1. INTRODUCTION

Recently, artificial intelligence (AI) in education has revolutionised teaching techniques and opened new possibilities for individualised and adaptable learning (Alarefi, 2024). Educational institutions are increasingly using AI-driven solutions such as intelligent tutoring systems, automated grading software, and data-informed learning analytics to boost student engagement, instructional effectiveness, and learning pathways (Cukurova, 2025). AI’s potential benefits in education are well known, but the factors determining students’ acceptance of these technologies, and their academic and cognitive outcomes are examined in limited studies. Understanding these elements is crucial to optimising AI in education while reducing risks like overreliance on automation and ethical problems including data privacy and algorithmic prejudice (Samala et al., 2024).

Previous well established models and theories such as The technology acceptance model (TAM) by Davis (1989) and the unified theory of acceptance and use of technology (UTAUT) by Venkatesh et al. (2003) show that perceived usefulness (PU), ease of use, and social influence strongly influence students’ acceptance of digital learning tools. AI has drawbacks, including algorithmic decision-making, data security concerns, and potential consequences on critical thinking (Mouta et al., 2024). Over-reliance on AI-driven suggestions may reduce students’ problem-solving autonomy and creativity (Roll & Wylie, 2016), although other research show that real-time feedback and personalised information might increase learning efficiency (Chen et al., 2020). AI implementation’s psychological and social effects on student motivation, classroom dynamics, and teacher-student relationships require further study.

Despite growing global interest in AI in education, most of the research has focused on Western and East Asian contexts, leaving many cultural and educational systems unprepared for new technologies (Assaf et al., 2024). This study fills this gap by examining student AI use and its implications in Saudi Arabia, which is undergoing rapid digital change as part of Vision 2030. As part of Vision 2030, Saudi Arabia is investing in AI and digital education to diversify the economy and build a knowledge-based economy (Alarefi, 2023a). The Saudi Ministry of Education has implemented intelligent learning platforms, AI-driven tutoring systems, and digital literacy programs (Alarefi, 2023b). These programs’ success depends on students’ willingness to use AI technologies and their impact on educational performance and cognitive skill development.

In the context of Saudi Arabia, several factors can contribute to the usage of AI, and these based on TAM and UTAUT can be related to the usefulness of the usage as well as the ease of use along with the influence of others, such as other students, teachers, and schools, as well as the community. Students’ self-efficacy is a critical factor that can determine the usage (Ali & Allawi, 2024). However, there are limited studies that have examined the AI usage by students. In Saudi Arabia, despite strong Internet penetration and a tech-savvy youth (Alanzi et al., 2023), legacy educational methods and reluctance to change may limit AI-driven usage for educational purposes (Ghanem & Alshahrani, 2021). Ethical concerns regarding data privacy and AI decision-making may also affect students’ trust in these technologies (AlMulhim, 2023). Understanding these contextual factors is essential for creating AI-driven educational tools that match local values and learning styles.

This research provides Saudi Arabian policymakers and educators with regional views on AI in education to improve the global discussion. This study will identify Saudi students’ AI adoption motivations and assess its academic and cognitive effects, offering evidence-based recommendations for improving AI integration in Saudi classrooms. The findings contribute to ethical and pedagogical discussions about AI in education, ensuring that technology innovation supports sustainable and fair learning. The next sections discuss the literature review followed by the research methodology, findings, discussion, and conclusion.

2. LITERATURE REVIEW

3. RESEARCH METHODOLOGY

This quantitative study employed a cross-sectional survey design to examine the predictors and consequences of AI tool adoption among university students in Saudi Arabia. The target population comprised undergraduate students enrolled in public universities across Saudi Arabia. A multi-stage sampling approach was implemented to ensure representative geographical coverage and disciplinary diversity. First, four universities were purposively selected from different regions. These institutions were chosen based on their regional distribution, student population size, and established e-learning infrastructure. Within each university, stratified random sampling was employed to select participants proportionally across academic disciplines (humanities, health sciences, and business) and year of study (1st to 4th year). The sample size was determined using Cochran’s formula for finite populations, targeting a 95% confidence level with a 5% margin of error. This approach yielded a required sample of 1,200 participants, with 300 students recruited from each university.

The study utilized a structured questionnaire developed through an extensive four-phase process. The questionnaire items were adapted from validated instruments in technology acceptance literature while incorporating AI-specific constructs. The survey comprised six sections. The first is the demographic information and it includes three questions regarding age, gender, and major. The second section include the variables of the study. PU was measured using six items, and PEOU was measured using five items adopted from Davis (1989). Social influence includes four items and was adopted from Venkatesh et al. (2003). For trust in AI, it was adopted from Lian (2015) while measurement of ethical concern was adopted from Bond and Bedenlier (2019). Further, the adoption of AI was adopted from Venkatesh et al. (2003). Measurement of self-efficacy was adopted from Alam (2021) while measurement of academic performance was adopted from Chen et al. (2022) and measurement of cognitive skill development (five items) was adopted from Kayali and Alaaraj (2020).

The second stage included the validation of the instrument. The initial English version was reviewed by a panel of five experts. The content validity index was calculated, with all scales exceeding the recommended threshold of 0.80. Experts provided feedback on item clarity, relevance, and cultural appropriateness. The third stage included translation of the questionnaire. The questionnaire underwent rigorous translation into Arabic following the forward-backward translation protocol. Two bilingual academics independently translated the instrument, followed by reconciliation of discrepancies. The fourth state was the pilot testing. The Arabic questionnaire was pilot tested with 40 students (10 from each university). Reliability analysis showed excellent internal consistency (CA ranging from 0.83 to 0.92 across scales). Minor adjustments were made to improve item clarity based on pilot feedback.

Data collection occurred over an 8-week period (between November 2024 and January 2025) through an online survey platform. The researcher coordinated with university administrators to distribute the survey through official channels. A link was sent to the students, and they were reminded weekly. The final sample included 357 completed responses, representing a 29.7% response rate. The sample characteristics mirrored the target population in terms of gender distribution, academic level, and discipline representation.

A comprehensive data cleaning protocol was implemented. Cases with more than 20% missing items (n=18) were excluded. For remaining missing values, multiple imputation using chained equations was employed. Little’s missing completely at random (MCAR) test indicated the data was MCAR (χ²=32.18, p=0.14). The outliers was examined, and this resulted in removing 22 responses. Multivariate outliers were identified using Mahalanobis distance (p<0.001). After verifying these were legitimate responses, extreme values were winsorized to the 95th percentile. All variables showed acceptable skewness (<|1.0|) and kurtosis (<|1.0|). Variance inflation factors ranged from 1.12 to 2.87, well below the threshold of 5, indicating no concerning multicollinearity among predictors. Comparison of early (n=192) and late respondents (n=125) revealed no significant differences in demographic characteristics or scale scores (all p>0.05), suggesting minimal non-response bias. In total 317 responses were valid and complete.

4. FINDINGS

4.3. Structural Model

In accordance with recent partial least square-structural equation modeling (PLS-SEM) evaluation standards (Hair et al., 2023), the model fit was assessed using several global fit indices. The standardized root mean square residual value was 0.056, which is below the recommended threshold of 0.08, indicating a good overall model fit. The normed fit index was 0.91, which also meets the minimum criterion of 0.90, further supporting model adequacy. Additionally, the discrepancy measures d_ULS and d_G were 0.225 and 0.193, respectively, and fall within acceptable limits for PLS-SEM analysis, indicating no significant model misfit. Although Chi-square (χ²=1,985.56) is reported, it is primarily descriptive in the context of PLS-SEM and not interpreted in the same way as in covariance based-structural equation modeling. These indicators collectively demonstrate that the structural model exhibits an acceptable fit with the empirical data, supporting the robustness of the hypothesized relationships. Table 3 (Hair et al., 2023) shows the results of testing the model fit for the structural model.

 

Table 3

Model fit indices for the structural model

Model fit index	Value	Recommended threshold	Reference	Interpretation
Standardized root mean square residual	0.056	<0.08	Hair et al. (2023)	Good model fit
Normed fit index	0.91	>0.90	Hair et al. (2023)	Acceptable fit
d_ULS (unweighted least squares discrepancy)	0.225	As low as possible	Hair et al. (2023)	Acceptable
d_G (geodesic discrepancy)	0.193	As low as possible	Hair et al. (2023)	Acceptable
Chi-square (χ2)	1,985.56	Not significant preferred	Hair et al. (2023)	Descriptive only in PLS-SEM

[i]

PLS-SEM, partial least square-structural equation modeling.

 

The structural model was assessed by examining the R-square (R2) and F-square (F2) as well as the path coefficient. As can be seen in Table 4, the R-square of AI adoption is 0.695, indicating that the five predictors can explain 69.5% of the variation in AI adoption. AI adoption can explain 29% of academic performance and 24.2% of cognitive skill development. After including the mediator self-efficacy, the R-square of self-efficacy as shown in Fig. 1 accounted to 63.4%. The R-square of the adoption of AI increased to 70.9% while the academic performance remained the same at 29% and increased to 24.3% for the cognitive skill development. The F-square for all significant paths are above 0.02 while for non-significant paths, they are less than 0.02.

 

Table 4

Result of hypotheses testing

Hypotheses	Path	B	Std.	T	p-value	F2	R2
H1	PU→Adoption of AI	0.308	0.049	6.276	<0.05	0.12	0.695
H2	PEOU→Adoption of AI	0.062	0.050	1.249	0.212	0.01
H3	Trust in AI→Adoption of AI	0.298	0.066	4.508	<0.05	0.11
H4	Social influence→Adoption of AI	0.287	0.062	4.588	<0.05	0.10
H5	Ethical concern→Adoption of AI	0.014	0.030	0.460	0.645	0.00
H6	PU→Self-efficacy→Adoption of AI	0.088	0.025	3.488	<0.001	0.03	0.709
H7	PEOU→Self-efficacy→Adoption of AI	0.057	0.020	2.895	0.004	0.02
H8	Trust in AI→Self-efficacy→Adoption of AI	0.035	0.018	1.931	0.054	0.01
H9	Adoption of AI→Academic performance	0.539	0.028	19.482	<0.05	0.21	0.290
H10	Adoption of AI→Cognitive skill development	0.492	0.052	9.548	<0.05	0.19	0.242

[i]

B, coefficient; Std., standard deviation; T, T-value; F2, F-square; R2, R-square.

 

Fig. 1

Structural model. PU, perceived usefulness; PEOU, perceived ease of use; AI, artificial intelligence.

 

Based on the direct effect model and the mediating effect model, Table 4 shows the results of testing the hypotheses. H is the hypothesis, B is the coefficient, Std. is the standard deviation, T is the T-value, p is the p-value, F2 is the effect size or F-square, and R2 is the R-square.

Our analysis supports seven of the ten hypotheses. PU has a strong positive effect on AI adoption (B=0.308, p<0.05) indicating that H1 is supported. The second hypothesis is rejected (H2), which is related to the effect of PEOU on adoption of AI. Trust in AI has a strong effect on the adoption of AI (B=0.298, p<0.05). Thus, H3 is supported. For H4, it is supported because the social influence significantly predicted the adoption of AI (B=0.287, p<0.05). For H5, it is rejected. Ethical concerns have no effect on adoption of AI. The non-significant effect of ethical concerns suggests students may compartmentalize these worries when making practical adoption decisions, prioritizing immediate academic benefits over longer-term ethical considerations.

In terms of the mediation role of self-efficacy, it was confirmed with PU and PEOU but not with trust in AI. This indicates that part of the relationship between PU and PEOU can be explained by self-efficacy while the self-efficacy has no role between trust in AI and adoption of AI. The mediation effects of self-efficacy are confirmed for PU and ease of use, highlighting students’ confidence as a key mechanism translating beliefs into adoption behaviour. Therefore, H6 and H7 are supported while H8 is rejected. The lack of mediation for trust through self-efficacy indicates these constructs operate independently in this sample. Students’ trust in AI systems appears unrelated to their confidence in using them. For the effect of AI adoption on educational outcome, it has a significant effect on academic performance (B=0.539, p<0.05) and cognitive skill development (B=0.492, p<0.05). Therefore, H9 and H10 are supported.

5. DISCUSSION

This study’s findings offer important theoretical contributions to technology acceptance literature in several ways. The supported hypotheses generally validate the applicability of TAM in AI adoption contexts within Saudi higher education. The strong predictive power of PU aligns with Davis (1989)’s original propositions, suggesting these core TAM constructs remain relevant for emerging educational technologies. In addition, social influence as a significant predictor coordinates with conventional wisdom from UTAUT (Venkatesh et al., 2003). This finding suggests that social norms may operate similarly in Saudi academic environments compared to Western contexts where UTAUT was developed.

The rejection of hypotheses related to PEOU and ethical concerns warrants deeper theoretical reflection. The non-significant effect of PEOU suggests that, in the Saudi higher education context, students may already possess sufficient digital literacy and exposure to technology, which reduces the salience of usability issues. This aligns with findings from studies in digitally advanced societies, where once baseline technological competence is achieved, ease of use becomes less critical than PU or social endorsement in predicting adoption. Regarding ethical concerns, the absence of significant influence may reflect a combination of cultural and contextual factors. Students may prioritize the immediate academic benefits of AI tools over abstract ethical risks, particularly when trust in state institutions and regulatory frameworks is high. Additionally, cultural and religious values, such as reliance on institutional authority and Islamic ethical orientations, may mitigate individual apprehensions about privacy and algorithmic fairness. These insights suggest that while global AI ethics debates are prominent, in practice, students in this context tend to compartmentalize such concerns, giving precedence to pragmatic educational gains.

The dismissal of ethical concerns by students, despite the prominence of global debates on AI ethics, requires a nuanced interpretation. In the Saudi context, several interrelated factors may account for this pattern. First, cultural norms and the collectivist orientation of society may lead students to defer to institutional and governmental oversight, assuming that AI systems deployed in universities are already aligned with ethical and legal standards. Second, religious and moral frameworks, particularly Islamic digital ethics, emphasize trust in authority and compliance with institutional rules, which may reduce the salience of individual-level ethical questioning. Third, the relatively low awareness of technical risks, such as data privacy breaches or algorithmic bias, may limit students’ ability to critically evaluate AI systems. Finally, the pragmatic orientation of students, who are more concerned with immediate academic benefits, further explains why ethical issues were deprioritized. These findings suggest that ethical awareness campaigns, integrated into digital literacy and AI literacy programs, could help students better engage with the ethical dimensions of AI in education while maintaining trust in institutional safeguards.

The significant mediating role of self-efficacy (H6, H7) extends Bandura (1986)’s social cognitive theory by demonstrating its relevance for AI technology adoption. However, the lack of mediation for trust (H8) suggests that confidence in AI systems may function differently than trust in human instructors or conventional technologies, highlighting a need for theoretical refinement regarding trust mechanisms in AI-enabled learning environments. The non-significant effect of ethical concerns contradicts much contemporary AI ethics literature, possibly indicating a disconnect between abstract ethical considerations and practical adoption decisions among students.

6. IMPLICATIONS

This study improves theoretical understanding of educational technology adoption. It initially strengthens TAM by confirming its core concepts and defining key AI adoption boundaries. The strong predictive power of PU supports TAM, proving its relevance for advanced technology such as AI. The non-significant effect of PEOU challenges the TAM assumptions, suggesting that adoption decisions may be less influenced by ease of use as users gain technical proficiency.

The present research improves theoretical understanding by including AI trust as a major predictor, surpassing traditional TAM components. This supports earlier work calling for models that account for AI systems’ unique traits, notably their opaque decision-making processes. Human-AI interactions require innovative theoretical frameworks to explain trust formation since users must rely on system outputs without understanding the algorithms. The significant impact of social influence supports UTAUT theory in collectivist education. This finding suggests that social norms may similarly affect technology adoption in numerous cultures, even though reference groups (e.g., peers versus instructors) and their proportionate effect may differ. Ethics’ limited impact weakens assumptions in AI ethics literature, demonstrating a disjunction between theoretical ethical norms and implementation choices that requires additional theoretical research.

Beyond validating established frameworks such as TAM, UTAUT, and social cognitive theory, this study highlights the need for incorporating additional constructs that capture the unique challenges of AI adoption in educational contexts. Future research could integrate variables such as techno-stress, which reflects the psychological strain arising from overexposure to AI-driven platforms, and AI literacy, which measures students’ ability to critically understand and evaluate AI systems. Both constructs are increasingly relevant as universities adopt advanced digital infrastructures. Similarly, resistance to change may offer explanatory power in contexts where traditional learning methods remain dominant and cultural norms shape technology acceptance. Embedding these constructs into future models would not only extend theoretical novelty but also create a culturally adaptive framework tailored to Middle East and North Africa (MENA) education systems, thereby advancing the literature beyond the boundaries of classical adoption theories.

The findings reveal many AI adoption measures for institutions. Institutions must emphasise academic benefits while using AI technology since PU affects acceptance. This may include discipline-specific AI application presentations, peer demonstration programs where early adopters discuss their discoveries, and practical demonstrations of how AI tools may improve learning. AI systems need trust, therefore universities must build it. Transparency standards for institutional AI tools, clear descriptions of AI system capabilities and constraints, and outlets for student algorithmic concerns are practical ways. These methods may help pupils trust AI in education. The partial mediation of self-efficacy underlines the necessity for AI technology instruction to increase student confidence. Institutions should offer incremental skill-building programs like practical workshops and supervised practice instead of only technical instruction. Customised programs must focus on how AI may improve academic learning. Although ethical concerns did not significantly affect adoption decisions, student answers show that institutions should not ignore ethical education. Institutions should incorporate AI ethics into digital literacy courses, create academic AI use norms, and host ethical discussions on AI in education. AI adoption positively impacts academic achievement and cognitive skill development, suggesting that institutions should proactively integrate AI into curriculum to optimise learning. This may entail revising exams to reflect AI-enhanced skills, giving faculty development on AI pedagogy, and helping teachers integrate AI technology to maximise educational advantages.

7. CONCLUSION

This study advances our understanding of AI adoption in higher education by validating core technology acceptance principles, while identifying important cultural specificities in the Saudi context. The findings suggest that while traditional adoption factors remain relevant, the unique characteristics of AI technologies and local educational cultures require theoretical refinement and tailored implementation approaches. By addressing the identified limitations in future research, scholars can develop more comprehensive models of educational technology adoption in the AI era. For practitioners, the results highlight both opportunities and challenges in leveraging AI tools to enhance teaching and learning outcomes. The findings of this study have several limitations. The focus on the Saudi context limits the findings to Saudi Arabia and undergraduate students.

Another key limitation of this study lies in the measurement of academic performance and cognitive skills, which relied exclusively on self-reported data. While self-assessments provide valuable insights into students’ perceptions, they are inherently subjective and may be influenced by social desirability bias or limited self-awareness. The absence of objective performance indicators, such as grade point average records, standardized test scores, or instructor evaluations, reduces the robustness of outcome assessment. Future research should aim to triangulate self-reports with objective academic data and faculty assessments, thereby enhancing measurement validity and allowing for a more comprehensive evaluation of the educational impact of AI adoption.

A further limitation concerns the scope of the sample, which was restricted to undergraduate students in public universities. While this focus ensured a clear and context-specific understanding of AI adoption in Saudi higher education, it limits the generalizability of the findings. Patterns of adoption may differ in private universities, where resources and institutional policies may be more flexible, or among graduate students, who often engage with AI in more specialized and research-intensive ways. Moreover, the exclusive national focus does not capture potential cross-cultural variations across MENA or other global contexts. Future studies should therefore broaden the sampling frame to include private institutions, graduate cohorts, and cross-national comparisons in order to validate and extend the present findings.

Future studies should examine whether these patterns hold in different national contexts. This study adopted a cross-sectional design. Therefore, it cannot capture how adoption patterns evolve over time. Further studies are recommended to conduct longitudinal research because it could reveal important developmental trajectories in AI tool usage. In term of the measurements of the variables, academic performance can be measured objectively or subjectively. Similarly, measures of cognitive skill development relied on self-reports rather than objective assessments. Future work should incorporate standardized academic and cognitive tests. The focus on public universities may not reflect patterns at private institutions. Additionally, graduate students were not included in the sample. Future research should investigate why ease of use is less impactful in this context compared to Western studies. In addition, future studies can examine the disconnect between ethical concerns and adoption behaviour. Furthermore, future studies should explore instructor perspectives on AI adoption to complement student data. Additional variables can be added as moderators or mediators. These include the AI literacy, perceived risk, technostress, and IT knowledge of students and governmental policies to support the adoption of AI and its impact on educational outcomes.