Unraveling EFL students’ technology self-efficacy, emotions and behavioral intentions in leveraging large language models: the serial mediation of perceptions and attitudes Authors: Zhaoyang Shan, Manman Shan, Lanyun Ding, Yonggang Su Published: 2026 ABSTRACT AbstractWith the growing momentum to integrate artificial intelligence (AI) into education, large language models (LLMs) have shown great potential for English as a Foreign Language (EFL) learning. However, the relationships between EFL students’ technology self-efficacy, emotions toward LLM-assisted English learning, and their behavioral intentions remain underexplored. Guided by the Technology Acceptance Model (TAM), this study explores how technology self-efficacy and emotions (i.e., perceived enjoyment and technology anxiety) influence behavioral intentions, with perceptions (i.e., perceived ease of use and perceived usefulness) and attitudes serving as mediating variables among Chinese EFL university students. A total of 677 valid questionnaires were collected and structural equation modeling (SEM) was employed for data analysis. The results revealed that: (a) technology self-efficacy positively predicted perceived enjoyment while negatively predicting technology anxiety; (b) technology self-efficacy positively predicted perceptions, which in turn predicted attitudes and ultimately behavioral intentions; (c) perceived enjoyment positively predicted perceived usefulness, which subsequently predicted attitudes and ultimately behavioral intentions; and (d) technology anxiety negatively predicted perceived ease of use, which in turn predicted attitudes and ultimately behavioral intentions. Overall, this study validated an extended TAM framework in the context of LLM-assisted language learning, showing how cognitive and emotional factors influence behavioral intentions through perceptions and attitudes in a serial manner. These findings offer valuable insights for integrating LLMs into foreign language learning.Supplementary InformationThe online version contains supplementary material available at 10.1186/s40359-026-04812-y.Keywords: Technology self-efficacy, Perceived enjoyment, Technology anxiety, Large language models, English learning, Structural equation modeling Introduction Large language models (LLMs) are generative AI systems capable of generating and processing human-like text by learning from large datasets [1]. Their advanced capabilities in comprehension, production, and interaction have garnered attention in artificial intelligence in education (AIED), particularly in language learning [2]. Tools such as OpenAI’s ChatGPT, Baidu’s ERNIE Bot, Doubao, and DeepSeek are reshaping second language (L2) learning by offering personalized, interactive learning environments [3]. As English remains the most widely learned foreign language globally, LLMs are increasingly integrated into English learning for practice and support [4, 5]. Despite the rising scholarly interest, most research on LLMs in foreign language education remains conceptual or qualitative [6, 7], with limited quantitative evidence on the factors influencing English as a Foreign Language (EFL) learners’ adoption of LLMs in English learning. The Technology Acceptance Model (TAM) [8] provides a widely used framework for deciphering technology adoption, including core variables (e.g., perceived usefulness), outcome variables (e.g., behavioral intention), and external variables (e.g., subjective norm), with the external and outcome variables mediated by the core variables [9]. While other frameworks, such as the Unified Theory of Acceptance and Use of Technology (UTAUT) [10] and the Stimulus-Organism-Response (SOR) model [11], also provide useful insights, TAM is preferred here due to its simplicity, focus on key variables, and successful application in similar research contexts, making it a more suitable choice for this study. TAM has been applied across domains such as healthcare [12] and classroom instruction [13]. In language learning, it has been employed to investigate students’ perceptions and adoption of language learning applications and platforms [14]. Considering the external variables in technology acceptance, emerging studies have highlighted that cognitive factors such as technology self-efficacy (TSE) [15], along with emotional factors including perceived enjoyment (PE) [16] and technology anxiety (TA) [17], are crucial to learners’ adoption of language learning technologies. Additionally, research has further demonstrated that perception-related variables such as perceived ease of use (PEU), and attitude toward use (ATU) can predict behavioral intention (BI) to adopt AIED tools [18–20]. It can thus be acknowledged that existing studies have yet to integrate the three core components (TSE, TA, and PE) to comprehensively reveal how cognitive and emotional factors influence EFL students’ BI to use LLMs in English learning. Moreover, research is still lacking on the potential serial mediating roles of perceptions and attitudes between cognition, emotions, and BI in LLM-based English learning. Grounded in TAM, this study synthesizes TSE, PE and TA to explore how EFL university students’ cognition and emotions influence their BI to utilize LLMs in English learning, while also investigating the serial mediating effects of PU, PEU, and ATU through structural equation modeling (SEM). Although prior TAM-based research has increasingly incorporated emotional variables, these factors are often examined in isolation from cognitive beliefs and primarily within conventional educational technologies. The present study addresses this gap by emphasizing the need for an integrated cognitive-affective perspective, particularly in the context of LLMs, where users’ experiences are inherently dynamic, interactive, and emotionally contingent. By extending TAM in this way, this study advances theoretical understanding of technology acceptance by integrating cognitive and emotional factors and clarifying the serial mediating roles of PU, PEU, and ATU in LLM-assisted learning contexts. Practically, the findings offer actionable implications for educators, instructional designers, and policymakers seeking to align LLM-based tools with learners’ psychological needs and pedagogical goals. At a broader level, the findings contribute to ongoing discussions on the future of language education by illustrating how LLM use shapes learners’ cognitive and emotional experiences, thereby providing insights into the evolving role of generative AI in language learning. In addition, this study contributes to the growing body of research on the role of emotions in language learning and teaching by demonstrating how emotional factors interact with cognitive beliefs to influence technology acceptance. The study also responds to recent calls for examining the psychological and contextual factors shaping the future of language teaching and learning [21]. Literature review The evolution of TAM TAM, initially proposed by [22], has evolved significantly over the past decades. Its core constructs, perceived ease of use (PEU) and perceived usefulness (PU), explain users’ behavioral intentions (BI) and actual use (AU). Attitude toward use (ATU) mediates the relationships between PEU and BI as well as PU and BI, ultimately influencing AU. Both PU and PEU shape users’ ATU [8], as shown in Fig. 1, and the constructs are defined in Table 1. TAM is based on the Theory of Reasoned Action (TRA), which posits that attitudes and beliefs about a behavior influence intentions to perform it [23]. In TAM, PU and PEU guide attitudes and intentions while ATU reflects the affective dimension of technology use [22]. The framework categorizes core variables (PEU, PU, ATU), outcome variables (BI, AU), and external variables such as facilitating conditions [9]. Over time, TAM has incorporated additional external variables, including social norms [24], self-efficacy [25], emotions [26, 27], and social presence [28], to better explain technology adoption, especially in educational contexts. TAM [22] Table 1Definitions of core constructs [22]ConstructDefinitionsBIA user’s likelihood to engage in using a technology.ATUA user’s positive or negative evaluation of using a technology.PUThe degree to which a user believes a technology would enhance his/her performance.PEUThe degree to which a user believes using a technology would be free of effort. Definitions of core constructs [22] In educational research, TAM has been utilized to explore the application of e-learning platforms and materials among teachers and students. In terms of platform types, previous studies have examined a variety of e-learning platforms, such as mobile-assisted learning tools and AI-mediated systems [29–33]. Specifically, a study investigated 557 Chinese EFL university students’ continued use of mobile-assisted language learning tools [30], while other research examined 234 Iranian EFL teachers’ acceptance of ChatGPT for English teaching [29]. Similarly, another study of 97 Iraqi EFL teachers investigated how mobile learning served as a valuable resource for their professional development in higher education during the COVID-19 pandemic [32]. Considering learning outcomes, studies have found that technology-mediated language learning can significantly affect students’ learning performance [34–36]. For example, wearable VR-supported language learning tools have been shown to improve vocabulary retention [35], whereas communication apps like LINE and WhatsApp, with LINE primarily enhancing receptive skills (listening/reading) and WhatsApp improving productive skills (speaking/writing) among university students [36]. Moreover, a study reported that VR-based collaborative learning strategies could enhance overall learning achievement [34]. Furthermore, researchers have also explored different categories of influencing factors of technology acceptance by extending TAM [13, 27, 37]. These studies not only considered the classic constructs (PU, PEU, ATU, and BI), but also incorporated external variables such as task-technology fit [13], learning motivation [37], and emotional experience [27]. TAM holds a prominent position in research on language learning technologies, but its limitations have drawn increasing attention. TAM’s explanatory power for emotional factors and contextual variables is relatively limited [38]. Although various studies have begun incorporating emotional factors into extensions of TAM [16, 29], these factors are often treated as external variables in isolation, overlooking the interaction between cognitive factors (e.g. self-efficacy) and emotional factors (e.g., enjoyment and anxiety). Cognition and emotion, as key psychological determinants, are crucial in influencing learners’ decision-making processes [39], yet prior research has largely failed to explore how these factors interrelate and jointly impact technology acceptance, particularly in the context of foreign language learning. This oversight presents a critical gap in the literature, especially regarding the integration of LLMs in English language education. As an emerging AI technology, LLMs remain underexplored within the TAM framework, particularly the influencing factors and mechanisms of LLM acceptance among university students for English learning. A comprehensive examination incorporating both cognitive (e.g., self-efficacy) and emotional factors (e.g., enjoyment and anxiety) is notably scarce, leaving a critical gap in understanding of the psychological determinants of LLM integration in EFL contexts. Therefore, this study adopts TAM as the theoretical framework to scrutinize the relationships among EFL university students’ cognition, emotions, perceptions, attitudes and behavioral intentions to use LLMs for English learning. Specifically, this study further extends TAM with serial mediation analysis, exploring how cognitive (e.g., self-efficacy) and emotional factors (e.g., enjoyment and anxiety) influence perceptions of LLMs, which then shape attitudes and behavioral intentions in using LLMs for English learning. This approach enhances TAM’s explanatory power in the context of LLM adoption for EFL learning. Students’ technology self-efficacy in language learning Self-efficacy, a core concept in social cognitive theory, refers to “beliefs in one’s capabilities to organize and execute the courses of action required to produce given attainments” [40]. As a pivotal cognitive variable, it encapsulates students’ self-assessment of their ability to complete specific tasks [41]. In the field of educational technology, technology self-efficacy (TSE) was defined as “the belief in one’s ability to successfully perform a technologically sophisticated new task” [42]. TSE has been a critical factor for students in the acceptance and application of language learning technology [35, 43–46]. Research has found that AI self-efficacy positively predicts AI usage by reducing AI anxiety, and improving attitudes toward AI in the context of L2 learning [43]. Similarly, self-efficacy has been shown to be a critical determinant in students’ use of a wearable virtual reality English learning platform for specific purposes, suggesting educators should enhance students’ TSE to encourage their adoption of innovative technologies [35]. Another study further established that computer self-efficacy positively influences learners’ BI in foreign language learning [46]. Moreover, TSE not only affects students’ PEU, but also enhances their PU, thereby creating a positive technology adoption behavior in language learning [47–49]. A study of 75 Saudi EFL undergraduate learners revealed significant positive correlations between computer self-efficacy and key technology acceptance variables (i.e., PU and PEU) [47]. Then, another study involving 348 Chinese undergraduates during COVID-19 demonstrated that students’ computer self-efficacy positively influenced both PU and PEU of the technology [48]. A study of 303 Malaysian undergraduates in higher education elucidated TSE as a determinant factor for PU and PEU in technology-mediated learning environments [49]. Moreover, it is important to distinguish between TSE and language self-efficacy (LSE). LSE refers to learners’ beliefs in their ability to successfully perform language-specific tasks, such as speaking, writing, reading, and listening [50]. Research has shown that LSE also plays a significant role in language acquisition and can influence learners’ motivation, emotion, and overall language learning performance [51]. While TSE focuses on learners’ confidence in using technology for language learning, LSE refers to their beliefs in their ability to perform specific language-related tasks. In all, it is reasonable to hypothesize that TSE might directly influence both PE and PEU when EFL university students utilize LLMs for English learning. TSE as an external variable is a crucial leverage point for educational interventions, and incorporating TSE into an extended TAM framework helps comprehensively understand its effects on EFL students’ use of LLMs for English learning. Students’ emotional responses in technology-assisted language learning Research in the field of language learning indicates that emotions can significantly affect learners’ willingness to communicate [52, 53], engagement and performance [54]. Moreover, emotions are widely recognized as a foundational component of motivation in language learning [55], serving as an intensifier that energizes and directs learners’ behaviors toward their goals, thereby underscoring the critical importance of emotions in language learning processes [56, 57].Within educational technology contexts, various emotions related to technology use are equally influential [58]. Perceived enjoyment (PE) and technology anxiety (TA) represent the most extensively investigated emotional factors in educational technology research. PE is an intrinsic driver that promotes the adoption of new information technologies [59]. PE describes the degree to which individuals enjoy using a system and perceive it as contributing to their expected gains [59]. In other words, it reflects the degree of pleasure students derive from interacting with the educational technology. Several studies have revealed that PE can significantly influence PU [16, 17, 29]. One study found that PE influences the use of educational technology tools indirectly through PU [16], while another confirmed a significant direct impact of PE on PU in technology adoption for language learning [17]. Further evidence indicated that PE is a crucial determinant of PU [29]. Therefore, it is reasonable to posit that EFL university students’ PE in using LLMs for English learning could affect PU. Integrating the external variable of PE into TAM can effectively explore whether EFL students’ PU of LLMs is enhanced by a sense of enjoyment when using LLMs for English learning, thereby improving the explanatory power of their technology adoption behavior. TA describes students’ anxiety about using intelligent tutoring systems and concerns over academic performance [17]. Research has shown that TA can negatively affect PEU by increasing anxiety, when students interact with educational technology [17, 60, 61]. A study involving 528 senior secondary students in China validated that students’ TA reduces their PEU of intelligent tutoring systems for English learning [17]. Similarly, another study involving 245 Chinese university students revealed that computer anxiety significantly influences PEU when utilizing automated writing evaluation systems [60]. A study of 347 Chinese middle school students in K12 educational settings further demonstrated that students’ TA negatively affects their PEU of tablet computers [61]. As a result, it is reasonable to believe that EFL learners’ TA might negatively influence their PEU of LLMs for English learning. Incorporating TA as an additional external variable into TAM allows for a more comprehensive understanding of the psychological barriers that impede EFL learners’ BI when adopting LLMs-powered educational technologies. It is noteworthy that emotional responses, such as anxiety and enjoyment, are also linked to self-efficacy beliefs [62, 63]. This relationship can be theoretically explained by the Control-Value Theory (CVT) of achievement emotions [64], which posits that emotions arise from individuals’ appraisals of control over learning activities. In this regard, self-efficacy can be understood as a key source of control appraisal, whereby learners with higher self-efficacy are more likely to perceive greater control, experience more positive emotions (e.g., enjoyment), and less negative emotions (e.g., anxiety). Previous studies have demonstrated that TSE exhibits a positive correlation with PE [65, 66] while showing a negative association with TA [42, 67]. In other words, when using LLMs for English learning, EFL students with higher TSE, reflecting greater perceived control over the technology, are more likely to experience positive emotions (e.g., higher PE) and less likely to suffer from negative emotions (e.g., lower TA). Therefore, this study extends TAM by incorporating cognitive (e.g., TSE) and emotional (e.g., PE and TA) factors as the external variables to systematically investigate the psychological mechanisms underlying EFL university students’ BI to use LLMs for English learning. Hypothesized model development Based on the TAM framework and the above literature review, a comprehensive hypothesized model has been proposed, depicting the relationships among one cognitive variable (TSE), two emotional variables (PE and TA), two perception variables (PEU and PU), as well as ATU and BI, as illustrated in Fig. 2. The hypothesized model Research indicates that TSE not only positively influences students’ PEU but also enhances their PU, which in turn fosters positive technology adoption behavior in language learning [47–49]. Moreover, high TSE encourages positive emotions such as PE during usage [65, 66] while simultaneously reducing negative emotions like TA [43]. Based on these insights, this study proposes the following hypotheses in the context of university students’ use of LLMs for English learning: Hypothesis 1. EFL learners’ TSE positively affects their PU. Hypothesis 1. EFL learners’ TSE positively affects their PU. Hypothesis 2. EFL learners’ TSE positively affects their PEU. Hypothesis 2. EFL learners’ TSE positively affects their PEU. Hypothesis 3. EFL learners’ TSE positively affects their PE. Hypothesis 3. EFL learners’ TSE positively affects their PE. Hypothesis 4. EFL learners’ TSE negatively affects their TA. Hypothesis 4. EFL learners’ TSE negatively affects their TA. Research suggests that PE enhances users’ PU in language learning [16, 29]. On the other hand, TA, a negative emotional state, can impede users’ PEU for learning [17, 61]. Based on these findings, this study proposes the following hypotheses: Hypothesis 5. EFL learners’ PE positively affects their PU. Hypothesis 5. EFL learners’ PE positively affects their PU. Hypothesis 6. EFL learners’ TA negatively affects their PEU. Hypothesis 6. EFL learners’ TA negatively affects their PEU. According to the TAM, PEU is believed to positively influence PU, as users are more likely to perceive a technology as useful when it is easy to use [22]. Furthermore, PEU has been shown to directly shape users’ ATU by influencing their perception of the technology’s value in their context [68]. PU, in turn, has a direct impact on users’ ATU, as individuals tend to develop a more favorable attitude toward technology when they perceive it as useful [8]. A positive attitude toward technology is a significant predictor of BI, as individuals with a more positive ATU are more likely to intend to use the technology [69]. Based on these theoretical underpinnings, and considering the role of LLMs in supporting university students’ English learning, the following hypotheses are proposed: Hypothesis 7. EFL learners’ PEU positively affects their PU. Hypothesis 7. EFL learners’ PEU positively affects their PU. Hypothesis 8. EFL learners’ PEU positively affects their ATU. Hypothesis 8. EFL learners’ PEU positively affects their ATU. Hypothesis 9. EFL learners’ PU positively affects their ATU. Hypothesis 9. EFL learners’ PU positively affects their ATU. Hypothesis 10. EFL learners’ PU positively affects their BI. Hypothesis 10. EFL learners’ PU positively affects their BI. Hypothesis 11. EFL learners’ ATU positively affects their BI. Hypothesis 11. EFL learners’ ATU positively affects their BI. Technology acceptance in educational contexts is influenced by various cognitive and emotional factors. According to the TAM, core variables such as PU, PEU, and ATU mediate the relationship between external variables and outcome variables (i.e., BI) [9]. TSE, as a cognitive attribute, has been well established to play a crucial role in shaping learners’ perceptions and their overall acceptance of technology [48, 49]. Similarly, PE, the enjoyment learners experience while using technology, also plays a significant role in shaping their intentions to use educational tools [16]. Additionally, TA is another key emotional factor that negatively impacts learners’ BI to use technology [17]. Based on these findings and the TAM framework, TSE, PE, and TA are proposed as external variables in measuring university students’ use of LLMs for English learning. The following hypothesized paths and relationships are proposed: Hypothesis 12. EFL learners’ PU mediates the relationship between their TSE and BI. Hypothesis 12. EFL learners’ PU mediates the relationship between their TSE and BI. Hypothesis 13. EFL learners’ PU mediates the relationship between their PE and BI. Hypothesis 13. EFL learners’ PU mediates the relationship between their PE and BI. Hypothesis 14. EFL learners’ PU and ATU play a serial mediating role in the relationship between their TSE and BI. Hypothesis 14. EFL learners’ PU and ATU play a serial mediating role in the relationship between their TSE and BI. Hypothesis 15. EFL learners’ PEU and ATU play a serial mediating role in the relationship between their TSE and BI. Hypothesis 15. EFL learners’ PEU and ATU play a serial mediating role in the relationship between their TSE and BI. Hypothesis 16. EFL learners’ PU and ATU play a serial mediating role in the relationship between their PE and BI. Hypothesis 16. EFL learners’ PU and ATU play a serial mediating role in the relationship between their PE and BI. Hypothesis 17. EFL learners’ PEU and ATU play a serial mediating role in the relationship between their TA and BI. Hypothesis 17. EFL learners’ PEU and ATU play a serial mediating role in the relationship between their TA and BI. Hypothesis 18. EFL learners’ PEU and PU play a serial mediating role in the relationship between their TA and BI. Hypothesis 18. EFL learners’ PEU and PU play a serial mediating role in the relationship between their TA and BI. These hypotheses are developed based on the proposed model and are intended to test the relationships among cognitive, emotional, perception variables, and attitudes, and how they ultimately influence behavioral intentions to use LLMs for English learning. To further explore the underlying mechanisms that govern these relationships, the following research questions have been formulated: RQ1: What are the relationships between TSE, emotions (PE and TA), perceptions (PEU and PU), ATU, and BI? RQ1: What are the relationships between TSE, emotions (PE and TA), perceptions (PEU and PU), ATU, and BI? RQ 2: Do perceptions and ATU serve as serial mediators between TSE, emotions, and BI? RQ 2: Do perceptions and ATU serve as serial mediators between TSE, emotions, and BI? Methodology Research design This study adopted a cross-sectional survey design to examine the relationships between cognitive and emotional factors and EFL university students’ BI to use LLMs for English learning. Given the complex interplay among multiple latent constructs (e.g., cognition, emotion, perception, and attitude), structural equation modeling (SEM) was employed as the primary analytical technique to test the hypothesized relationships and assess the proposed serial mediation mechanism within an integrated TAM framework. The study targeted undergraduate students from three universities in northern China across different academic disciplines (e.g., engineering, humanities, and social sciences) to ensure sample diversity and enhance the generalizability of the findings. Data were collected using a self-report questionnaire measuring students’ cognitive beliefs, emotional experiences, perceptions, attitudes, and behavioral intentions toward the use of LLMs, such as ChatGPT, Doubao, and Deepseek in English learning. The questionnaire was administered online to facilitate accessibility and minimize potential response bias. Participants A total of 966 questionnaires were collected from undergraduate students at three universities in northern China. Cases with substantial missing responses were excluded during data screening, and remaining missing values were handled using listwise deletion in SPSS 26.0. To ensure alignment with the research objectives, only participants who reported using LLMs to assist their English learning were included in the analysis. The final sample consisted of 677 valid responses (males = 321, females = 356), with ages ranging from 18 to 25 years (M = 21.5, SD = 1.80). A total of 256 participants were excluded because they had never heard of or used LLMs for English learning. Additionally, 33 participants were excluded for providing incomplete responses or completing the questionnaire in an unrealistically short time (e.g., less than one-third of the median completion time), suggesting inattentive responding. The final sample size of 677 exceeds common recommendations for SEM, which typically suggest 10–20 cases per estimated parameter, providing sufficient statistical power for analyzing the proposed model. Among the participant, 183 (27.2%) were from humanities, 146 (21.5%) from social sciences, 134 (19.7%) from natural sciences, and 214 (31.6%) from engineering. With respect to LLM use experience, more than half of the participants (n = 347, 51.3%) reported using LLMs for English learning on a daily basis, indicating regular engagement with these tools. In terms of task types, translation was the most frequently reported use (n = 282, 41.7%), followed by writing tasks (n = 212, 31.3%). Vocabulary and grammar learning accounted for 18.5% (n = 125), whereas conversation practice was the least common (8.5%, n = 58). These patterns suggest that LLMs are primarily employed for text-based and form-focused learning activities rather than interactive communication practice. Instruments The questionnaire consisted of two main sections: demographic information and survey constructs. The demographic section collected information on participants’ age, gender, major, and other relevant background information. For the measurement of the main constructs, all scales were adapted to the context of LLM-assisted English learning through minor wording modifications, without altering their underlying theoretical constructs. Prior to the formal survey, a pilot test was conducted with a small group of students to refine item wording, ensure clarity, and confirm the reliability of each scale, particularly for constructs with a limited number of items (i.e., three-item scales). The complete measurement instruments are provided in Appendix. TSE Scale Participants’ TSE was measured using a 3-item scale adapted from [70], assessing their confidence in using LLMs for English learning. Items were rated on a 5-point Likert scale ranging from 1 (‘strongly disagree’) to 5 (‘strongly agree’), with higher scores indicating greater TSE. Sample items include: “I am confident in using LLMs to learn English” and “I can independently solve problems encountered when using LLMs for English learning.” The scale demonstrated acceptable reliability in the present study (Cronbach’s α = 0.771). TA Scale TA was measured using a 3-item scale adapted from [71], assessing participants’ anxiety and discomfort when using LLMs for English learning. Items were rated on a 5-point Likert scale ranging from 1 (‘strongly disagree’) to 5 (‘strongly agree’), with higher scores indicating greater TA. Sample items include: “I feel uneasy when I think about trying to use LLMs for English learning” and “I hesitate to use LLMs for English learning for fear of making mistakes that I can not correct.” The scale showed good reliability (Cronbach’s α = 0.819). PE Scale PE was measured using a 4-item scale adapted from [72], capturing participants’ positive emotional experiences when using LLMs for English learning. Items were rated on a 5-point Likert scale ranging from 1 (‘strongly disagree’) to 5 (‘strongly agree’), with higher scores indicating greater PE. Sample items include: “I find using LLMs to learn English enjoyable” and “Using LLMs helps me explore new English knowledge.” The scale demonstrated high reliability (Cronbach’s α = 0.889). PEU and PU Scale PEU and PU were measured using 4-item scales adapted from [8]. PEU assessed participants’ perceptions of the ease of using LLMs for English learning, while PU measured their perceptions of the usefulness of LLMs in improving English learning outcomes. Items were rated on a 5-point Likert scale ranging from 1 (‘strongly disagree’) to 5 (‘strongly agree’), with higher scores indicating greater PEU and PU. Sample items include: “Using LLMs for English learning is very easy for me” (PEU) and “Using LLMs improves my English learning efficiency” (PU). The reliability coefficients were acceptable for both constructs (PEU: Cronbach’s α = 0.831; PU: Cronbach’s α = 0.847). ATU Scale ATU was measured using a 3-item scale adapted from [23], assessing participants’ overall evaluative feelings toward using LLMs for English learning. Items were rated on a 5-point Likert scale ranging from 1 (‘strongly disagree’) to 5 (‘strongly agree’), with higher scores indicating greater ATU. Sample items include: “Using LLMs for English learning is a good idea” and “I have a positive attitude toward LLM-assisted English learning.” The scale demonstrated good reliability (Cronbach’s α = 0.844). BI Scale BI was measured using a 4-item scale adapted from [59], assessing participants’ intention to continue using LLMs for English learning. Items were rated on a 5-point Likert scale ranging from 1 (‘strongly disagree’) to 5 (‘strongly agree’), with higher scores indicating greater BI. Sample items include: “I am willing to use LLMs to learn English” and “I will continue using LLMs for English learning in the future.” The scale showed good reliability (Cronbach’s α = 0.846). Data collection and analytic strategies Data were collected through an online survey platform (Sojump, https://www.wjx.cn/) over a two-month period (April–June 2025). Participants were recruited using a convenience sampling strategy through university networks and course-based distribution. The survey link was disseminated via instructors and student groups, and participation was entirely voluntary. Before commencing, the purpose and procedure of the study were clearly explained at the beginning of the questionnaire. Informed consent was obtained prior to participation, and respondents were assured of their right to withdraw at any time, as well as the anonymity and confidentiality of their responses. No monetary compensation was provided. To mitigate potential common method bias, several procedural remedies were implemented, including assuring anonymity, emphasizing that there were no right or wrong answers, and separating different construct sections in the questionnaire. Data analysis was conducted in three steps. Firstly, SPSS 26.0 was utilized for preliminary analysis, including descriptive and correlation analysis. Secondly, AMOS 24.0 was utilized for SEM to conduct confirmatory factor analysis (CFA), and validate the hypothesized research model (see Fig. 3). Model fit was assessed using multiple indices, including the comparative fit index (CFI), Tucker–Lewis index (TLI), goodness-of-fit index (GFI), normed fit index (NFI), root mean square error of approximation (RMSEA), and standardized root mean square residual (SRMR). Thirdly, to test the mediation effects, the bias-corrected nonparametric percentile bootstrap method was employed, and it can provide accurate estimates of indirect effects. Parameter estimation was performed using the maximum likelihood (ML) method. The structural equation model of the constructs. TSE Technology self-efficacy, TA Technology anxiety, PE Perceived enjoyment, PU Perceived usefulness, PEU Perceived ease of use, ATU Attitude toward use, BI Behavioral intention Results Preliminary analysis Table 2 presented the means, standard deviations, and correlation results among the constructs. The mean values indicated participants’ ratings for each variable were generally consistent, with scores clustering around similar levels. One-sample t-tests showed that all the rating scores of the constructs were higher than 3 (ts(676) ≥ 5.04, ps < 0.05), indicating the trend of participants’ agreement on these constructs. In addition, skewness values for all constructs ranged from − 0.200 to − 0.059, and kurtosis values ranged from − 0.585 to − 0.310. All values fell well within the acceptable thresholds, indicating that the assumption of normality was satisfied [73]. Table 2Descriptive statisticsConstructMeanSDTSETAPEPUPEUATUBITSE3.440.81TA3.150.88− 0.337**PE3.350.850.365**− 0.313**PU3.490.800.235**− 0.278**0.304**PEU3.440.820.400**− 0.359**0.390**0.316**ATU3.390.860.380**− 0.316**0.336**0.309**0.394**BI3.470.820.399**− 0.315**0.408**0.140**0.427**0.388**Skewness( < ± 2)-0.200-0.089-0.059-0.063-0.126-0.142-0.193Kurtosis( < ± 7)-0.310-0.585-0.473-0.477-0.409-0.476-0.482**p < 0.01 Descriptive statistics **p < 0.01 TSE showed positive correlations with PE (r = 0.365, p < 0.01), PU (r = 0.235, p < 0.01), PEU (r = 0.400, p < 0.01), ATU (r = 0.380, p < 0.01), and BI (r = 0.399, p < 0.01), while showing a negative correlation with TA (r=-0.337, p < 0.01). TA was negatively correlated with PE (r=-0.313, p < 0.01), PU (r=-0.278, p < 0.01), PEU (r=-0.359, p < 0.01), ATU (r=-0.316, p < 0.01), and BI (r=-0.315, p < 0.01). PE was positively correlated with PU (r = 0.304, p < 0.01), PEU (r = 0.390, p < 0.01), ATU (r = 0.336, p < 0.01), and BI (r = 0.408, p < 0.01). PU presented positive correlations with PEU (r = 0.316, p < 0.01), ATU (r = 0.309, p < 0.01), and BI (r = 0.140, p < 0.01). PEU was positively correlated with ATU (r = 0.394, p < 0.01), and BI (r = 0.427, p < 0.01). ATU showed a positive correlation with BI (r = 0.388, p < 0.01). Model evaluation Table 3 showed the results of the measurement model evaluation, including factor loadings, average variance extracted (AVE), composite reliability (CR), and the square root of AVE. The factor loadings for all items in this study exceeded the recommended threshold of 0.60 [74], demonstrating strong relationships between the items and their respective constructs. To further verify the model’s convergent validity, AVE and CR values were calculated. The results presented that all constructs had AVE values exceeding the threshold of 0.50, and all constructs had CR values exceeding the recommended threshold of 0.70, demonstrating high internal consistency and convergent validity of the model. Then, to assess discriminant validity, this study adopted the method proposed by [75], comparing the square root of the AVE values with the correlation coefficients of the respective constructs. The results showed that the square root of the AVE values for each construct was greater than the correlation coefficients, thereby establishing discriminant validity. Table 3Measurement model evaluationConstructItemFactor LoadingsAVECRSquare Root of AVETSE30.724–0.7340.5290.7710.727TA30.763–0.7890.6010.8190.775PE40.790–0.8310.6670.8890.816PU40.712–0.7890.5830.8480.764PEU40.706–0.7800.5530.8320.744ATU30.796–0.8080.6440.8440.802BI40.753–0.7650.5780.8460.760 Measurement model evaluation Structural equation modeling This study utilized SEM to examine the relationships among the constructs, with model fit evaluated using standard indices. According to [76], acceptable thresholds are χ²/df between 1 and 3, CFI, and TLI, GFI, and NFI values > 0.90, and RMSEA < 0.08. The results presented that all indices met the evaluation criteria: χ²/df was 2.143, within the recommended range; RMSEA was 0.041, less than 0.08; SRMR was 0.093, which was within the acceptable range [77], particularly for larger sample sizes; and the values of CFI (0.961), GFI (0.941), NFI (0.930), and TLI (0.956) all exceeded 0.90, meeting the criteria for excellent fit. Based on Fig. 3; Table 4, the standardized path coefficients and their significance were as follows. TSE was found to be negatively associated with TA (β = -0.461, p < 0.001) and positively associated with PE (β = 0.495, p < 0.001) and PEU (β = 0.449, p < 0.001). TA was negatively related to PEU (β = -0.248, p < 0.001). Both PE and PEU were positively related to PU (β = 0.207, 0.242, p < 0.001), and PU and PEU were positively linked to ATU (β = 0.207, 0.446, p < 0.001), which in turn was positively associated with BI (β = 0.484, p < 0.001). However, the paths from TSE to PU (β = 0.090, p = 0.158) and from PU to BI (β = 0.008, p = 0.864) were not significant, indicating that other factors, such as attitudes and emotions, may play a more important role. Table 4Path analysisPathβS.E.C.R. P HypothesisR²TSE→PU (H1)0.0900.0651.4100.158Not Supported0.167TSE→PEU (H2)0.4490.0598.088***SupportedTSE→PE (H3)0.4950.0569.967***Supported0.251TSE→TA (H4)-0.4610.059-8.967***Supported0.233PE→PU (H5)0.2070.0464.038***SupportedTA→PEU (H6)-0.2480.046-4.998***Supported0.391PEU→PU (H7)0.2420.0564.145***SupportedPEU→ATU (H8)0.4460.0519.017***SupportedPU→ATU (H9)0.2070.0494.552***Supported0.344PU→BI (H10)0.0080.0470.1710.864Not SupportedATU→BI (H11)0.4840.0499.499***Supported0.128***p<0.001 Path analysis ***p<0.001 Mediation analysis Table 5 presented the results of the mediation effect tests. A mediation effect was considered significant if the 95% confidence interval did not include zero [78]. Notably, PU did not serve as significant mediators, as their 95% confidence intervals all crossed zero (TSE→PU→BI [-0.004, 0.029]; PE→PU→BI [-0.015, 0.026]). Serial mediation analysis revealed that PU and ATU (95% CI [0.009, 0.024]), as well as PEU and ATU (95% CI [0.012, 0.055]), acted as mediators in the relationship between TSE and BI. Then, PU and ATU (95% CI [0.011, 0.030]), and PEU and ATU (95% CI [-0.033, -0.007]) also serially mediated the relationships between PE and BI, TA and BI respectively, while PEU combined with PU did not act as serial mediators in the relationship between TA and BI. Table 5Bootstrap mediation effectPathEstimatesS.EBiasPercentileLowerUpper p LowerUpper p TSE→PU→BI (H12)0.0120.008-0.0040.0290.346-0.0040.0290.346PE→PU→BI (H13)0.0050.010-0.0150.0260.250-0.0150.0250.250TSE→PU→ATU→BI(H14)0.0150.0040.0090.0240.0000.0090.0240.000TSE→PEU→ATU→BI (H15)0.0280.0110.0120.0550.0000.0110.0530.000PE→PU→ATU→BI (H16)0.0200.0050.0110.0300.0000.0120.0300.000TA→PEU→ATU→BI (H17)-0.0160.006-0.033-0.0070.000-0.031-0.0060.000TA→PEU→PU→BI (H18)0.0020.003-0.0030.0080.193-0.0030.0080.192 Bootstrap mediation effect Discussion This study extended TAM by incorporating cognitive (i.e., TSE) and emotional factors (i.e., TA and PE) as external variables, and examined the relationships among these external variables, the core TAM constructs (PU, PEU, and ATU), and the outcome variable (BI) in the context of EFL students using LLMs for English learning. This extension of TAM represents a critical theoretical development, broadening its scope to include both cognitive and emotional factors, which have rarely been jointly explored in TAM, especially in the context of LLM-assisted language learning. The results indicated that TSE was positively associated with PE and negatively associated with TA. The serial mediation analysis further revealed that PU, PEU, and ATU sequentially mediated the effects of TSE, TA, and PE on BI. Specifically, TSE was associated with higher PU and PEU, which in turn were associated with higher ATU and BI. Similarly, PE was positively associated with PU, which was further associated with higher ATU and BI. In contrast, TA was negatively associated with PEU, which was associated with lower ATU and BI. The main findings are discussed in detail in the following section. The relationships between technology self-efficacy, emotions, perceptions, attitudes and behavioral intentions This study demonstrated that TSE was positively associated with PE and negatively related to TA in the context of LLM-assisted English learning, with H3 and H4 supported, which is partially consistent with previous research [43, 65]. Students with higher TSE are more likely to enjoy using LLMs for English learning because they feel confident in navigating them. On the other hand, this confidence helps reduce anxiety, creating a more positive and less stressful learning experience. In short, higher TSE strengthens individuals’ ability to handle technological challenges, reducing negative emotions and accumulating user experiences [43]. Within TAM, it revealed that cognitive (e.g., TSE) and emotional factors (e.g., TA and PE) as external variables are also related to each other. Regarding the influence of TSE and emotions on perceptions, this study confirmed that TSE significantly enhances students’ PEU of LLMs, H2 supported, and partially aligning with previous research [44, 45]. EFL students with higher TSE exhibit greater confidence in using LLMs, leading them to perceive these tools as more user-friendly and effective in improving language skills. Nonetheless, this study did not find that TSE could significantly influence PU, meaning H1 was not supported. While TSE was shown to facilitate PEU, it did not necessarily translate into stronger perceptions of the utilitarian value of LLMs (i.e., PU). This suggests that students’ confidence in handling LLMs for English learning (i.e., TSE) may primarily reduce operational barriers rather than shaping their beliefs about its academic utility. One possible explanation is that EFL students’ PEU are more strongly tied to learners’ emotional experiences [79], which can not be inferred merely from a sense of technical competence. Another possible explanation lies in the distinctive technological characteristics of LLMs. Unlike conventional educational technologies that offer task-specific functionalities, LLMs function as open-ended, generative systems whose outputs are highly dependent on user input and contextual interpretation. As a result, even learners with high TSE may not consistently perceive LLMs as useful for academic purposes, because the quality, relevance, and reliability of generated content can vary substantially. In other words, while students may feel technically confident in using LLMs, their belief in the academic utility of these tools (PU) may require more than just familiarity with the technology. Effective guidance, training, or instructional integration could be necessary to help students see how LLMs can be used meaningfully for language learning tasks. Additionally, this divergence from previous findings [47, 49] may also stem from the native language background (the aforementioned two studies: Saudi; Peninsular Malaysia), sample size (n = 75; n = 303), and the surveyed learning tools (Zoom; not specified in learning tools). Specifically, Chinese EFL learners in this study may tend to evaluate LLMs based on task performance in an exam-oriented context, which may weaken the link between TS and PU. Thus, this finding refines TAM by showing that the pathway from self-efficacy to usefulness is not universal but context-dependent. Students’ PE positively influences their PU of LLMs for English learning, supporting H5, and partially aligning with previous studies [16, 17, 29]. EFL students who find using LLMs enjoyable are more likely to perceive these tools as valuable for improving their English proficiency, fostering positive perceptions of the technology. In contrast, students’ TA significantly inhibits their PEU, supporting H6 and confirming several studies on its negative effect [58, 61]. In particular, EFL students with high levels of TA tend to negatively evaluate LLMs for English learning, perceiving it as complex and challenging to use. Conversely, EFL students with high levels of TSE and PE are more likely to find the technology easy to use and beneficial for learning. Together, these findings highlight that TAM can be expanded to incorporate psychological dimensions as external antecedents, enriching its explanatory power in technology-enhanced language learning contexts. With respect to the relationships between perceptions, attitudes, and behavioral intentions, the current study found that PEU has a significant positive effect on PU and ATU, supporting H7 and H8 and validating the similar results of recent studies [33, 47]. When EFL students perceive LLMs as easy to use, they are more likely to see them as useful and develop positive attitudes toward adoption. A recent mobile learning study also confirmed these relationships [33]. Moreover, this study found that PU has a significant positive effect on ATU, H9 supported, and in line with previous research [31], while PU does not have a significant direct effect on BI, as indicated by H10, which was not supported. That is, when EFL students perceive LLMs as applicable for English learning, they directly foster a more favorable attitude toward using LLMs, but their PU does not independently improve their BI. A possible explanation is that students may acknowledge the practical benefits of LLMs for English learning, yet such recognition does not independently motivate usage unless it is accompanied by favorable variables (e.g., attitudes) toward the technology. This finding represents an important contribution to extending TAM, as it suggests that PU may influence BI indirectly through the formation of positive attitudes, rather than acting as a direct predictor. In contrast, one study found that PU does not significantly predict ATU [19]. A possible explanation is that this study specifically examined LLMs for English learning, making students more aware of their usefulness, and thereby strengthening their ATU. However, the other study primarily investigated the factors influencing college learners’ intention to use AI in distributed EFL learning contexts [19], inevitably making the use orientation of AI obscure. This study also identified the direct positive effect of ATU on BI, supporting H11 and consistent with the findings of previous studies [18, 31]. EFL students with positive attitudes toward LLM-assisted English learning show stronger intentions to adopt. Likewise, research has shown that learners who hold favorable attitudes toward ChatGPT are more likely to use it outside the classroom [4]. This is probably because students with high levels of ATU perceive greater usefulness and enjoyment in LLM-assisted learning, which motivates them to translate positive attitudes into actual behavioral intentions. By reaffirming the central role of ATU in shaping BI, the finding also reinforces the core proposition of TAM. The role of perceptions and attitudes as serial mediators between technology self-efficacy, emotions, and behavioral intentions When examining the mediating roles of perceptions and attitudes between TSE, emotions, and BI, this study unveiled that students’ TSE and emotions toward using LLMs for English learning would influence their final BI through a serial mediation process of perceptions and attitudes. This is partially consistent with previous studies [25, 26]. The identification of serial mediation is a key theoretical contribution in this study, as it suggests that EFL students’ TSE and emotions do not directly shape BI. Instead, they affect perceptions first, which then influence attitudes, and ultimately determine BI. Specifically, when EFL students consider using LLMs in English learning, their TSE, and emotions first influence their perceptions of LLMs (i.e., PU and PEU). This cognitive evaluation subsequently shapes their attitudes, whether they use LLMs as necessary, or worth investing time and effort in. Ultimately, their attitudes determine their BI to integrate LLMs into their English learning practices. This pathway highlights the psychological variables-perception–attitude–intention chain appears within TAM. Intriguingly, students’ ATU emerged as a vital link between TSE, emotions, and BI, partially consistent with previous findings [25, 26]. Therefore, H12 and H13 were not supported. This likely reflects that students with more positive attitudes toward LLM-assisted English learning are more motivated to act on these attitudes, resulting in stronger intentions to use the technology. Furthermore, EFL students’ PU and PEU did not directly mediate the link between TSE, emotions, and BI, likely because their influence is also channeled through attitudes toward using LLMs for English learning. This finding contrasts with [80], who argued that ATU did not play a central role between external factors, perceptions, and BI. The divergence may result from differences in participants and contextual factors, including varying levels of technological familiarity and the specific learning context, as the present study focused on LLM-assisted English learning rather than general online courses. Moreover, it challenges the traditional TAM assumption that PU directly predicts BI. In the context of LLM-assisted English learning, students’ adoption decisions appear to be primarily driven by attitudes. TSE positively predicts students’ PEU and PU, which in turn are associated with ATU and BI, echoing previous research [25], with H14 and H15 supported. When EFL students feel confident using LLMs, they tend to believe these tools can effectively support their English learning (i.e., higher PU) and are easy to use (i.e., higher PEU). These perceptions foster more positive attitudes toward LLMs (i.e., higher ATU). As attitudes improve, students become more willing to integrate LLMs into their English learning, thereby strengthening their BI to adopt the technology. This reinforces TAM’s original proposition regarding the importance of PEU and PU in shaping ATU and BI, but it also extends the model by evidencing that TSE can initiate this causal chain. Then, PE only influences students’ BI by firstly affecting PU, which impacts ATU, and ultimately shapes their BI to adopt LLMs for English learning, with H16 supported. This partially supports the results of previous research [28]. EFL students who enjoy using LLMs are more likely to perceive them as beneficial for English learning (i.e., higher PU). This positive perception increases favorable attitudes toward using LLMs (i.e., higher ATU), which then translate into stronger BI. In contrast, TA was negatively associated with students’ PEU of LLMs for English learning, which was further related to ATU and subsequently to BI, as similarly reported by previous research [26], with H17 supported. When EFL students experience TA, they feel overwhelmed and uncertain about using LLMs, perceiving them as difficult to control in English learning, which lowers their PEU, making the tools seem more complex and frustrating (i.e., lower PEU). As a result, students’ attitudes toward using LLMs become negative (i.e., lower ATU), reducing their willingness to adopt LLMs for English learning (i.e., lower BI). It is noteworthy that the path (TA→PEU→PU→BI) was not significant, indicating that H18 was not supported, which further validates the mediating role of ATU in the relationship between emotions (e.g., TA) and BI. This suggests that students’ ATU serve as a crucial gateway through which TA ultimately impacts students’ adoption decisions, rather than through a purely perceptive process (i.e., PEU and PU). Therefore, this study identified the central role of TAM (i.e., ATU) while extending its theoretical scope to include psychological dimensions, particularly emotions, that have been rarely examined alongside cognitive factors in LLM-assisted language learning. Conclusion This study, within the TAM framework, highlights the pivotal role of attitudes (ATU) by identifying the serial mediation of perceptions (PU and PEU) and ATU between technology self-efficacy (TSE), emotions, and EFL students’ behavioral intentions (BI) to use LLMs for English learning. The results indicate that technology acceptance in LLM-assisted language learning is not driven by isolated cognitive or emotional factors, but by a sequential process in which perceptions and attitudes jointly mediate these influences. This extends TAM by demonstrating that its explanatory mechanism can be more effectively understood as a layered and interconnected pathway in AI-enhanced learning contexts. Moreover, this research provides practical implications for optimizing the design of LLMs and improving the adoption of educational technology. Implications This study makes a significant theoretical contribution to the extension of TAM by integrating both cognitive and emotional factors, which have not been sufficiently addressed in traditional TAM. By incorporating TSE and emotional factors such as PE and TA, the study provides a more comprehensive understanding of the complex psychological mechanisms involved in technology acceptance, especially in the context of LLM-assisted language learning. Then, the inclusion of serial mediation allows for a more nuanced exploration of how cognitive (e.g., TSE) and emotional factors (e.g., PE and TA) as external variables shape students’ perceptions (e.g., PU and PEU), which in turn influence their attitudes toward the technology and ultimately their BI. This approach advances TAM by offering a more detailed theoretical framework that acknowledges the interrelated roles of cognitive and emotional factors in shaping technology acceptance within AI-enhanced learning environments. Furthermore, this study demonstrates that TAM can be adapted and extended to better fit the language learning context, showing that cognitive and emotional factors are critical variables in understanding students’ acceptance of LLMs. By doing so, this study highlights the need to integrate psychological aspects such as TSE, PE and TA more systematically within TAM, thus extending its applicability beyond its traditional focus on PU and PEU. This represents a theoretical advancement of TAM in the field of LLM-assisted language learning, where emotional and cognitive aspects play a vital role in the adoption of technology. This study also offers practical implications for enhancing EFL university students’ English learning outcomes and informing technology-integrated pedagogical practices. First, the study indicates that TSE is positively associated with students’ PEU and PE, while being negatively associated with TA when using LLMs for English learning. These findings suggest that students’ confidence in using LLMs plays a key role in shaping both their cognitive and emotional experiences of technology use. Based on this, targeted interventions such as hands-on workshops, step-by-step tutorials, and scaffolded practice sessions should be implemented to build students’ confidence in using LLMs for tasks like essay drafting, grammar checking, and vocabulary learning. These interventions can include guided exercises where students gradually tackle increasingly complex tasks with LLM support, accompanied by real-time feedback. Second, the study underscores the critical role of TSE and emotions (i.e., PE and TA) in shaping EFL students’ BI toward using LLMs. This finding highlights the need for LLM designs and learning activities that are both intuitive and pedagogically meaningful. Developers and educators could simplify user interfaces, provide in-app guidance or prompts for effective LLM use, and embed features that clearly show learning gains (e.g., suggestions for improving sentence coherence or vocabulary variety) to boost student engagement. Third, the study emphasizes the pivotal role of ATU in influencing students’ BI to use LLMs for English learning. Future initiatives should actively cultivate positive attitudes toward LLMs as legitimate learning partners. Practical strategies include integrating LLM-assisted tasks into regular coursework, such as adaptive writing exercises, role-play or AI-mediated conversational practice, demonstrating their educational value. Teachers can model effective use cases in class, explicitly discuss the pedagogical advantages of LLMs, and highlight ethical considerations, such as verifying AI-generated content and avoiding overreliance. In the end, this study offers guidance for higher education institutions in implementing technology-assisted English teaching. Universities should establish comprehensive support systems that combine technical training workshops, emotional and motivational support, and ongoing professional development for teachers. Institutional policies should promote responsible and ethical use of LLMs, ensuring that students not only become proficient users but also critical evaluators of AI-generated outputs, able to judge the accuracy, appropriateness, and reliability of the content. Limitations and future research The current research still has several limitations. First, the study’s sample is limited to non-English major university students in China, which may not fully represent the broader population of EFL learners. The learning needs, perceptions, and cultural contexts of this group might differ significantly from those of English majors or students from different cultural or educational backgrounds. To address this limitation, future research should broaden the sample to include students from various majors (e.g., English majors) as well as students from diverse cultural and geographical contexts to enhance the generalizability and external validity of the findings. Second, this study adopts a cross-sectional design and relies on self-reported data, which limits causal inference and may introduce common method bias (CMB). In addition, behavioral intention was the only outcome measure, and actual LLM usage data were not collected. Although behavioral intention is a strong predictor of technology use, the intention–behavior gap could not be examined in this study. Future research is encouraged to adopt longitudinal or experimental designs and incorporate objective usage data (e.g., system logs or interaction records) to better capture dynamic changes and strengthen ecological validity. Third, the present model primarily focuses on serial mediation pathways and does not include all possible direct effects among constructs. Specifically, based on the TAM framework and a synthesis of relevant prior research, this study conceptualizes a fully mediated model to examine the sequential mechanisms linking cognitive and emotional factors to behavioral intention. However, the possibility of direct effects among certain variables can not be ruled out. Future studies may consider incorporating these direct paths to further validate and refine the proposed model structure. Finally, while this study incorporates key emotional variables such as PE and TA, there is room to explore additional emotional constructs that could further illuminate the emotional landscape of technology adoption in language learning. Drawing on theories of academic emotions [81], future research could examine emotions such as hope, pride, and shame, and how these emotions impact students’ engagement with LLMs. Supplementary Information Supplementary Material 1. Acknowledgements The authors wish to thank all the participants who took part in this study. Special thanks are extended to five anonymous reviewers for their insightful comments. Abbreviations Artificial Intelligence Artificial Intelligence in Education Large Language Models English as a Foreign Language Technology Acceptance Model Technology Self-efficacy Technology Anxiety Perceived Enjoyment Perceived Usefulness Perceived Ease of Use Attitude toward Use Behavioral Intention Authors’ contributions Z.S. (Zhaoyang Shan): methodology, data curation, formal analysis, visualization, and writing—original and final drafts; M.S. (Manman Shan): research design refinement and statistical modeling; L.D. (Lanyun Ding): literature review and language polishing; Y.S. (Yonggang Su): conceptualization, theoretical framework, supervision, writing—review and editing, and funding acquisition. Funding This research was supported by the Social Science Key Project of Shandong Province (22BYYJ01). Data availability The datasets generated and analysed during the current study will be made available upon reasonable request. Declarations Ethics approval and consent to participate This study was conducted in accordance with the Declaration of Helsinki and approved by the Institutional Review Board (IRB) of the School of Foreign Languages and Literature at Shandong University. Informed consent was obtained from all participants involved in the study. Consent for publication Not applicable. 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 Associated Data Supplementary Materials Supplementary Material 1. Data Availability Statement The datasets generated and analysed during the current study will be made available upon reasonable request.