Introduction

The significant advancements in artificial intelligence (AI) have presented opportunities for development and reform across various fields, and education is no exception. Notably, the emergence of Chat Generative Pre-Trained Transformer (ChatGPT) highlights the immense potential and broad prospects of AI in education (Chen et al. 2024; Wang et al. 2024). ChatGPT can generate human-like text, providing personalized learning content while offering an interactive learning experience based on users’ input (Gao et al. 2023; Kung et al. 2023; Ryong et al. 2023; Berdejo-Espinola and Amano 2023). Research indicates that ChatGPT can act as a virtual learning assistant, particularly in language learning, by enabling learners to interact in their target language, ask questions, request explanations, and seek linguistic assistance (Qu and Wu 2024). Its applications include speaking practice, grammar correction, and text translation, which enhance learners’ skills in listening, reading, speaking, and writing (Halaweh 2023; Extance 2023; Annamalai et al. 2023).

Despite the demonstrated potential of ChatGPT, learning Chinese as a Second Language (CSL) presents unique challenges due to the language’s complex tonal system, pictographic characters, and culturally embedded expressions (Cai and Lee 2015). Prior research highlights several advantages of using ChatGPT in international Chinese language education: (1) it can interpret learners’ inputs in standard Mandarin, assisting in pronunciation standardization; (2) it effectively recognizes and distinguishes confusing vocabulary, helping learners accurately understand word usage in different contexts; (3) it offers robust error analysis, enabling learners to correct linguistic errors; and (4) it adjusts text complexity to match learners’ comprehension levels, providing customized reading materials (Wei et al., 2025). These features emphasize the potential of ChatGPT in addressing the specific needs of CSL learners. However, most existing research has focused on general university student populations (Wang et al. 2024), as they are typically early adopters of emerging technologies like ChatGPT (Shahzad et al. 2024; Polyportis and Pahos 2024; Budhathoki et al. 2024). Although studies on language learners’ acceptance of ChatGPT are gradually increasing, such as those by Qu and Wu (2024) and Liu and Ma (2023), which examine its adoption among English as a Foreign Language (EFL) learners, the majority of these studies focus on EFL contexts and provide limited insights into the acceptance and usage of ChatGPT among CSL learners (Li et al. 2024). Therefore, research focused on CSL learners is particularly urgent and necessary, as it helps better understand the needs and dynamics of this specific learning group.

Furthermore, the adoption of ChatGPT raises concerns about the generation of inaccurate information, potential plagiarism, and risks to user data privacy. During personalized learning, ChatGPT collects large amounts of personal data, learning records, and user behavior. Misuse, leaks, or inadequate protection of this data can lead to significant trust issues, ultimately diminishing learners’ confidence in AI tools and affecting their usage behaviors. Although researchers have extended the Technology Acceptance Model (TAM) to assess users’ attitudes toward ChatGPT, studies have primarily focused on factors such as AI literacy (Wang et al. 2024), anxiety (Budhathoki et al. 2024), and personal innovativeness (Strzelecki et al. 2024). However, few studies have incorporated Trust (TRU) as a variable within their models. Shahzad et al. (2024) extended the TAM by adding perceived intelligence and perceived trust, confirming the significant role of TRU among Chinese university students in their awareness, acceptance, and adoption of ChatGPT. Similarly, Polyportis and Pahos (2024) expanded the meta-UTAUT framework by incorporating TRU as an essential factor, identifying it as a significant positive antecedent of users’ attitudes. Both studies have established the importance of TRU among university students regarding the acceptance of ChatGPT, it is equally critical for CSL learners. Therefore, introducing the TRU variable to expand the model and investigate CSL learners’ acceptance and adoption of ChatGPT is of great significance.

Recent studies have further indicated that the explanatory power of traditional TAM constructs may vary across different stages of technology adoption. In particular, when technologies become widely adopted and integrated into users’ everyday practices, the predictive strength of Perceived Ease of Use (PEOU) and Perceived Usefulness (PU) on Intention to use (ITU) appears to weaken. For example, Wang et al. (2022) and Lin and Yu (2023) found that learners’ PEOU no longer significantly determined their behavioral intentions in mature usage contexts. Similarly, Deng and Yu (2023) as well as Fan and Wang (2023) reported that PU may fail to exert a strong effect on ITU when users are already familiar with the technology and have incorporated it into routine learning practices. Furthermore, a study examining Thai higher education students’ willingness to use ChatGPT revealed that neither PU nor PEOU significantly influenced learners’ ITU toward ChatGPT (Shaengchart et al. 2023). These findings suggest that TAM has boundary conditions in ubiquitous adoption settings, where other constructs—such as habit, social influence, or hedonic motivation—may become more salient. Building on this stream of literature, our study not only extends the TRAM framework by integrating the dimension of TRU but also provides new insights into the limitations of TAM in post-adoption contexts, especially in the case of CSL learners’ use of ChatGPT.

In light of these gaps, this study introduces TRU as a key variable to expand the TAM-based Technology Readiness and Acceptance Model (TRAM). By employing Structural Equation Modeling (SEM), we aim to assess how TRU and Technology Readiness (TR) factors influence CSL learners’ acceptance and usage of ChatGPT in higher education. This study addresses three critical questions: (1) How do CSL learners perceive and accept ChatGPT as a language learning tool? (2) What roles do TRU and TR play in shaping Attitude toward use (ATU) and ITU? (3) Under mature-use conditions, do PEOU and PU still predict ITU—either directly or indirectly via ATU? These findings are expected to offer insights and implications for integrating ChatGPT into international Chinese language education. The remainder of this paper is structured as follows: Section 2 presents a literature review and formulates research hypotheses. Section 3 outlines the methodology. Section 4 reports the results of the data analysis. Section 5 discusses the findings and their implications. Section 6 offers pedagogical insights and recommendations. Finally, the main conclusions, limitations, and directions for future research are provided.

Theoretical framework and hypotheses development

The integration of ChatGPT into Chinese learning environments raises important questions about its acceptance and use by learners. Research on this topic is crucial, as CSL learners’ willingness to engage with ChatGPT can significantly impact their language acquisition outcomes (Li et al. 2024). Understanding how CSL learners perceive and utilize this technology will provide insights into its role in enhancing language learning.

Theoretical framework

The emergence of new technologies has prompted significant transformations in the field of international Chinese language education. Various established theories and analytical frameworks have been developed to examine individual acceptance and usage tendencies. For example, Barrett et al. (2020) Adapted the TAM to investigate CSL learners’ attitudes towards Hubs by Mozilla, a multi-user VR learning environment. Similarly, Yang and Lou (2024) combined the Self-Determination Theory (SDT) and TAM to explore CSL learners’ attitudes towards mobile learning technologies. Drawing on this approach, the current study incorporates SDT as a complementary theoretical lens to better understand learners’ intrinsic motivation in using ChatGPT. Specifically, autonomy (the sense of control over learning), competence (the perceived ability to achieve learning goals), and relatedness (the feeling of connectedness with others or the learning environment) are expected to influence learners’ PU and PEOU, thereby shaping their acceptance of AI-based tools. More recently, Li et al. (2024) apply the TAM to analyze CSL learners’ acceptance of ChatGPT in oral language practices and its influencing factors. Other frameworks, such as the Unified Theory of Acceptance and Use of Technology (Ahmed et al. 2023) and the Extended TAM Model (Liang and Li 2023) have been effectively used to analyze factors affecting the adoption of digital learning resources in the education of CSL learners.

ChatGPT’s influence in the field of education is increasingly significant, as it introduces new possibilities for learning methods. The acceptance and use of CSL learners towards ChatGPT are closely linked to its effective implementation in actual learning environments. When examining the acceptance and use of new technology by learners, it is essential to consider not only the technology itself but also the individual psychological and emotional factors involved. Therefore, attention must be directed towards the system-specific perceptions of CSL learners regarding ChatGPT. Additionally, the overall attitudes and general inclinations of CSL learners towards ChatGPT should also be assessed (Kampa 2023).

Lin et al. (2007) incorporated TR into the widely used TAM framework, thus giving rise to the TRAM (see Fig. 1). Within the TRAM framework, TR significantly shapes users’ perceptions regarding the ease of use and usefulness of new technologies, which in turn influences their acceptance and adoption of these technologies (Kampa 2023; Nigatu et al. 2024). Parasuraman (2000) defines TR as the tendency of individuals to accept and use new technologies to achieve personal and work-related goals, identifying four key influencing factors: Optimism (OPT), Innovativeness (INN), Discomfort (DIS), and Insecurity (INS). Individuals exhibiting OPT and INN are generally more willing to adopt new technologies and maintain a positive attitude, whereas those experiencing DIS and INS are more likely to reject the use of new technologies. On the other hand, TAM is a widely employed model for understanding user acceptance and usage of technology (Davis 1986; Davis 1989; Venkatesh 2000), derived from rational action theory (Fishbein and Ajzen 1977) and the theory of planned behavior (Ajzen 1987). The model delineates two primary determinants of user acceptance: PU, which refers to the belief that a technology can enhance productivity, and PEOU, which reflects perceptions of how easy it is to use the technology (Davis 1989). These factors significantly influence users’ attitudes towards technology, ultimately impacting their actual usage. Thus, both TR and TAM can be effectively utilized to explore users’ acceptance of new technologies (Kuo et al. 2013; Huang et al. 2015; Jin 2019; Sivathanu 2019; Kampa 2023).

Fig. 1
Fig. 1
Full size image

Technology readiness and technology acceptance model (TRAM).

The TRAM emerges as a comprehensive framework for explaining human behavioral intentions, offering a perspective that encompasses both system-specific cognition and general inclinations in human decision-making. This theory has been extensively utilized in domains such as sports (Cui 2022) and health care. For example, Nigatu et al. (2024) revealed that the TRAM provides a robust framework for understanding health professionals’ intentions to adopt teleradiology. In the education domain, the TRAM has been widely utilized to effectively predict higher education students’ acceptance and use of new technologies (Yusuf et al. 2021; Kampa 2023; Yusuf et al. 2024). Yusuf et al. (2021) and Kampa (2023) apply the TRAM to analyze university students’ readiness to adopt E-Learning and M-Learning, respectively. Yusuf et al. (2024) demonstrated that TR significantly influences students’ readiness and acceptance of social media technology in blended learning environments within private higher education institutions. This study aims to further investigate the application and expansion of the TRAM within the educational domain.

Although the TRAM has been widely employed in the field of education, its application in explaining and forecasting the acceptance and use of ChatGPT among CSL learners during their language acquisition process has not been thoroughly investigated. This creates an important research gap, as grasping the factors that shape the behavioral intentions of CSL learners regarding the use of ChatGPT is crucial for its successful incorporation into language learning activities and educational settings. Consequently, this study aims to explore the factors influencing the behavioral intentions of CSL learners to utilize ChatGPT by applying the TRAM model, thereby addressing this notable gap.

At the same time, some recent studies remind us that TAM may not work equally well in every situation. When a tool becomes very common or part of daily practice, the traditional predictors in TAM—PEOU and PU—seem to lose much of their power to explain intention to use (Shaengchart et al. 2023; Alshammari and Babu 2025; Lee et al. 2025). In other words, once learners already know that a system is useful and easy enough to handle, these factors no longer play a major role in shaping their decision to keep using it. This has led researchers to bring in other perspectives, such as UTAUT2, which adds variables like habit and hedonic motivation (Venkatesh et al. 2012). By referring to these insights, the present study not only expands TRAM with the inclusion of TRU but also speaks to how TAM may need to be reconsidered in settings where technologies like ChatGPT are already widely accepted.

The importance of this study stems from its ability to enhance the breadth of TRAM research and shed light on the acceptance and utilization of ChatGPT within the realm of international Chinese language education. A quantitative research methodology is utilized to uncover the primary factors that affect CSL learners’ engagement with ChatGPT, emphasizing technological perceptions and individual attitudes, which contributes to a comprehensive understanding of their attitude and acceptance. Furthermore, the study conducts a detailed analysis of the underlying technological and personal factors based on the results of SEM. Ultimately, this research offers relevant educational recommendations and policies for the use of ChatGPT among CSL learners in educational contexts.

hypothesis development

Earlier studies have highlighted the significance of TRAM as a fundamental theoretical framework for understanding and forecasting human behavioral intentions. However, in the context of this research, specific challenges are faced by CSL learners when utilizing ChatGPT. These challenges necessitate appropriate modifications and adjustments to the TRAM. Therefore, this study incorporates TRU factors into the TRAM framework to enhance the model’s explanatory and predictive capacities, as illustrated in Fig. 2.

Fig. 2
Fig. 2
Full size image

Hypothesized model of CSL learners’ ChatGPT acceptance and use.

Optimism (OPT)

OPT is a psychological framework that encourages individuals to view new developments, particularly in technology, in a positive light (Parasuraman 2000; Kuo et al. 2013). Optimistic individuals are more likely to focus on the conveniences and advantages offered by technology rather than the potential risks and negative consequences. Previous research has consistently validates the significance of OPT in rapidly changing technological environments, as it promotes proactive engagement with new technologies. For example, Yusuf et al. (2021) demonstrated the significant influence of OPT on PEOU and PU in the context of Online Learning or E-Learning based on Learning Management System. Meanwhile, Kampa (2023) also examine the effects of students’ OPT on their PEOU and PU regarding M-Learning, and reveal that users are generally able to recognize the potential benefits of technology, leading to a positive attitude and greater acceptance of new technologies. Hence, the following hypothesis is proposed:

H1a: OPT has a significant and positive impact on PEOU.

H1b: OPT has a significant and positive impact on PU.

Innovativeness (INN)

INN is one of the positive factors within TR, significantly influencing individuals’ PEOU and PU of new technologies (Chen and Lin 2018). Individuals with strong innovative capabilities are often driven by intrinsic motivation to actively explore and learn about new technologies (Parasuraman 2000). These early adopters typically maintain a positive attitude towards the learning curves and operational complexities associated with new technologies, thus facilitating their acceptance and usage (Kim and Chiu 2019; Nigatu et al. 2024). Numerous scholars have empirically validated the impact of INN on PEOU and PU. For example, Yusuf et al. (2024) and Sivathanu (2019) both confirmed the influence of INN on PEOU and PU. Based on these observations, the following hypothesis is proposed:

H2a: INN has a significant and positive impact on PEOU.

H2b: INN has a significant and positive impact on PU.

Discomfort (DIS)

DIS is a negative psychological state characterized by a feeling of lack of control and confusion when using new technologies (Parasuraman 2000). It is often accompanied by anxiety and frustration, which can negatively affect users’ understanding and utilization of technology (Walczuch et al. 2007; Nigatu et al. 2024). Consequently, individuals experiencing DIS tend to have doubts and confusion regarding the PEOU and PU of new technologies. Previous research has validated the negative impact of DIS on technology adoption. For example, Cui (2022) revealed that DIS significantly negatively influenced the PEOU and PU of mobile fitness apps. Therefore, the following hypothesis is proposed:

H3a: DIS has a significant and negative impact on PEOU.

H3b: DIS has a significant and negative impact on PU.

Insecurity (INS)

INS is one of the negative factors within TR. Individuals who experience INS typically lack trust in new technologies, which creates barriers to their acceptance and usage (Parasuraman 2000; Cui 2022). When users feel insecure, they perceive the use of new technologies as potentially hazardous, citing concerns such as data breaches, technical malfunctions, or other risks (Davis 1993). This mindset leads to feelings of tension and unease regarding the operation of new technologies, resulting in the perception that such technologies are complex or difficult to master. Previous research indicates a significant negative correlation between INS and PEOU and PU. For example, Nigatu et al. (2024) demonstrated the significant influence of INS on PEOU and PU in the context of health care. Therefore, the following hypothesis is proposed:

H4a: INS has a significant and negative impact on PEOU.

H4b: INS has a significant and negative impact on PU.

Trust (TRU)

TRU is introduced in this study as a novel construct to extend the TRAM framework. It generally represents users’ confidence in key aspects such as technology, data security, and privacy protection. TRU plays an especially pivotal role when AI technologies are applied within educational contexts (Qin et al. 2020). It has long been recognized as a key driver of technology adoption and has a direct positive influence on users’ willingness to accept and utilize new tools (Wang et al. 2024). When users perceive AI technologies as trustworthy, they are more likely to affirm their effectiveness, safety, and reliability. As revealed by Wu et al. (2011), TRU significantly moderates the relationships among user-related and technology-related factors within the TAM structure. In this context, TRU serves as a psychological enabler that strengthens the connection between users and emerging technologies, thereby increasing their motivation to explore and learn how to use these tools (Raffaghelli et al. 2022; Ali et al. 2023). As a result, anxiety and discomfort during technology use can be reduced, thereby enhancing both PEOU and PU. A substantial body of theoretical and empirical research supports the integration of TRU into the TAM framework (Chircu et al. 2000; Gefen 2000). Moreover, Nazaretsky et al. (2022) demonstrated that a well-designed professional development program can effectively enhance teachers’ trust in educational technology, which in turn improves their readiness to adopt AI-based tools. Shahzad et al. (2024) further confirmed the significant impact of TRU on Chinese university students’ awareness, acceptance, and actual usage of ChatGPT. Based on these findings, the study formulates two hypotheses:

H5a: TRU has a significant and positive impact on PEOU.

H5b: TRU has a significant and positive impact on PU.

The effect of perceived ease of use on perceived usefulness

PEOU significantly influences PU (Davis 1989; Davis et al. 1989; Venkatesh 2000). In other words, users tend to view user-friendly technologies as more useful. This perspective suggests that when the barriers to using new technologies are low and users can operate them easily, they can achieve their goals more quickly and obtain desired outcomes. Such experiences lead users to form positive perceptions of new technologies, subsequently acknowledging their usefulness (Wang et al. 2022; Nigatu et al. 2024). For example, in the education field, Lin and Yu (2023) revealed that PEOU is a significant positive predictor of students’ PU of digital academic reading tools on computers. Therefore, the following hypothesis is proposed:

H6: PEOU has a significant and positive impact on PU.

The effect of perceived ease of use and perceived usefulness on attitude to use

PEOU and PU are considered predictive factors for users’ attitude toward using (ATU) a technology (Chen and Lin 2018; Liang and Li 2023). When users perceive new technologies as useful, they tend to believe that using these technologies will result in higher productivity or better learning outcomes, which further enhances their favorable attitude toward the technology. Similarly, when users find new technologies easy to operate and use, they are more likely to develop a positive attitude toward usage, influencing their ultimate willingness to adopt the technology. Previous research indicates that PEOU and PU significantly affect users’ ATU. For example, Yu et al. (2023) examined the effect of Chinese university students’ PEOU and PU on their attitudes towards blended learning, and revealed that both variables can positively affect their ATU. Therefore, we propose the following hypothesis:

H7: PEOU has a significant and positive impact on ATU.

H8: PU has a significant and positive impact on ATU.

The effect of perceived ease of use, perceived usefulness and attitude toward use on intention to use

According to the TAM, PEOU and PU not only directly influence users’ intention to use technology but also indirectly affect final usage decisions through the mediating variable of ATU (Zhou et al. 2021; Yu et al. 2023). Previous research highlights the substantial empirical relationships among PU, PEOU, ATU, and Intention to use (ITU). For example, Lin and Yu (2024) suggested that the traditional hypotheses regarding the relationships among PEOU, PU, and ITU were supported in the context of higher education students using AI-generated virtual teachers in language learning videos. However, the predictive effect of PU on ITU is notably inconsistent, as a considerable amount of research indicates that PU does not significantly predict ITU (Wang et al. 2022; Deng and Yu 2023). In light of these inconsistencies, we intend to test the predictive effect of PU on ITU through H10 and investigate whether these traditional hypotheses are supported in the context of CSL learners. Therefore, we propose the following hypothesis:

H9: PEOU has a significant and positive impact on ITU.

H10: PU has a significant and positive impact on ITU.

H11: ATU has a significant and positive impact on ITU.

Methods

Research instruments

Separate sections comprised the survey. In the first section, participants were asked to provide consent for involvement and to share personal details, including their gender, age, educational level, years of learning Chinese and country. We assured the participants that their data would be strictly confidential and anonymous, and would be used exclusively for academic research purposes, and that they would formally proceed to the questionnaire only upon their consent. The second section comprised 33 items (see Supplementary Table S1), each rated on a 5-point Likert scale, with 1 representing strongly disagree and 5 representing strongly agree. To improve measurement validity and precision, the study enlisted the expertise of five professionals specializing in international Chinese language education, language testing, and linguistics to review, modify, and adapt the items, references for them were as follows: OPT, INN, DIS, INS (Parasuraman 2000; Parasuraman and Colby 2014; Chen and Lin 2018); TRU (Gefen 2000; Pavlou 2003); PEOU, PU, ATU, ITU (Davis 1989; Venkatesh 2000; Liang and Li 2023). For constructs that have been less researched, the design of our questions might have raised concerns. However, following the validation of previous studies on the acceptance of ChatGPT, the questionnaire items were further refined to better suit the current research context. Subsequently, a combination of exploratory and Confirmatory Factor Analysis (CFA) was employed to conduct a thorough examination of the results. The third section included an open-ended question designed to understand how learners utilize ChatGPT for Chinese language learning. Moreover, two scholars proficient in both Chinese and English performed back-translation of the items.

Research procedure

Before formally distributing the questionnaire, we had conducted a pilot study on 54 CSL learners. The analysis revealed that the questionnaire had a great internal consistency and reliability, as evidenced by its overall Cronbach’s Alpha values of 0.92 and each construct Cronbach’s Alpha values varying from 0.74 and 0.91. The finalized version of the questionnaire was made available on an online survey platform (www.wjx.cn), enabling us to share it via a hyperlink or a QR code for online distribution. Utilizing purposive sampling methods, we distributed the questionnaire on Chinese social media platforms such as WeChat to recruit interested CSL learners who had studied at Chinese universities and had used ChatGPT. The criteria for recruitment were as follows: (1) participants must have been CSL learners; (2) participants should have utilized ChatGPT for their Chinese language studies; (3) participants must have expressed a willingness to take part in this research and agreed to share their responses with academic researchers.

During this survey, 379 questionnaires were gathered, with 48 responses deemed incomplete or unclear. This left 331 valid responses, leading to an effective response rate of 87.34%. To ensure the quality of the sample, several rigorous measures were implemented. First, one screening question was posed to identify a representative sample among the 379 participants: “Have you ever used ChatGPT to assist in learning Chinese? (Yes/No)”. Consequently, participants who did not use ChatGPT for learning Chinese were deemed ineligible for this study, and their responses were classified as invalid. Second, we filtered the open-ended responses in the survey, removing answers that did not provide substantial information, such as “I don’t know”, or blank answers (Lin and Yu 2023). Third, responses displaying flat-lined patterns (consistent ratings across multiple questionnaire items) were eliminated due to indications of inattentiveness (Che et al. 2023). Eventually, the final dataset contained 331 valid responses.

Research methodology

The research model was validated using SEM to examine and estimate the relationships between variables (Hair et al. 2011). SEM is a multivariate statistical analysis technique that integrates path analysis and factor analysis, allowing for the simultaneous handling of multiple dependent relationships and enabling researchers to test complex theoretical models (Hair et al. 2010; Kline 2015; Kline 2023). To ensure the stability and reliability of the SEM analysis, the sample size adhered to established recommendations, with previous research suggesting a range of 200–500 participants for complex models (Yu et al. 2023). Additionally, data quality was critically assessed. Normality tests (Kline 2015) confirmed the distributions of the variables, and a Harman’s single-factor test was conducted to evaluate Common Method Bias (Podsakoff et al. 2003), ensuring no significant methodological biases affected the findings. The SEM analysis followed a two-step approach. First, a measurement model was established using CFA to validate the constructs’ reliability and validity. Second, the structural model was used to test the hypothesized relationships among variables. Model fit indices were employed to evaluate the adequacy of the model. The evaluations and validations were performed using AMOS 26.0 and SPSS 28.0 software. Path coefficients, standardized regression weights, and their statistical significance were interpreted to test the study’s hypotheses. The relevant results and detailed interpretations are reported in the next section.

Results

Participants

The demographic information of the valid participants is presented in Table 1. The survey results indicate that a substantial proportion of respondents are female, accounting for 75.8%, while male CSL learners comprise 24.2% of the sample. This imbalance in gender representation may be related to the higher prevalence of females in humanities disciplines. Regarding the age distribution of participants, those under 22 years old represent 48.9%, while respondents aged 22–30 account for 40.5%, and learners over 30 years old comprise only 10.6%. In terms of educational level, 62.8% of respondents hold a bachelor’s degree, 26.0% possess a master’s degree or higher, and 11.2% have completed high school or lower education. The majority of respondents have studied Chinese for more than three years (58.3%), with 28.7% studying for 1–3 years and 13.0% studying for less than one year. Furthermore, it should be noted that all respondents studied Chinese at three universities located in Beijing, Shanghai, and Guangzhou. Participants predominantly come from Asia (68%), Europe (18.1%), and other regions (13.9%), indicating that the sample represents a broader community of CSL learners.

Table 1 The demographic information (N = 331).
Full size table

Normality test and common method bias

Due to the use of maximum likelihood estimation method based on the assumption of multivariate normality in the SEM approach, the normality of the distributions for all nine variables in the survey was examined. The results indicate that skewness range is [−0.801, 0.441] and kurtosis range is [−1.089, 0.514]. According to Kline (2015), when the absolute value of skewness is less than 3 and the absolute value of kurtosis is less than 10, the distribution is considered acceptable for normality. Additionally, due to the data collection method or the use of a single source, there is a potential for common method bias. To address this concern, a Harman’s single-factor test was conducted using SPSS software to thoroughly evaluate the possibility of common method variance in this study. The results show that a single factor accounts for only 35.85% of the variance, whereas the conventional threshold for common method bias is 50% (Podsakoff et al. 2003). Thus, the data collected for this study do not exhibit any issues related to common method bias (MacKenzie and Podsakoff 2012). After confirming the necessary preconditions for SEM, we validated the reliability and validity of the research instrument, along with conducting assessments of model fit and path analysis. To enhance the reliability of the results, two items (DIS2 and INS1) were removed due to their factor loadings being below 0.5 (see Table 2). The subsequent tests used the revised survey data, excluding these items, to evaluate our proposed model.

Table 2 Confirmatory factor analysis results (N = 331).
Full size table

Confirmatory factor analysis

Reliability serves as a crucial metric for evaluating the internal consistency of measurement models. After analyzing the valid responses using SPSS 28.0, the questionnaire exhibited a high overall Cronbach’s Alpha of 0.92, with each construct above the 0.5 minimum threshold (Fornell and Larcker 1981) and ranging from 0.70 to 0.88, demonstrating excellent reliability. Validity refers to the precision of the measurement models in accurately capturing the intended content. To verify the validity of the measurement models, CFA was performed using AMOS 26.0. When assessing convergent validity, three aspects were considered: firstly, standardized factor loadings should exceed 0.5 for subsequent interpretive studies (Fornell and Larcker 1981); secondly, the Composite Reliability (CR) must exceed 0.6 (Bagozzi and Yi 1988; Hair et al. 2011); and lastly, it is recommended that the Average Variance Extracted (AVE) be more than 0.5 (Hair et al. 2010).

According to Table 2, every construct had a CR value between 0.69 and 0.88, which was higher than the advised value of 0.6. These metrics imply that there is an excellent internal consistency among the measurement items. However, the AVE values for the INN and DIS constructs are 0.36 and 0.46, respectively, which is below the 0.5 recommended cutoff. Fornell and Larcker (1981) state that AVE might offer a more cautious evaluation of the measurement model’s validity, and “on the basis of CR alone, the researcher may conclude that the convergent validity of the construct is adequate, even though more than 50% of the variance is due to error” (p. 46) (Lam 2012). This combination of AVE around 0.4 and CR exceeding 0.6 has been accepted, cited, and applied in recent research (Lam 2012; Chen and Lin 2018; Teo and Lee 2010; Lin and Yu 2023). In line with this perspective and consistent with Fornell and Larcker’s (1981) argument that a construct can still demonstrate adequate convergent validity when CR exceeds 0.6, even if AVE is below 0.5, we maintain that the convergent validity of the constructs in this study remains acceptable.

That said, we acknowledge that the relatively low AVE values for INN and DIS may suggest weaker convergent validity for these constructs compared to others. Additionally, while most standardized factor loadings exceeded the recommended threshold of 0.5, two items (DIS2 and INS1) fell below this value and were therefore removed from the final measurement model to enhance its overall fit and validity.

Concerning discriminant validity, a variable is considered to have excellent discriminant validity if the coefficient of Pearson correlation between two variables is lower than the square root of AVE (Fornell and Larcker 1981; Hair et al. 2010). Table 3 illustrates that the square root of AVE (the bold values) for each construct is greater than the correlation coefficients between the constructs, which indicated a satisfactory degree of discriminant validity for the measurement model.

Table 3 Discriminant validity.
Full size table

Model fit indices

Using the Maximum Likelihood method, the SEM constructed for this study was estimated, and the model fit was determined using AMOS 26.0. According to Table 4, a strong fit for the model was indicated by the Chi-square value (CMIN) /Degree of Freedom (DF) value of 2.153, which was less than the CMIN/DF criterion (Bagozzi and Yi 1988). The Adjusted goodness of fit index (AGFI) exceeded 0.8, with a value of 0.854. Surpassing the 0.9 threshold, the model and the data showed a strong match, as evidenced by the Comparative fit Index (CFI) of 0.915, the Tucker-Lewis Index (TLI) of 0.904, and the Incremental Fit Index (IFI) of 0.915 (Lai et al. 2023; Hair et al. 2011; Bagozzi and Yi 1988). Moreover, the Root Mean Square of Error Approximation (RMSEA) was 0.059, which reached the recommended threshold (Hu and Bentler 1999). Therefore, all estimates complied with the suggested criteria, demonstrating that the hypothetical model fit the data perfectly.

Table 4 Model fit measurement.
Full size table

Structural model

Standardized path coefficients and coefficients of determination (R²) were examined to evaluate the structural model. R² is a metric that ranges from 0 to 1, with higher values reflecting greater explanatory and predictive capability. The results indicate that the independent variables in this study explain 74% of the variance in PEOU, 84% in PU, 90% in ATU, and 76% in ITU (see Fig. 3). These values demonstrate that the model constructed in this research possesses strong explanatory and predictive power.

Fig. 3: Path analysis and R2.
Fig. 3: Path analysis and R2.
Full size image

***p < 0.001, **p < 0.01, *p < 0.05. The solid lines represent supported hypotheses.

The research hypotheses were examined using AMOS 26.0. Table 5 presents the outcomes of the hypothesis testing and illustrates the hypothesized model with standardized regression weights (see Fig. 4). Out of the sixteen hypotheses, ten are validated, specifically: H1a, H1b, H2a, H3a, H4b, H5a, H5b, H6, H8, and H11. However, no significant findings substantiate H2b, H3b, H4a, H7, H9, and H10. Furthermore, AMOS 26.0 was utilized for mediating analysis with 5000 bootstrap samples and a bias-corrected percentile approach for a 95% confidence interval. The results reveal no significant mediating pathways. A detailed analysis and discussion are presented in the next section.

Fig. 4
Fig. 4
Full size image

The extended TRAM with standardized regression weights.

Table 5 Hypothesis testing results.
Full size table

Discussion

Optimism (OPT)

This study shows that OPT has a significant positive impact on both PEOU and PU (H1a, H1b). The findings are consistent with the results of surveys conducted by Yusuf et al. (2021) and Kampa (2023) regarding the acceptance and use of E-Learning and M-learning among higher education students. The positive correlation between OPT and both PEOU and PU indicates that the more optimistic CSL learners are, the more they perceive ChatGPT as easy to use and beneficial. Optimistic CSL learners focus more on the conveniences and advantages that ChatGPT provides, allowing them to proactively embrace new technologies.

Innovativeness (INN)

INN was found to have a significant effect on PEOU (H2a) but not on PU (H2b). This result is in line with findings from Seong and Hong (2022) and Nigatu et al. (2024), who argued that INN tends to encourage users to try and engage with new technologies rather than change their judgment of usefulness. For CSL learners, being innovative may help them explore ChatGPT and learn how to use it more smoothly, yet whether the tool is genuinely useful relies more on the system’s available functions and content quality than on the learners’ own traits.

Discomfort (DIS)

In this study, DIS was negatively related to PEOU (H3a) but not to PU (H3b). A similar pattern was observed by Kampa (2023), showing that DIS often shapes judgments about ease of operation but does not necessarily undermine the recognition of usefulness. For CSL learners, even if they sometimes feel pressured or frustrated when starting with ChatGPT, once they are able to control the system they can still see its value in supporting their studies.

Insecurity (INS)

In TR, INS is a negative readiness factor. As expected, INS was associated with lower PU (H4b). By contrast, INS showed a positive association with PEOU (H4a). Similar patterns have been observed (e.g., Yusuf et al. 2021; Cui 2022), suggesting this is not anomalous. A straightforward interpretation is that in settings where ChatGPT use is already routine, privacy or reliability concerns may diminish perceived benefits yet do not make the interface feel harder to operate; if anything, users who persist in using the tool report it as easy to use. This reading is descriptive rather than causal, and it aligns with the idea that risk awareness and ease of operation can co-exist in mature-use contexts (Rashid 2025).

Trust (TRU)

TRU is a new factor introduced in this study. The research findings indicate that the TRU of CSL learners in ChatGPT significantly affects both their PEOU and PU (H5a, H5b). Previous studies have confirmed the impact of TRU on PEOU and PU in various domains, including electronic voting systems and digital currencies (Mannonov and Myeong 2024; Palos-Sanchez et al. 2021; Folkinshteyn and Lennon 2016). However, in the field of educational technology, although research by Shahzad et al. (2024) demonstrates that the moderating effect of TRU significantly enhances students’ adoption of ChatGPT, few studies have discussed the direct relationship between TRU, PEOU, and PU. Our study provides empirical support for this relationship, demonstrating that TRU can significantly enhance learners’ PEOU and PU of AI tools. Moreover, the standardized path coefficients of TRU on PEOU (0.519) and PU (0.443) exceed those of the four TR variables (OPT, INN, DIS, INS). This finding indicates that the TRU variable possesses strong predictive and explanatory power and highlights the usefulness and efficacy of the TRU variable proposed in this study. To some extent, this validates the success of the extended TRAM proposed in this study.

Components of TAM

In summary, the SEM results indicate that PEOU has a significant positive effect on PU (H6). Learners’ awareness of ChatGPT’s benefits also enhances ATU (H8). Consistent with TAM-based research, ATU significantly predicts ITU (H11), supporting the continued relevance of core TAM components in AI-assisted contexts. In our sample, PU explains a large share of ATU (R² = 0.90), indicating that PU is a stable and strong predictor of attitude.

However, not all hypothesized paths were supported. Specifically, PEOU does not significantly predict CSL learners’ ATU (H7). Similar results have been reported in language-learning settings (Liu and Ma 2023), suggesting that perceiving ChatGPT as easy to operate does not necessarily translate into a more favorable attitude toward its use.

More importantly, PEOU and PU did not significantly predict ITU among CSL learners (H9, H10). This finding departs from earlier TAM evidence (Chen and Lin 2018; Beldad and Hegner 2017; Davis 1993), but it is consistent with recent observations in educational technology showing that the effects of PEOU and PU on intention are not uniform across contexts. For example, studies in higher education report non-significant PEOU → ITU effects (Yu et al. 2023; Fan and Wang 2023), and others document non-significant PU → ITU links (Lin and Yu 2023; Wang et al. 2022). Notably, a study with Thai higher-education students found that neither PU nor PEOU significantly predicted learners’ ITU (Shaengchart et al. 2023), directly supporting our pattern in a comparable educational context. In our data, mediation tests did not yield any indirect effects, reinforcing the conclusion that PEOU and PU—at least in our setting—do not operate through attitude-type mechanisms to shape intention.

Taken together, these results point to a mature-use interpretation: when ChatGPT is already familiar and routinely available in CSL learning, PEOU and PU may become normalized—that is, recognized by most learners but no longer decisive for intention. In such conditions, ITU is more likely shaped by readiness- and context-proximal factors that better reflect the distinctive complexity of educational use. Under this reading, PEOU and PU function more as necessary but insufficient conditions, which explains their lost predictive power for ITU in our sample. This is a theory-consistent account rather than a speculative one: it aligns with recent educational findings that report attenuated or non-significant PEOU/PU → ITU links where generative AI has moved from novelty to routine, and it matches our null mediation evidence showing that classical attitudinal pathways did not carry PEOU/PU effects to intention in this sample.

Overall, while these outcomes may initially appear to contradict earlier TAM studies, they offer a precise, evidence-based refinement for AI-supported CSL. In mature-use settings, ITU is no longer primarily driven by PEOU or PU. In the case of CSL learners, once ChatGPT is widely accepted, PEOU and PU cease to be key determinants of ITU. This indicates the need to revise or extend TAM in post-adoption settings and to give more attention to other influences, such as learners’ habits, TR or TRU of use. Future acceptance models for CSL should therefore incorporate these dimensions explicitly rather than assuming uniformly positive PEOU/PU effects on ITU.

Pedagogical implications and suggestions

Studies have shown that TRU is essential to the research model. To further enhance the trust level of CSL learners, two key strategies are essential. First, when CSL learners use ChatGPT to assist in learning Chinese, it records users’ feedback and usage habits, potentially posing risks of personal information and data exposure (Zhang et al. 2023). Therefore, relevant departments need to formulate regulations, policies, and laws to avoid data security risks and protect the personal rights of CSL learners, promoting the safe use of technologies such as ChatGPT under scientific supervision. Second, providing accurate output information is an effective support for building trust. ChatGPT may output a mix of true and false or fabricated information, leading to CSL learners being unable to effectively distinguish factual errors, thus causing cognitive biases. Consequently, it is essential to continuously improve deep learning and machine learning algorithms (Lai et al. 2023) and develop specialized large language models to effectively improve the accuracy and precision of professional knowledge information output.

Research has identified that inhibiting factors in TR may cause CSL learners to develop technology anxiety. To address this situation, higher education institutions should integrate advanced information technologies such as language intelligence and virtual reality into classroom teaching. This integration aims to innovate pedagogical approaches, which are intended to facilitate CSL learners’ adaptation to intelligent teaching methodologies and enhance their digital literacy. Currently, relevant authorities should endeavor to establish a supportive digital learning environment, thereby enabling CSL learners to overcome anxiety regarding digital learning tools. Moreover, reducing dependency on technology is critical. Educational institutions must establish and enforce sensible policies and guidelines, guiding CSL learners toward the responsible use of tools like ChatGPT. Teachers, as supervisors and intermediaries, should ensure that ChatGPT is used as a learning aid rather than a shortcut for completing assignments or cheating.

Research indicates that the positive attitude of CSL learners towards the use of ChatGPT for learning Chinese will directly impact their intention to use it. To enhance CSL learners’ attitudes towards use, ChatGPT can provide diverse information such as text, sound, images, and videos during the process of learning Chinese. This comprehensive approach includes simulating dialogues related to specific scenarios, generating materials for reading at different proficiency levels, and correcting Chinese grammar errors, all of which aim to enhance CSL learners’ listening, speaking, reading, and writing skills, thereby improving learning efficiency. Furthermore, it is important to develop a corpus of prompts for the application of ChatGPT in international Chinese language education. Finally, it is essential to actively utilize ChatGPT’s feedback and evaluation mechanisms to promptly correct errors and avoid the forgetting effect. Providing timely feedback helps CSL learners better accomplish Chinese learning tasks. Leveraging ChatGPT’s role-playing capabilities by setting it as a teacher or mentor and providing personalized guidance based on dialogue data can enhance interactive learning experiences, thereby stimulating CSL learners’ desire and positive attitude towards use.

Conclusions

Major findings

This study employed SEM method to explore the factors influencing the acceptance and use of ChatGPT as a tool for assisting CSL learners in Chinese language acquisition. The key findings of this study are as follows: (1) Theoretical contributions: This study advances the understanding of technology acceptance by extending the TRAM with the integration of TRU, offering a more comprehensive framework for exploring technology adoption in educational contexts. Specifically, the integration of TRU provides a novel dimension for evaluating how TRU impacts both PEOU and PU, which are central constructs in TAM. By focusing on the context of international Chinese language education, this research fills a critical gap, addressing the lack of studies applying TAM or TRAM to the unique challenges and opportunities presented by Chinese language learning with ChatGPT. The study also contributes to the growing literature on AI applications in education by providing empirical evidence on how CSL learners’ TR and TRU shape their attitudes and intentions toward using ChatGPT; (2) Path analysis: The analysis identifies the most significant standardized path coefficients as follows: PU on ATU (0.950), ATU on ITU (0.593), and TRU on PEOU (0.519). Among these, PU emerges as the most stable and robust predictor of ATU; (3) Hypothesis testing: Three hypotheses derived from traditional assumptions of the TAM are strongly supported. However, contrary to expectations, PEOU does not significantly predict ATU, and neither PEOU nor PU significantly affects ITU; (4) Effects of TR: The four dimensions of TR (OPT, INN, DIS, INS) significantly affect the acceptance of ChatGPT through TAM variables. specifically, INN and DIS do not predict PU, whereas INS negatively relates to PU and shows a positive association with PEOU; (5) Role of TRU: As a newly introduced variable, TRU plays a pivotal role in extending the TRAM framework. TRU positively and significantly influences CSL learners’ PEOU and PU, highlighting its importance in understanding their acceptance of ChatGPT.

Beyond extending TRAM with TRU, this study highlights a critical theoretical implication: the diminishing predictive role of PEOU and PU on ITU in mature use contexts. This suggests boundary conditions for TAM and calls for its reconsideration in post-adoption studies.

Limitations and directions for future research

It is essential to acknowledge that current research has certain limitations:

Limited use of qualitative data

First, the study did not sufficiently emphasize qualitative data. Although an open-ended question was included in the questionnaire design, no qualitative analysis such as thematic coding or narrative synthesis was conducted to extract meaningful insights. Specifically, this item was primarily used as a screening tool to assess participants’ actual use of ChatGPT for Chinese language learning. Participants were asked to briefly describe how they utilized ChatGPT in their studies; however, the collected responses were generally brief, lacked depth, and showed a high degree of content homogeneity. As a result, this textual data was not systematically analyzed and contributed minimally to understanding learners’ subjective experiences. Future research should incorporate richer qualitative methods, such as interviews or focus groups, to capture deeper insights into learners’ perceptions, attitudes, and usage patterns. This would facilitate a more robust integration of qualitative and quantitative findings and enhance the interpretability of results.

Sample characteristics and generalizability

This study was limited to CSL learners studying at higher education institutions in China, with a relatively small sample size and notable demographic imbalances (e.g., gender, age, education level, learning years, and country of origin). These factors may affect the generalizability and external validity of the findings. Nevertheless, normality tests and common method bias assessments were conducted, suggesting that the impact of these limitations did not substantially compromise the overall validity of the results. Future studies should consider employing stratified sampling or expanding the sample to include more diverse populations to improve the cross-cultural applicability of the findings.

Variability in participants’ ChatGPT usage

Another potential limitation lies in the variability of participants’ prior exposure to ChatGPT. The study did not set a threshold for minimum usage duration, and some participants may have formed perceptions based predominantly on limited firsthand experience or preconceived notions. This could have influenced their responses regarding acceptance, perceived usefulness, and behavioral intentions. To address this, future research could impose a minimum usage criterion or classify participants based on their usage frequency or duration. This would allow for a more nuanced and objective assessment of attitudes and interaction patterns.

Psychometric properties of the measurement model

Some constructs exhibited AVE values slightly below the commonly recommended threshold of 0.5 (e.g., INN = 0.36, DIS = 0.46). However, CR remained acceptable (0.69–0.88), and standardized factor loadings ranged between 0.522 and 0.888, indicating overall acceptable convergent validity. Given that these values are still within acceptable ranges according to Fornell and Larcker’s (1981) criteria, the measurement model can be considered psychometrically sound. Nonetheless, future studies should consider refining item wording, increasing the number of items, or collecting larger datasets to enhance the psychometric precision of the instrument.

Limited number of items per construct

Although the measurement instrument was adapted from validated scales and demonstrated adequate psychometric properties through CFA, each construct was measured by only 3–4 items. While Cronbach’s alpha, CR, and AVE values were acceptable, a larger item pool could improve the reliability, dimensionality, and generalizability of the constructs across varied learner populations. Future research is encouraged to expand the item sets and further validate the scales across diverse contexts, including longitudinal and cross-cultural settings, to enhance the stability and robustness of the theoretical model.