tao, W., Yang, J. & Qu, X. Utilisation of, perceptions on, and intention to use AI chatbots among medical students in China: National cross-sectional study. JMIR Med. Educ. 10 , e57132. 10.2196/57132 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 41. Breitwieser, M. et al. AI in fracture detection: A cross-disciplinary analysis of physician acceptance using the UTAUT model. Diagnostics 15 , 2117. 10.3390/diagnostics15162117 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 42. Li, Q. & Qin, Y. AI in medical education: Medical student perception, curriculum recommendations and design suggestions. BMC. Med. Educ. 23 , 852. 10.1186/s12909-023-04700-8 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 43. Alqaisi, N. et al. Assessing ChatGPT adoption in Jordanian medical education: A UTAUT model approach. BMC. Med. Educ. 25 , 750. 10.1186/s12909-025-07336-y (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 44. Joseph, A. M. et al. Predicting the employment of teledentistry in clinical practice by the Saudi dental community using a theoretical model. Digit. Health. 10 , 20552076241253740. 10.1177/20552076241253739 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 45. Cohen, J. Statistical Power Analysis for the Behavioral Sciences 2nd edn. (Lawrence Erlbaum Associates, 1988). [ Google Scholar ] 46. Wolf, E. J., Harrington, K. M., Clark, S. L. & Miller, M. W. Sample size requirements for structural equation models: An evaluation of power, bias, and solution propriety. Educ. Psychol. Meas. 10.1177/0013164413495237 (2013). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 47. Kline, R. B. Principles and Practice of Structural Equation Modeling 4th edn. (Guilford Press, 2015). [ Google Scholar ] 48. Lynn, M. R. Determination and quantification of content validity. Nurs. Res. 35 , 382–385. 10.1097/00006199-198611000-00017 (1986). [ PubMed ] [ Google Scholar ] 49. Polit, D. F. & Beck, C. T. The content validity index: Are you sure you know what’s being reported? Critique and recommendations. Res. Nurs. Health. 29 , 489–497. 10.1002/nur.20147 (2006). [ DOI ] [ PubMed ] [ Google Scholar ] 50. Taber, K. S. The use of Cronbach’s alpha when developing and reporting research instruments in science education. Res. Sci. Educ. 48 , 1273–1296. 10.1007/s11165-016-9602-2 (2018). [ Google Scholar ] 51. von Elm, E. et al. The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: Guidelines for reporting observational studies. Ann. Intern. Med. 147 , 573–577. 10.7326/0003-4819-147-8-200710160-00010 (2007). [ DOI ] [ PubMed ] [ Google Scholar ] 52. Almalki, M., Alkhamis, M. A., Khairallah, F. M. & Choukou, M. A. Perceived artificial intelligence readiness in medical and health sciences education: A survey study of students in Saudi Arabia. BMC. Med. Educ. 25 , 439. 10.1186/s12909-025-06995-1 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 53. Eysenbach, G. Improving the quality of web surveys: The Checklist for Reporting Results of Internet E-Surveys (CHERRIES). J. Med. Internet Res. 6 , e34. 10.2196/jmir.6.3.e34 (2004). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 54. Field, A. Discovering Statistics Using IBM SPSS Statistics 5th edn. (SAGE Publications, 2018). [ Google Scholar ] 55. Cheung, G. W. & Rensvold, R. B. Evaluating goodness-of-fit indexes for testing measurement invariance. Struct. Equ. Model. 9 , 233–255. 10.1207/S15328007SEM0902_5 (2002). [ Google Scholar ] 56. Podsakoff, P. M., MacKenzie, S. B., Lee, J. Y. & Podsakoff, N. P. Common method biases in behavioral research: A critical review of the literature and recommended remedies. J. Appl. Psychol. 88 , 879–903. 10.1037/0021-9010.88.5.879 (2003). [ DOI ] [ PubMed ] [ Google Scholar ] 57. Hair, J. F., Black, W. C., Babin, B. J. & Anderson, R. E. Multivariate Data Analysis 8th edn. (Cengage Learning, 2019). [ Google Scholar ] 58. Henseler, J., Ringle, C. M. & Sarstedt, M. A new criterion for assessing discriminant validity in variance-based structural equation modeling. J. Acad. Mark. Sci. 43 , 115–135. 10.1007/s11747-014-0403-8 (2015). [ Google Scholar ] 59. Alwadi, M. A. M., Alosail, L. A., Alroweilly, D. A., AlJubair, A. M. & Talakey, A. A. Evaluating the educational environment among Saudi board dental residents at King Abdulaziz Medical City in Riyadh using PHEEM tool. BMC Med. Educ. 24 , 1006. 10.1186/s12909-024-06001-0 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 60. Chen, F. F. Sensitivity of goodness of fit indexes to lack of measurement invariance. Struct. Equ. Model. 14 , 464–504. 10.1080/10705510701301834 (2007). [ Google Scholar ] 61. Zhao, X. et al. Factors influencing medical students’ adoption of AI educational agents: An extended UTAUT model. BMC Med. Educ. 25 , 1678. 10.1186/s12909-025-08234-z (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 62. Liu, F. et al. Assessing clinical medicine students’ acceptance of large language model: Based on technology acceptance model. BMC Med. Educ. 24 , 1251. 10.1186/s12909-024-06232-1 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 63. Acosta-Enriquez, B. G. et al. Acceptance of artificial intelligence in university contexts: A conceptual analysis based on UTAUT2 theory. Heliyon 10 , e38315. 10.1016/j.heliyon.2024.e38315 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 64. Swathi, K. S., Aswathy, S., Kavitha, T. C., Chadaga, K. & Sampathila, N. Behavioural intentions to adopt artificial intelligence in healthcare: Exploring the perception of healthcare professionals. J. Technol. Behav. Sci. 10.1007/s41347-025-00483-5 (2025). [ Google Scholar ] 65. Dai, Q. et al. Structural equation modeling for influencing factors on behavioral intention to adopt medical AI among Chinese nurses: A nationwide cross-sectional study. BMC Nurs. 24 , 1084. 10.1186/s12912-025-03748-9 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 66. Abdullah, S., Hasan, S. R., Asim, M. A., Khurshid, A. & Qureshi, A. W. Exploring dental faculty awareness, knowledge, and attitudes toward AI integration in education and practice: A mixed-method study. BMC Med. Educ. 25 , 691. 10.1186/s12909-025-07259-8 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 67. El-Masri, M. & Tarhini, A. Factors affecting the adoption of e-learning systems in Qatar and USA: Extending the Unified Theory of Acceptance and Use of Technology 2 (UTAUT2). Educ. Technol. Res. Dev. 65 , 743–763. 10.1007/s11423-016-9508-8 (2017). [ Google Scholar ] 68. AlQaifi, F. & Tengilimoglu, D. Factors affecting dentists’ intention to adopt artificial intelligence: An extension of the Unified Theory of Acceptance and Use of Technology (UTAUT) model. J. Health Organ. Manag. 39 , 1–18. 10.1108/JHOM-03-2025-0159 (2025). [ DOI ] [ PubMed ] [ Google Scholar ] 69. Sarhan, S., Badran, A., Ghalwash, D., Gamal Almalahy, H. & Abou-Bakr, A. Perception, usage, and concerns of artificial intelligence applications among postgraduate dental students: Cross-sectional study. BMC Med. Educ. 25 , 856. 10.1186/s12909-025-07544-6 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 70. BinJwair, A. & Alamer, A. Digital drivers: How the acceptance and use of technology model shapes university students’ behaviour toward generative AI. Acta Psychol. (Amst.) 261 , 105930. 10.1016/j.actpsy.2025.105930 (2025). [ DOI ] [ PubMed ] [ Google Scholar ] 71. Leung, A., Fine, P., Tonni, I. & Louca, C. Artificial intelligence: A modern tool for assessing dental students?. J. Dent. 165 , 106273. 10.1016/j.jdent.2025.106273 (2026). [ DOI ] [ PubMed ] [ Google Scholar ] 72. Rincón, E. H. H. et al. Mapping the use of artificial intelligence in medical education: A scoping review. BMC Med. Educ. 25 , 526. 10.1186/s12909-025-07089-8 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 73. AlFarabi, A. S. et al. The diagnostic performance of large language models and oral medicine consultants for identifying oral lesions in text-based clinical scenarios: Prospective comparative study. JMIR AI 4 , e70566. 10.2196/70566 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 74. Tokgöz Kaplan, T. & Cankar, M. Evidence-based potential of generative artificial intelligence large language models on dental avulsion: ChatGPT versus Gemini. Dent. Traumatol. 41 , 178–186. 10.1111/edt.12999 (2025). [ DOI ] [ PubMed ] [ Google Scholar ] 75. Dermata, A., Arhakis, A., Makrygiannakis, A., Giannakopoulos, K. & Kaklamanos, E. G. Evaluating the evidence-based potential of six large language models in paediatric dentistry: A comparative study on generative artificial intelligence. Eur. Arch. Paediatr. Dent. 26 , 527–535. 10.1007/s40368-025-01012-x (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 76. AlGoraini, Y. et al. Perceptions of large language models in medical education and clinical practice among pediatric emergency physicians in Saudi Arabia: A multiregional cross-sectional study. Front. Public Health 13 , 1634638. 10.3389/fpubh.2025.1634638 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 77. Vision 2030. Health Sector Transformation Program. https://www.vision2030.gov.sa/en/explore/programs/health-sector-transformation-program (2016). 78. Vision 2030. Human Capability Development Program 2021–2025. https://www.vision2030.gov.sa/en/explore/programs/human-capability-development-program (2021). Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supplementary Material 1 (26.6KB, docx) Data Availability Statement The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request. Articles from Scientific Reports are provided here courtesy of Nature Publishing Group ACTIONS View on publisher site PDF (1.8 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Large language model use in dental education: a cross-sectional multi-country study.#439247
• Large language model use in dental education: a cross-sectional multi-country study.#437398
• Prompt Design at Scale: How Format, Instruction Count, and Context Length Shape Instruction Adherence and Hallucination in Large Language Models#386910
• Exploring the attitudes, perceptions, and experiences of Swiss dental students in adopting simulated clinical teaching with haptic simulators: a qualitative study.#654191
• Preferences of Filipino and Foreign College Students Towards Online Translation Tools#87813