GenAI is transforming higher education by offering students innovative approaches to accessing knowledge and enhancing their learning processes (Strzelecki and ElArabawy, 2024). GenAI, such as OpenAI’s ChatGPT and Baidu’s Ernie Bot, have transformed the learning process by providing personalized support, thereby enhancing students’ learning performance (Wang et al., 2024). These tools foster deeper cognitive processing and enhance students’ interest in problems-solving (Jin et al., 2025a). For instance, these tools can generate case studies, role-playing scenarios, and debate topics, which stimulate students’ ideas (Chan and Hu, 2023; Lodge et al., 2023). By using GenAI, students can enhance their ability to organize learning tasks, monitor progress, and receive immediate feedback, all of which are essential components of self-regulated learning. Furthermore, GenAI can provide customized learning paths to enable students to engage in adaptive learning experiences (Yu et al., 2025). This contributes to improved time management, higher engagement, and increased motivation, as learners gain a greater sense of control over their learning processes. Recognizing these benefits, educators have increasingly integrated GenAI tools to foster dynamic learning environments that promote active learning, and critical thinking (Tang et al., 2025).

In the context of Mainland China, the rapid integration of GenAI into higher education have fostered a distinctive environment for student learning. GenAI, such as OpenAI’s ChatGPT and domestic platforms like Baidu’s ERNIE Bot, are being increasingly adopted in Chinese universities, providing functions such as intelligent question-and-answer (Q&A) systems, academic writing support, and content generation. Surveys data indicate that GenAI-assisted functions, including paper editing and Q&A systems, have reached a penetration rate of approximately 84% in Chinese universities (Zhang and Wang, 2025). This widespread adoption reflects China’s policy emphasis on advancing educational digitalization and enhancing university students’ digital literacy. However, Chinese higher education is characterized by unique cultural and institutional features that affect how students engage with GenAI-assisted learning (Huang et al., 2019). For example, Chinese students are typically socialized within teacher-centered and exam-oriented educational environments, which may limit their capacity for self-regulation in interacting with GenAI tools (Yan et al., 2024). Moreover, concerns regarding academic integrity and data privacy are pronounced in Mainland China, where institutional trust and adherence to national AI regulations play a central role in shaping student adoption behavior (Teo and Huang, 2019). Therefore, investigating Chinese university students’ self-regulated learning in GenAI-assisted environments is timely and significant, as it captures an educational ecosystem characterized by distinctive cultural, institutional, and technological dynamics.

However, the integration of OpenAI’s ChatGPT and Baidu’s Ernie Bot into higher education also presents risks and challenges, particularly concerning self-regulated learning (Jin et al., 2025b). A primary concern is the potential overreliance on these tools, which may hinder students’ active engagement in critical thinking and reflection (Strzelecki, 2024). By relying on GenAI for immediate responses, students may bypass essential cognitive processes, which can impair their ability to set learning goals and monitor progress (Zhou et al., 2024). Furthermore, sensitive information disclosed during interactions may be vulnerable to breaches or misuse, potentially eroding students’ confidence in employing these tools (Chiu, 2024). Also, GenAI-generated content may carry biases from its training data, potentially resulting in inaccuracies that distort students’ decision-making (Farrokhnia et al., 2024). This limitation undermines students’ ability to critically evaluate generated content and hinders the development of their self-regulation learning. Concerns about academic integrity also arise because GenAI tools can generate assignments, essays, or creative works, undermine the authenticity of student submissions and hinder the development of independent learning (Bhullar et al., 2024). Finally, university students with limited technical proficiency may encounter technological barriers that hinder their effective use of GenAI, thereby constraining their self-regulation learning (Fergus et al., 2023).

To address these risks and foster their self-regulated learning, universities should focus on cultivating students’ digital literacy, enabling them to use GenAI effectively while retaining autonomy (Wang et al., 2024). Establishing clear ethical standards and safeguarding data privacy are essential for creating a responsible educational environment (Cotton et al., 2024). Ensuring transparency in AI algorithms and content generation further enables students to critically evaluate GenAI outputs and ensures that these tools support their self-regulation learning (Arthur et al., 2025).

In education, self-regulation refers to learners’ ability to actively manage their cognitive, emotional, and behavioral processes in the learning process (Kim et al., 2023). It involves setting goals, monitoring progress, adjusting learning strategies, and maintaining motivation. In the context of GenAI-assisted learning, self-regulation is particularly critical. Using GenAI tools such as ChatGPT or Baidu’s ERNIE Bot requires learners to effectively balance tool interaction with the active management of their own learning processes (Wang, 2024). Self-regulation encompasses not only traditional cognitive strategies but also the ability to adapt to the distinctive affordances of GenAI tools (e.g., content generation, feedback loops, and personalized learning; Hew et al., 2025). These tools support learners in organizing their learning process, while they also require students to actively monitor their interactions with the technology to prevent over-reliance (Zhang and Tur, 2024). Therefore, in this study, self-regulation is defined as students’ capacity to manage their learning experiences in GenAI-assisted learning environments by adjusting their strategies, behavior, and emotions to achieve the intended learning outcomes (Xu et al., 2025).

To understand and measure self-regulation in GenAI-assisted learning, Liaw and Huang (2013) proposed a three-tier model, which has become an important framework in e-learning research (Figure 1). This model comprises three interrelated tiers that shape how learners regulate their behaviors in technology-supported learning. The first tier highlights individual characteristics (e.g., self-efficacy and anxiety) and environmental factors (e.g., information system quality), which influence learners’ cognitive and affective responses during the learning process. High-quality GenAI tools (e.g., OpenAI’s ChatGPT, Baidu’s Ernie Bot) can alleviate learners’ cognitive load and offer more personalized learning experiences, thereby fostering their self-regulation learning. The second tier encompasses cognitive and affective factors, including perceived satisfaction and usefulness. These factors shape learners’ perceptions of the tools they use, thereby influencing their cognitive and affective responses. The model posits that learners who perceive tools as satisfying and useful are more likely to engage in self-regulation learning. Finally, the model highlights the behavioral intention tier, which is driven by perceived self-regulation. In this study, this tier reflects the learner’s intention to continue using GenAI, based on their usefulness and satisfaction. In other words, learners who can effectively regulate their learning with GenAI are more likely to sustain their use of these tools in the future.

Although TAM and UTAUT are widely applied to examine technology acceptance, they primarily emphasize perceived ease of use, perceived usefulness, and behavioral intention, focusing on the technological aspects of users’ interactions with tools. However, these models overlook the cognitive, emotional, and behavioral dynamics that are crucial for understanding self-regulated learning in the context of educational technologies (Şimşek et al., 2025; Liu et al., 2025b). In contrast, Liaw and Huang’s (2013) three-tier self-regulation model offers a more comprehensive framework by integrating individual characteristics, emotional responses, and behavioral intention, providing a deeper understanding of how learners interact with GenAI tools such as ChatGPT and Baidu’s ERNIE Bot in higher education. This model extends the TAM/UTAUT frameworks by addressing the psychological and emotional factors that influence self-regulation in technology-mediated learning environments. While TAM and UTAUT primarily examine user acceptance based on technology perceptions, the three-tier model incorporates learners’ self-efficacy, anxiety, and emotional responses, which are particularly important in GenAI-assisted learning. It highlights the role of individual traits and contextual factors in shaping how students engage with GenAI tools, balancing cognitive, emotional, and behavioral aspects to achieve their learning goals. The novelty of applying the three-tier model in the context of GenAI-assisted learning lies in its ability to address the interplay between learners’ personal characteristics, their emotional responses, and the quality of the GenAI tools they use. Unlike the more limited scope of TAM/UTAUT, this model provides a richer, more nuanced understanding of how self-regulation influences learners’ use of GenAI tools. This theoretical framework offers valuable insights into the psychological, emotional, and contextual factors that shape learners’ self-regulation and how they perceive and interact with GenAI tools in their academic learning environments. Consequently, this expanded framework contributes to a deeper theoretical understanding of self-regulated learning in GenAI-enhanced education.

SEM is a powerful statistical technique for examining complex relationships between observed and latent variables (Anderson and Gerbing, 1988). Unlike traditional regression analysis, SEM enables researchers to test simultaneous causal pathways among multiple variables, including direct and indirect effects (Fan et al., 2016). This is particularly valuable for investigating constructs such as self-regulation, which encompass interrelated cognitive, emotional, and behavioral factors that cannot be directly measured. In this study, SEM was employed to construct a theoretical framework for examining the factors that influence learners’ self-regulation in GenAI-assisted learning environments. By using SEM, this study examined how individual characteristics (e.g., self-efficacy and anxiety), environmental factors (e.g., information system quality), and cognitive and affective responses (e.g., perceived satisfaction and usefulness) interrelate to shape learners’ self-regulation and behavioral intentions. This approach allows for the evaluation of both the direct effects of these factors and the modeling of their indirect effects. The use of SEM in this study is justified by its ability to handle complex data structures, accommodate measurement error, and model latent variables, which are central to this research (Fornell and Larcker, 1981). Unlike other modeling techniques, SEM offers a more nuanced understanding of how multiple factors, including learners’ emotional and cognitive responses, influence their learning behaviors. Given the dynamic and personalized nature of GenAI (e.g., OpenAI’s ChatGPT and Baidu’s Ernie Bot), SEM enables the capture of the complex interactions between learners and these tools, as well as their potential to support or hinder self-regulated learning. Overall, SEM serves as an appropriate and essential methodological approach for exploring the psychological, emotional, and contextual factors that influence self-regulation in GenAI-assisted learning. It provides a holistic understanding of how interactions with GenAI tools influence learners’ behaviors, forming a cornerstone for this study’s design.

Bandura (1986) defined perceived self-efficacy as an individual’s confidence in their ability to successfully accomplish a specific task. In this study, perceived self-efficacy refers to learners’ confidence in using GenAI tools to achieve their learning goals (Börekci and Uyangör, 2025). Kumar (2021) reported that the effect of learners’ self-efficacy on chatbot-assisted learning was limited. However, some studies have suggested that higher self-efficacy can enhance learners’ learning performance and persistence, as it is associated with a more positive attitude toward this learning approach (Chang et al., 2022; Divekar et al., 2022). For instance, Lee et al. (2022) found that GenAI not only supported students in understanding knowledge but also offered immediate solutions to the problems they encountered, thereby enhancing their learning confidence. Therefore, perceived self-efficacy is crucial for GenAI-assisted learning.

Computer anxiety is an uncomfortable emotional state characterized by nervousness, worry, and apprehension (Doll and Torkzadeh, 1988). In this study, perceived anxiety refers to an individual’s tendency to feel uneasy or fearful about the current or potential use of GenAI (Sallam et al., 2025). This anxiety may be exacerbated by the need for users to learn new terminology and understand unfamiliar applications associated with GenAI (Caporusso, 2023). Also, issues such as user privacy breaches and virus intrusions increase the risks of internet usage, further intensifying learners’ anxiety (Barbeite and Weiss, 2004). Existing research has indicated a negative correlation between perceived anxiety and the frequency of GenAI use (Zhu et al., 2024). Therefore, perceived anxiety is a critical factor influencing the use of GenAI.

In this study, information system quality is defined as the extent to which the information generated by GenAI accurately conveys the intended meaning (Baabdullah, 2024). It encompasses the technical performance of the system, the quality of the generated content, and the effectiveness of its practical applications. Mun and Hwang (2024) indicated that the information system quality of ChatGPT significantly influenced learners’ intention to use them. Similarly, Tlili et al. (2023) found that the reliability and flexibility of GenAI were crucial for ensuring stable and efficient operation. Furthermore, the content quality of GenAI is reflected in the accuracy and completeness of its generated information, which is intended to provide users with high-quality responses (Goli et al., 2023). For example, a high level of information system quality can effectively encourage learners to engage with live chat features during the learning process (McLean and Osei-Frimpong, 2019). Therefore, information system quality is a crucial determinant of GenAI-assisted learning.

An interactive learning environment refers to a setting that enhances learning experience by facilitating interactions among learners, instructors, peers, and learning systems (Liaw, 2008). GenAI fosters interactive learning by responding to learners’ queries, which lies at the core of interactive engagement (Rospigliosi, 2023). In GenAI-assisted learning, synchronous and asynchronous features create dynamic communication channels. Synchronous communication enables real-time interaction between teachers and students, while asynchronous communication supports flexible participation without simultaneous engagement (Sharma et al., 2007). These interaction modes not only enable learners to share information but also allow them to access valuable knowledge autonomously. Because learning inherently occurs in social contexts, interactions among learners, instructors, and peers offer critical opportunities for knowledge construction (Liaw and Huang, 2013). Consequently, an interactive learning environment is essential for GenAI-assisted learning.

The usefulness of learning systems is recognized as a critical factor influencing learning performance (Virvou and Katsionis, 2008). In this study, perceived usefulness is defined as the extent to which learners perceive GenAI as effective, efficient, and satisfying in supporting their learning (Kim et al., 2025). Prior research has shown that interactive learning environments were key determinants of learners perceived usefulness in e-learning systems (Liaw and Huang, 2007). Börekci and Uyangör (2025) found that perceived self-efficacy significantly impacted perceived usefulness. Zheng W. et al. (2024) and Zheng Y. et al. (2024) also suggested that the information system quality significantly affected perceived usefulness. Furthermore, Sallam et al. (2025) demonstrated a negative correlation between perceived anxiety and perceived usefulness. Based on these findings, we propose the following hypotheses:

Perceived satisfaction is considered a key indicator of the effectiveness of learning systems (Virvou and Katsionis, 2008). In this study, it is defined as the degree of comfort and contentment learners experience in using GenAI (Kim et al., 2025). Existing studies have shown that perceived self-efficacy, information system quality, interactive learning environments, and perceived usefulness significantly influenced perceived satisfaction (Kim and Ong, 2005; Liaw, 2008). Specifically, Liaw et al. (2007) found that perceived self-efficacy played an important role in shaping learners’ satisfaction with e-learning systems. Liu et al. (2009) reported that learner satisfaction increases when e-learning environments provide more interactive activities. Further, the information system quality is closely associated with learners’ satisfaction (Almufarreh, 2024; Wut and Lee, 2022). Al-Sharafi et al. (2023) also shown that learners are satisfied with GenAI-assisted learning when the tools meet their needs. Moreover, prior studies have identified a negative correlation between perceived anxiety and perceived satisfaction (Sun et al., 2008; Tsai, 2009). Based on these findings, the following hypotheses were formulated:

Existing research has found a strong correlation between perceived satisfaction and perceived self-regulation (Kramarski and Gutman, 2006). For example, Ji et al. (2025) suggested that perceived satisfaction enhanced learners’ self-regulation, which in turn increases their intention to continue using GenAI. Zhou et al. (2024) also found that perceived usefulness was a crucial factor in promoting self-regulation in GenAI-assisted learning environments. Vighnarajah et al. (2009) also suggested that interactive e-learning environment could enhance learners’ self-regulation. Therefore, the following hypotheses were proposed:

This study defines behavioral intention as learner’s willingness to continue using GenAI (Cai et al., 2023). Existing studies have suggested that perceived satisfaction and usefulness could significantly influence learners’ behavioral intention in GenAI-assisted learning (Jung and Jo, 2025; Mohamed Eldakar et al., 2025). For example, Al-Sharafi et al. (2023) found that when students perceive AI chatbots as effective in assisting their learning tasks, their satisfaction also increases, enhancing their likelihood of continued use. Furthermore, Ma and Lee (2019) found that learners perceived usefulness and self-regulation were positively correlated with their behavioral intentions. Based on the above findings, we proposed the following hypotheses and research model (Figure 2).

This study aimed to systematically test the hypotheses of the research model using a survey methodology. The questionnaire comprised three sections, with the first section collecting participants’ demographic information, including gender, age, educational level, major, and experience using GenAI. The second section comprised eight scales, each corresponding to a construct in the research model. The third section included an open-ended question designed to elicit participants’ views on the advantages, disadvantages, and suggestions for GenAI-assisted learning. This section was primarily based on the frameworks developed by Liaw (2008) and Liaw and Huang (2013), with relevant items revised to fit the context of this study. Therefore, the final research instrument consisted of 29 items measured on a seven-point Likert scale, ranging from 1 (strongly disagree) to 7 (strongly agree). Items numbered 1–3 evaluated perceived self-efficacy. Items 4–7 measured perceived anxiety. The information system quality was gauged through Items 8–11. Items 12–15 were formulated to assess the interactive learning environments, whereas Items 16–18 appraised their perceived satisfaction. Items 19–21 quantified the perceived usefulness. Items numbered 22–25 evaluated perceived self-regulation. Lastly, Items 26–29 determined the behavioral intention. The final version of the questionnaire is presented in Table A1.

Before the formal survey, a pilot survey was conducted to ensure the adaptability of the questionnaire items to the actual survey. However, the discriminate validity was not satisfactory between perceived self-efficacy and perceived satisfaction. Therefore, one of the items related to perceived self-efficacy was removed according to the results. Another item related to perceived usefulness was also removed due to its unacceptable factor loading.

To further improve the scientific validity of the scale, two experts with substantial experience in higher education and educational technology were invited to review the content of the questionnaire. The first expert is a professor specializing in technology integration in teacher education, with over 20 years of experience in developing and validating research instruments. The second expert is a senior researcher in psychometrics who has collaborated on several large-scale studies involving survey validation. These experts provided detailed feedback on the relevance, clarity, and validity of the questionnaire items, ensuring alignment with the study’s objectives and the theoretical framework. For example, in this study, the definition of “use” was based on the expected frequency of participants’ use of GenAI in current learning, rather than their future use. All items were carefully screened and adjusted for content validity. The items were then subjected to a back-translation validation process to ensure cross-cultural consistency and linguistic accuracy. Specifically, one researcher translated from English to Chinese and then another researcher translated the Chinese back to English.

This study was conducted in Jiangsu Province, an economically developed coastal province in eastern Mainland China. Participants included undergraduate and graduate students majoring in math, etc. The Questionnaire Star website¹ was used to pose the questionnaire. Counsellors from the school of mathematics at several universities in Jiangsu assisted us in sharing the link to the questionnaire with their university students via WeChat and Tencent QQ. While these recruitment methods helped us efficiently reach participants, it is important to acknowledge that these platforms may limit the diversity of the sample. WeChat and Tencent QQ are popular among younger, tech-savvy students, which may lead to an overrepresentation of certain demographic groups, such as more active and engaged students. Additionally, counselors from specific universities may have led to a concentration of participants from a limited number of institutions, potentially affecting the generalizability of the findings to students from other provinces or fields of study. At the beginning of the questionnaire, respondents were informed that they have the right to withdraw, confidentiality, anonymity and data protection. Participants voluntarily completed the survey and were entered into a small gift card lottery as appreciation for their participation. They were also informed about anonymity issues, explaining that the data would be used solely for research purposes. Ethical guidelines approved by the first author’s university academic committee were followed during data collection. Ethical approval was obtained under protocol number yzumath20240801 on August 1, 2024.

After data cleaning, the final analysis included data from 607 university students (e.g., prospective mathematics teachers) with experience using GenAI, aged 18–27. Participants were selected based on their use of GenAI, with popular tools being Baidu’s Ernie Bot and OpenAI’s ChatGPT. The demographic characteristics show that 37.4% of the respondents identity as males and 57.3% of the respondents are undergraduates. Additional details regarding the demographic characteristics of the respondents are presented in Table 1, with descriptive statistics.

While we believe that these samples provide valuable insights, we also acknowledge that these recruitment methods introduce the possibility of selection bias. These samples may not fully represent the broader university student population, especially in terms of demographic diversity. Future work could benefit from expanding the recruitment methods to include a more varied range of platforms and institutions, which would enhance the representativeness of the sample and improve the external validity of the findings.

Once data collection was completed, we downloaded the data sets from the WJX platform for further analysis. The data were stored in SPSS 26 and saved in SAV format for normality testing.

Skewness and kurtosis indices were examined to assess the normality of each item, which is a prerequisite for SEM analysis. A dataset is generally considered normally distributed when the absolute skewness is below 3 and the absolute kurtosis is below 10 (Kline, 2023). The distribution of all scale items was examined using SPSS 26. Then, SPSS 26 and AMOS 28 were employed to evaluate the reliability and validity of the measurements and to test the proposed structural model. We also conducted a confirmatory factor analysis (CFA) using AMOS 28 to assess the measurement model. Prior to the CFA, the Kaiser–Meyer–Olkin (KMO) measure and Bartlett’s test of sphericity were examined to ensure sampling adequacy. The model included eight latent variables, and standardized factor loadings were evaluated to confirm construct validity. A factor loading threshold of 0.50 was adopted, and items with loadings below this threshold were removed to improve the reliability and validity of the measurement model.

Subsequently, Cronbach’s alpha (α), Average Variance Extracted (AVE), and Composite Reliability (CR) were calculated for each variable based on the retained items. Specifically, Cronbach’s α should exceed 0.80, the AVE should be greater than 0.50 and the CR should be above 0.70 (Hair et al., 2012). To assess the discriminant validity of the constructs, we employed the “Validity and Reliability Test” plugin in Amos, which incorporates both the Fornell–Larcker criterion and the Heterotrait–Monotrait (HTMT) ratio of correlations.

After confirming the validity and reliability of the constructs, we used Amos 28 to perform standardized model estimation via the maximum likelihood method. The analysis generated path coefficients (β) and significance levels (p), which were used to test the research hypotheses and evaluate the mediating effects. According to Baron and Kenny (1986), mediating effects can be classified into three types: full mediation, partial mediation, and suppressing mediation. In this study, mediation effects were tested using the bias-corrected and accelerated bootstrap method, with 5,000 resamples, as recommended in recent SEM practice (Hu et al., 2025; Tang et al., 2025). The explanatory power of the model for the outcome variables was evaluated by calculating the squared multiple correlations (R²). According to Hair and Alamer (2022), R² values of 0.25, 0.50, and 0.75 are considered to represent weak, moderate, and substantial explanatory power, respectively.

Model fit indices in this study were evaluated based on widely accepted criteria, with the chi-square divided by degrees of freedom (χ²/df, CMIN/DF) expected to be ≤ 3.0 (Hayduk, 1987). The root mean square error of approximation (RMSEA) was expected to be ≤ 0.05 (Hair et al., 2017). The incremental fit index (IFI), Tucker–Lewis index (TLI), and comparative fit index (CFI) were required to be ≥ 0.90 (Bagozzi and Yi, 1988). The standardized root means square residual (SRMR) was considered acceptable if ≤ 0.08 (Hu and Bentler, 1999).

To gain deeper insight into learners’ experiences with GenAI-assisted learning, we included one open-ended interview question at the end of the questionnaire: “Please tell us about the advantages, disadvantages and suggestions of GenAI.” This question was designed to elicit participants’ reflective opinions about GenAI, including perceived strengths, limitations, and potential improvements. The goal was to capture the subjective aspects of their experiences that might not be fully revealed through quantitative scales.

We conducted a thematic analysis on the open-ended responses to explore learners’ perceptions. All responses were reviewed and manually coded into three main themes—advantages, disadvantages, and suggestions—as aligned with the structure of the prompt. Each theme was further categorized into sub-themes based on lexical frequency and semantic similarity. Representative expressions were extracted for illustration.

The coding process was conducted using NVivo 12, and all codes were discussed and validated by two researchers to enhance reliability. The process involved iterative comparison and resolution of discrepancies through discussion.

The results of the normality test are presented in Table 2. The skewness values ranged from −0.921 to 0.161, and the kurtosis values ranged from −1.014 to 1.890. These results indicate that the skewness and kurtosis of the dataset fall within acceptable limits, suggesting that all 29 items across the eight variables approximate a normal distribution.

The proposed model demonstrated an excellent fit to the collected data. As shown in Table 3, the model fit indices were as follows: CMIN/DF = 1.865, RMSEA = 0.038, IFI = 0.973, TLI = 0.969, CFI = 0.973, and SRMR = 0.040. All six indices met the recommended thresholds, indicating a satisfactory model fit.

As shown in Table 4, all items were satisfactorily loaded onto the eight factors corresponding to the proposed variables. Both the composite reliability (CR) and average variance extracted (AVE) values for all constructs exceeded the recommended thresholds. Additionally, the CFA results indicated a KMO value of 0.934, and Bartlett’s test of sphericity was significant (χ² = 11,507.763, p < 0.001). Cronbach’s alpha values for all variables ranged from 0.811 to 0.946, indicating satisfactory internal consistency. Collectively, these results confirmed that the measurement model met the prerequisites for subsequent SEM analysis.

Furthermore, The Fornell-Larcker criterion and the Heterotrait-Monotrait Ratio (HTMT) were employed to assess discriminant validity. As shown in Table 5, the square roots of the AVE for each construct exceeded the corresponding inter-construct correlations, indicating satisfactory discriminant validity (Fornell and Larcker, 1981). This result indicated that the study achieved satisfactory discriminant validity according to the Fornell–Larcker criterion. To further confirm these findings, the Heterotrait–Monotrait (HTMT) ratio of correlations was calculated, reflecting the average inter-item correlations across constructs compared with those within the same construct (Henseler et al., 2015). As shown in Table 6, all HTMT values were below the recommended threshold of 0.90, indicating acceptable discriminant validity (Hair et al., 2019).

The explanatory power of the hypothesized model is presented in Table 7. The R² values for perceived usefulness, perceived satisfaction, perceived self-regulation, and behavioral intention were 0.625, 0.642, 0.590, and 0.496, respectively. The results indicate that the model demonstrates moderate to strong explanatory power for perceived usefulness, perceived satisfaction, and perceived self-regulation, while the relatively lower R² value for behavioral intention. This suggests that the model accounts for only about half of the variance in behavioral intention, leaving a significant portion unexplained. The unexplained variance may be attributed to additional factors that were not considered in the current model. For example, social influence (e.g., peers, colleagues, or societal norms) may play a significant role in shaping students’ behavioral intentions, especially in technology adoption contexts (Zheng W. et al., 2024; Zheng Y. et al., 2024). Furthermore, platform trust could be another critical factor, as users’ trust in the GenAI platform may strongly affect their intention to use it (Wang et al., 2025). Also, perceived ease of use and previous experience with similar technologies could contribute to explaining the remaining variance in behavioral intention (Unal and Uzun, 2021). These factors, which were not included in this study, could provide a more comprehensive understanding of the factors influencing behavioral intention.

Overall, while the model shows a strong fit for several key variables, the relatively lower R² for behavioral intention highlights the need for further exploration of other potential influences, such as social influence and platform trust, that might contribute to the formation of behavioral intention.

Additionally, the results of the path analysis showed that information system quality and interactive learning environments contributed more to perceived usefulness than perceived self-efficacy and perceived anxiety. Information system quality and perceived usefulness explained and predicted perceived satisfaction more significantly than perceived self-efficacy, perceived anxiety, and interactive learning environments. Second, perceived usefulness had a greater impact on perceived self-regulation than perceived satisfaction and interactive learning environments. Perceived usefulness and perceived self-regulation emerged as stronger predictors of behavioral intention than perceived satisfaction.

Table 8 presents the mediating effects of perceived satisfaction, perceived usefulness and perceived self-regulation. The path of perceived anxiety on perceived satisfaction was not significant, as were the paths of perceived satisfaction and perceived usefulness on behavioral intention. All other mediating paths exhibited partial mediation effects. Figure 3 illustrates the final structural model.

Table 9 shows the key themes that emerged from the qualitative analysis. For advantages, learners frequently highlighted intelligent responses, fast and efficient interaction, and personalized learning support. Disadvantages focused on dependency on the tool, potential inaccuracies, and operational difficulties such as needing precise prompts. Suggestions included calls to enhance efficiency, improve personalization, strengthen accuracy, and promote broader awareness and usability of GenAI.