The concept of technology acceptance originated with Fishbein and Ajzen’s Theory of Reasoned Action (TRA) in 1975, which defines it as an individual’s intention to adopt a certain technology to accomplish specific tasks. The fundamental assumption of the theory is that users’ beliefs influence their behavioral intentions, and these intentions, in turn, determine their actual behavior toward using a given technology (Taylor and Todd, 1995).

Building on TRA, Ajzen proposed the TPB, which posits that an individual’s behavioral intention is influenced by three key factors: attitude toward the behavior, subjective norm, and perceived behavioral control. The more positive one’s attitude, the stronger the perceived social norms and behavioral control, and the stronger the behavioral intention—ultimately increasing the likelihood of performing the behavior (Ajzen, 1991). TPB emphasizes that behavioral intentions are not formed arbitrarily but are shaped by personal attitudes, social expectations, and perceived control. Subjective norms play an essential role in decision-making, as individuals often experience social pressure when determining whether to engage in a particular behavior.

Davis (1989) later applied TRA and TPB to the field of information systems to explain users’ acceptance of information technologies, thus developing the TAM. TAM explains the factors influencing an individual’s adoption of a particular technology within an organizational or social context (Davis, 1989). It identifies two core determinants of technology adoption intention: perceived usefulness (PU) and perceived ease of use (PE) which jointly determine users’ behavioral intentions.

Perceived usefulness refers to a potential user’s subjective evaluation of the extent to which using a particular technology or product will enhance task performance. Perceived ease of use denotes the degree to which a user believes that using a given technology will be free of effort (Abdullah et al., 2016; Chen et al., 2025). In the context of GenAI, perceived usefulness mainly reflects its capacity to assist in mastering theoretical knowledge, building cognitive frameworks, filtering relevant information, facilitating academic learning, supporting foreign-language reading, aiding programming, and fostering interdisciplinary understanding (Liu and Zhang, 2025). Perceived ease of use encompasses aspects such as ease of access (e.g., no registration or payment requirements), intuitive operation, high response speed, accuracy and comprehensiveness of output, multimodal support, adaptability, and privacy protection.

Accordingly, this study proposes the following hypotheses:

Following the theoretical logic of TAM and TPB, behavioral intention serves as a direct antecedent of actual usage behavior. Individuals with stronger behavioral intentions are more likely to translate these intentions into real-world actions. Thus, the following hypothesis is proposed:

Furthermore, social norms often exert a significant influence on technology adoption decisions. Peer and institutional expectations have been shown to shape individuals’ behavioral intentions to integrate AI into their learning processes (Kostić-Ljubisavljević and Samčović, 2024; Al-Mamary et al., 2024). Based on this, the following hypothesis is proposed:

The Technology Resistance Theory posits that technology adoption decisions depend not only on perceived benefits but also on perceived risks. The PRT, introduced by Bauer at Harvard University in 1960, was initially applied to market research to explain phenomena such as information search, brand loyalty, reference groups, and purchasing decisions (Bettman, 1973). Its core idea is that consumers’ decision-making is accompanied by subjective perceptions of uncertainty regarding outcomes, and perceived risk consists of both uncertainty and consequence severity (Barach, 1969). In the context of GenAI, perceived risk can be divided into three categories:

(1) Perceived Personal Risk (PPR).

When individuals begin using new technologies, they often encounter personal concerns that may hinder adoption. This evaluation process involves critical thinking (Walter, 2024), whereby users assess information produced by GenAI tools such as ChatGPT or Midjourney before forming judgments and making decisions (Shi et al., 2020; Li, 2025; Huang et al., 2025). Students may perceive personal academic risks such as exposure to misleading information, reduced critical or creative thinking, weakened expression and analytical ability, or the use of inaccurate or fabricated data (Zhang L. et al., 2025; Zhang X. et al., 2025; Huang and Wu, 2025; Chan and Lee, 2025).

(2) Perceived Academic Environmental Risk (PER).

Beyond individual concerns, academic environmental risks also affect adoption intentions. Although GenAI technologies have transformed academic practices, they simultaneously pose challenges, such as undermining academic integrity and reducing the effectiveness of plagiarism detection systems (Huang et al., 2025). Zhou’s natural and longitudinal experimental studies demonstrated that using ChatGPT may diminish sustained creative output and lead to content homogenization (Zhou et al., 2025).

(3) Perceived Social Risk (PSR).

Featherman and Pavlou (2003) argued that higher perceived social risks related to a product or service may reduce users’ willingness to recommend it, particularly under peer influence. In the case of GenAI, social risks include concerns over information authenticity, employment displacement, academic misconduct, and technological advancement surpassing regulatory oversight (Hamed et al., 2024; Salari et al., 2025).

Accordingly, the following hypotheses are proposed:

The Knowledge–Attitude–Practice model is one of the most widely used frameworks to explain how individual knowledge and beliefs influence behavioral change, particularly in the context of health behavior research. The model suggests that knowledge serves as the precursor to attitude formation, and attitudes, in turn, drive behavioral practice (Andrade et al., 2020; Salama et al., 2025). Knowledge reflects awareness and understanding of relevant information; attitude represents an individual’s positive or negative evaluation toward an object; and practice denotes the habitual or intentional behaviors exhibited in real-world contexts (Zhang et al., 2023). Typically, individuals acquire new knowledge that shapes their attitudes, which subsequently influence their behavioral actions (Ritter et al., 2025).

In the context of GenAI, this study examines university students’ levels of cognitive understanding—such as their awareness of GenAI concepts, familiarity with machine learning principles, understanding of AI “hallucination” phenomena, and attitudes toward AI-generated content—to assess how these knowledge components influence their behavioral intentions. Accordingly, the following hypothesis is proposed:

The machine learning process consists of three key steps: feature input, model construction, and performance evaluation (Qian et al., 2024; Qiu et al., 2025). We selected input variables that are likely to have a significant impact on students’ use of GenAI tools. Specifically, these include measurement variables under the latent constructs that significantly influence the students’ behavioral intentions path in the structural equation model. During the model construction phase, we employed eight machine learning algorithms: Linear Regression, Lasso Regression, Ridge Regression, K-Nearest Neighbors, Support Vector Regression, Random Forest Regression, XGBoost, and Gradient Boosting Regression.

To ensure optimal performance, we followed rigorous training protocols in applying machine learning algorithms. Hyperparameter tuning is a critical step in achieving optimal model performance. We used a grid search strategy to identify the best combination of parameters and fine-tuned them using cross-validation. Specifically, the dataset was randomly partitioned into k folds for k -fold cross-validation, where each iteration trained the model on k − 1 folds and validated it on the remaining fold. We set k to 10, used root mean square error (RMSE) as the optimization metric (Reid et al., 2015; Van Den Hoogen et al., 2019; Qiu et al., 2025), and reported the coefficient of determination (R²) to indicate the model’s goodness-of-fit between predicted and actual values. All machine learning models were implemented using Python (version 3.13).

Interpretable Machine Learning techniques were employed to enhance the transparency, credibility, and explanatory power of the predictive models. These methods facilitate a deeper understanding of the model’s internal logic, support debugging and optimization processes, and help identify key variables that most strongly influence students’ behavioral intentions toward GenAI adoption (Linardatos et al., 2020).

Two complementary IML approaches were applied to assess feature importance. The first method was based on node-splitting analysis derived from tree-based ensemble models. Specifically, the built-in get_booster() and get_score() functions in the XGBoost package (Python) were used to quantify feature importance according to the frequency and gain associated with each variable’s contribution to model improvement. This method provides a global view of variable importance within the decision tree structure. The second method employed SHAP (SHapley Additive exPlanations), a game-theoretic approach that decomposes model predictions into additive feature contributions (Jia et al., 2024; Qian et al., 2024). Using the Explainer module from the SHAP package in Python, we computed both global and local interpretations. Globally, SHAP values rank the overall importance of input features; locally, they quantify each feature’s marginal contribution to an individual prediction. A positive SHAP value indicates that a feature positively contributes to the likelihood of GenAI usage, whereas a negative value implies an inhibiting effect. By aggregating SHAP values across samples, this approach enables robust importance ranking and reveals potential interaction effects between features. A variable is considered a core influencing factor only when it is identified by two different interpretive methods simultaneously.

Collectively, these interpretable modeling strategies constitute the second stage of the proposed two-step SEM–XML framework. Following the SEM-based hypothesis testing and path analysis in the first stage, the IML-based importance analysis provides a data-driven validation and refinement of the conceptual model. This systematic integration of path analysis and factors identification strengthens the methodological robustness of the study and yields deeper insights into the behavioral mechanisms underlying GenAI adoption in higher education.

According to established methodological standards (Albayati, 2024; Leguina, 2015; Diamantopoulos and Siguaw, 2000; Yu et al., 2025), it is necessary to evaluate the reliability, validity, factor loadings, convergent validity, composite reliability and discriminant validity of all measurement items.

Based on prior literature, the reliability and construct validity of each measurement item were first examined. Reliability reflects the internal consistency of the measurement results—if an instrument produces stable results under the same conditions, it is considered reliable. Higher reliability indicates stronger inter-item correlations, implying that the items measure the same latent construct consistently. Reliability is typically assessed using Cronbach’s α, with a threshold value above 0.70 indicating acceptable reliability. Construct validity determines whether the measurement instrument accurately captures the theoretical construct which intends to measure. It reflects the correspondence between the observed indicators and the underlying latent variables. Higher construct validity implies that the items can effectively differentiate between distinct latent constructs, making the data suitable for factor analysis. The Kaiser–Meyer–Olkin (KMO) measure of sampling adequacy is commonly used, with a minimum acceptable value of 0.60 for each construct.

To further verify the absence of common method bias, a Harman Single-Factor Test was conducted. The test results showed that the first principal component explained 34.2% of the total variance, which is below the 40% threshold, indicating that common method bias is not a concern in this study.

As shown in Table 4, all constructs exhibit strong reliability and validity. Specifically, all Cronbach’s α values exceed 0.70; except for Perceived Academic Environmental Risk (0.795) and Subjective Norm (0.727), all others exceed 0.85, indicating robust internal consistency. Furthermore, all KMO values are greater than 0.70, confirming the suitability of the data for factor analysis and indicating that the items effectively represent their respective constructs. The square root of the AVE for each latent variable is greater than the standardized correlation coefficients outside the diagonal (see Table 5), which further supports the good discriminant validity of the latent variables.

To further assess whether observed variables effectively measure the latent constructs, Confirmatory Factor Analysis (CFA) was conducted. The model’s quality was evaluated using factor loadings, Squared Multiple Correlations (SMC), Composite Reliability (CR) and Average Variance Extracted (AVE) (Lu et al., 2025). (1) Factor loading represents the standardized correlation between an observed variable and its corresponding latent construct. (2) SMC indicates the proportion of variance in an observed variable explained by the latent factor. (3) Convergent validity measures the degree of agreement among items assessing the same construct, typically assessed by AVE, where a value above 0.50 indicates satisfactory convergence. (4) CR provides a more accurate estimate of internal consistency than Cronbach’s α during CFA, with values above 0.70 considered acceptable.

Measurement items with factor loadings below 0.70 or SMC values below 0.50 were removed and renumbered. The revised indicators are presented in Table 6. All constructs achieved CR > 0.70 and AVE > 0.50, confirming adequate convergent validity and internal consistency reliability.

Drawing from the theoretical framework and the results of the CFA, a Structural Equation Model (SEM) was constructed using AMOS 26.0 for model estimation. Figure 3 illustrates the SEM structure and standardized path coefficients, while Table 7 presents the model fit indices.

Model-data fit was evaluated by examining fit indices. RMSEA (Root Mean Square Error of Approximation) measures how well the model fits the population covariance matrix. CFI (Comparative Fit Index) compares the fit of the hypothesized model to a baseline model. IFI (Incremental Fit Index) also evaluates model fit by comparing the hypothesized model to a baseline. TLI (Tucker–Lewis Index) assesses model fit by comparing the chi-square value of the model to a null model. NFI (Normed Fit Index) compares the chi-square value of the proposed model to a baseline. The obtained fit indices were compared with criterion values to interpret model-data fit. The criteria were as follows: χ / df ≤ 2.00 , RMSEA ≤ 0.05 , CFI ≥ 0.95 , IFI ≥ 0.95 , TLI ≥ 0.95 , NFI ≥ 0.95 indicated a perfect fit; χ / df ≤ 3.00 , RMSEA ≤ 0.08 , CFI ≥ 0.90 , IFI ≥ 0.90 , TLI ≥ 0.90 , NFI ≥ 0.90 indicated a good fit; χ / df ≤ 5.00 , RMSEA ≤ 0.08 , CFI ≥ 0.85 , IFI ≥ 0.85 , TLI ≥ 0.85 , NFI ≥ 0.85 indicated an acceptable/moderate fit (Tabachnick et al., 2007; Brown, 2015; Kline, 2023; Delibalta et al., 2025; Hu and Bentler, 1999). In Table 7, all model fit indices meet or exceed recommended standards, except for the slightly lower TLI, indicating overall satisfactory model fit ( χ /df = 2.863, RMSEA = 0.043, CFI = 0.923, IFI = 0.954, TLI = 0.895, NFI = 0.933). These results suggest that the model is suitable for explaining both the relationships between latent and observed variables and the structural paths among latent constructs.

As shown in Table 8, the results of this paper found that:

From the perspective of the hypothesis testing results, three key findings emerge:

(1) Perceived ease of use and perceived usefulness significantly promote behavioral intention and actual use.

Perceived ease of use and perceived usefulness both show significant positive effects on students’ behavioral intention and on their reported actual use. When tools are easy to operate and clearly help students complete or improve academic work, students are more likely to intend to use them and to incorporate them into real study and research activities. This finding is consistent with the TAM and emphasizes that both functional value and usable design jointly drive adoption in higher-education settings.

Implications: Developers and educators should prioritize enhancing both the user experience and the functional utility of GenAI tools. This can be achieved by simplifying operational procedures, offering guided tutorials, integrating discipline-specific applications, and incorporating GenAI into digital literacy training. These steps will help build user confidence and promote sustained engagement.

(2) Subjective norms and technology awareness significantly enhance adoption intention.

Subjective norms, reflecting perceived social expectations from teachers and peers, and technology awareness, representing understanding of GenAI’s capabilities, both significantly impact adoption intention. These findings suggest that students’ behavior is shaped by both social influence and technical understanding. It highlights the importance of peer endorsement and technological literacy as key drivers of technology acceptance in educational contexts.

Implications: Institutions should foster positive social influence through peer modeling, faculty demonstrations, and case-sharing activities. Simultaneously, they can improve technological literacy by organizing training workshops and integrating GenAI into the curriculum. Embedding GenAI in real academic tasks can also strengthen its perceived legitimacy and encourage sustained use in higher education.

(3) Perceived personal and academic risks inhibit adoption, while social risks are insignificant.

Both perceived personal academic risk and perceived academic environmental risk negatively influence GenAI adoption, while perceived social risk does not have a statistically significant effect. Students primarily express concerns about academic dependency, integrity violations, and institutional penalties, rather than societal judgment. This suggests that, in academic settings, students are more focused on institutional norms and academic outcomes than on social approval. These results are consistent with prior studies, which show that individual and institutional academic risks exert a stronger deterrent effect than generalized social risks.

Implications: Universities and technology providers should establish clear ethical guidelines and frameworks for academic integrity, offer training on responsible usage, and demonstrate legitimate academic applications of GenAI. Addressing students’ concerns about academic outcomes and providing clarity on institutional policies will be more effective than focusing on social approval. This approach will help build trust and encourage sustained engagement with GenAI technologies.

To enhance the interpretability and robustness of the results, a series of supervised machine learning (ML) models were developed using Python and relevant libraries, including scikit-learn, XGBoost, and NumPy. These models were designed to concretize the relative importance of factors influencing students’ actual usage of GenAI tools, enabling a quantitative representation of abstract constructs. Specifically, eight regression algorithms were employed to ensure model diversity and comparative reliability: Linear Regression (LR), Lasso Regression (Lasso), Ridge Regression (Ridge), K-Nearest Neighbors (KNN), Support Vector Regression (SVR), Random Forest Regression (RFR), XGBoost (XGB), and Gradient Boosting Regression (GBR).

Each algorithm captures unique relationships between the independent variables (i.e., influencing factors) and the dependent variable (actual usage behavior). Linear models (LR, Lasso, Ridge) serve as parametric benchmarks for identifying linear associations and regularization effects. Nonlinear models such as KNN, SVR, and RFR handle complex, multidimensional data, while ensemble-based algorithms (XGBoost and GBR) combine multiple weak learners to achieve higher predictive accuracy and generalization. The modeling process followed standard ML procedures, including data preprocessing, feature standardization, and a 70:30 train-test split. All models were trained and validated on the same dataset to ensure comparability. Model performance was primarily assessed using the coefficient of determination (R²), along with learning curves to evaluate convergence and generalization.

In the first stage of SEM path analysis, it was found that PSR had an insufficiently significant impact on BI. As a result, the corresponding measurement variables for this latent variable were excluded from this stage. The explanatory variables inputted into the model consisted of the remaining complete latent variables—TA, PE, PU, SN, PAR, and PER—along with their corresponding measurement variables (see Table 4). The dependent variable, AB, was represented by the mean of its measured variables.

Hyperparameter optimization is crucial for achieving optimal model performance. Grid search with k-fold cross-validation was employed to identify the best combination of parameters for each model. R²-based learning curves were generated to assess both training and validation performance.

Upon evaluation, XGBoost demonstrated superior predictive accuracy and convergence (test R² = 0.9867; cross-validation R² = 0.9812), outperforming other algorithms (see Figure 4). Therefore, XGBoost was selected as the final model for factor concretization. Optimal hyperparameters were n_estimators = 290, learning_rate = 0.1, max_depth = 10, gamma = 0, min_child_weight = 2, subsample = 1, and colsample_bytree = 0.4. However, we acknowledge that the deep tree structure and lack of subsampling may increase the risk of overfitting. Given that the primary goal of this study is to concretize the core influencing factors, overfitting is less of a concern for the current analysis.

To further analyze the model’s prediction results, we first selected the top 10 features based on their importance scores and visualized them, as shown in Figure 4. By comparing the rankings from two different methods, we found that 8 of the top 10 features overlapped, further confirming their importance in predicting the model’s results. The average SHAP (Shapley Additive Explanations) values revealed the most influential factors in predicting students’ actual usage behavior.

As shown in Figure 5, TA2 (understanding of deep learning, large models, and related technical principles) ranked among the top three in both ranking methods, indicating that an understanding of technology is crucial in influencing students’ use of GenAI tools. In addition, TA1 (ability to distinguish between the concepts of AI, AGI, and GenAI) also ranked highly, highlighting that students’ basic understanding of these technologies plays a significant role in determining their use of GenAI tools. PE1 (ease of use and accessibility), PU1 (helping to master theoretical knowledge), and PU2 (supporting the establishment of knowledge frameworks) emphasized the importance of the tool’s usability and educational functionality in students’ decisions to use GenAI tools. At the same time, features such as PAR4 (risk of false data), PER2 (failure of plagiarism detection systems), and PER3 (duplication of academic results) ranked highly, reflecting students’ strong concern about academic integrity when using these tools.

Specifically, TA1 (ability to distinguish between AI, AGI, and GenAI concepts) ranked among the top three in both methods, reflecting that, in the context of AI technologies gradually permeating education and daily life, students’ ability to clearly differentiate between different types of AI impacts their willingness to adopt and effectively use these tools. Educators should focus on strengthening students’ understanding of basic AI concepts to lower the barriers to technology acceptance, build students’ confidence in using GenAI, and ultimately improve its effectiveness in actual learning. TA2 (understanding of deep learning, large models, and related technical principles) ranked highly, suggesting that students’ understanding of the fundamental principles behind deep learning and large models significantly impacts their acceptance and frequency of GenAI tool use. As generative AI technology continues to be applied in education, students’ awareness of the complexity and innovation behind these technologies becomes a key factor influencing their usage decisions.

PE1 (ease of use and accessibility), PE1 (helping to master theoretical knowledge), and PUT (supporting the establishment of knowledge frameworks) were all ranked among the top 10 in both methods, emphasizing the critical role of tool usability and perceived usefulness in students’ decisions to adopt GenAI tools. In practice, students are more likely to choose tools that are easy to use and directly aid their learning. Therefore, the usability and intuitiveness of a product will directly influence its popularity and effectiveness.

The high ranking of features like PAR4 (risk of false data), PER2 (failure of plagiarism detection systems), and PER3 (duplication of academic results) reveals students’ heightened concern about the academic credibility and integrity of generated content. In the current academic environment, issues related to the generation of false data and the failure of plagiarism detection systems have become widespread concerns. Students are increasingly worried about the authenticity and academic credibility of generated content, which may lead them to use these tools more cautiously or even limit their usage due to fears of academic misconduct.

Overall, the core factors influencing university students’ actual use of GenAI tools are not only related to technological understanding and the ease of use of the tools, but also deeply connected to multidimensional aspects such as academic integrity, data security, and learning outcomes. As GenAI technology evolves, a focus on foundational knowledge dissemination, balancing technological innovation with academic standards, ensuring usability and utility of tools, and managing potential risks will be key issues to address in future education on GenAI literacy for university students.