Introduction

In the process of learning English as a Foreign Language (EFL), learners’ psychological states play a critical role in determining their learning outcomes. With the rapid advancement of technology, AI-assisted learning environments have become increasingly prevalent, reshaping traditional teaching and learning paradigms. Artificial intelligence (AI) technologies, including intelligent tutoring systems, adaptive learning platforms, natural language processing (NLP) applications, and generative AI models, are now widely used in education, including language instruction1,2,3. These tools provide learners with personalized, adaptive, and interactive learning experiences, promoting engagement and reducing cognitive load by facilitating vocabulary learning, grammar correction, and speaking practice4,5,6.

Among the various factors influencing EFL learning outcomes, motivation is consistently identified as a key predictor of language achievement and sustained engagement in learning activities7. Highly motivated learners tend to exhibit greater participation in language classrooms, emphasizing the need to understand what promotes and sustains such motivation in today’s technologically enhanced learning environments8. This study is also informed by principles of positive psychology, which focus on fostering learners’ strengths, motivation, and well-being as a means to enhance their educational engagement9. In language learning contexts, positive psychology encourages autonomy-supportive environments that help learners internalize learning goals and sustain meaningful effort—making it a relevant lens for examining motivation in technology-enhanced classrooms10.

While the technological potential of AI is well recognized, its actual impact in educational settings depends heavily on learners’ acceptance of these tools. AI acceptance refers to learners’ perceptions of the usefulness, ease of use, trust, emotional response, and intention to adopt AI technologies for learning11. It is a psychological construct grounded in models such as the Unified Theory of Acceptance and Use of Technology11 and has become an essential focus in educational technology research. Pan and Wang12 emphasized that in the Chinese EFL context, teachers’ AI literacy is crucial for empowering them as digital citizens, suggesting that AI acceptance is not only a student-centered issue but also a key factor for educators. This emphasizes the importance of understanding AI acceptance across different stakeholders in educational settings.

Recent empirical studies have provided valuable insights into AI acceptance in various educational contexts. For example, Zou et al.13 examined students’ acceptance of an AI-based voice assessment tool and found that perceived usefulness and enjoyment significantly predicted their intention to use it. Chai et al.14 reported that students’ acceptance of educational platforms was positively associated with their engagement in online courses. Roth and Tengler15 focusing on pre-service teachers, emphasized the importance of trust in AI and ethical concerns in shaping users’ willingness to integrate AI into teaching. These studies confirm that learners’ and educators’ acceptance of AI is a key determinant of its educational effectiveness.

In EFL classrooms, where sustained motivation and engagement are critical to successful language acquisition, understanding AI acceptance is particularly important. Learners who view AI tools as accessible, valuable, and aligned with their learning needs are more likely to use them meaningfully and consistently16. Conversely, limited acceptance—due to perceived complexity, lack of trust, or unclear educational benefit—may prevent the effective use of even the most advanced tools17.

This study focuses on how learners accept and interact with AI tools in EFL learning environments. Rather than treating AI use as a purely technological phenomenon, we conceptualize AI acceptance as a learner-centered psychological construct that functions as an external contextual factor influencing internal motivation and learning behavior. Drawing on self-determination theory, this study proposes that favorable perceptions of AI tools can support the satisfaction of learners’ psychological needs, thereby enhancing autonomous motivation. In turn, heightened motivation is expected to promote greater behavioral engagement in language learning. By examining these relationships through a structural equation modeling approach, this study explores whether learners’ motivation mediates the relationship between AI acceptance and behavioral engagement, and whether AI acceptance has a direct influence on engagement as well.

While previous studies have explored AI usage in education, few have investigated the underlying psychological mechanisms—particularly how AI acceptance translates into actual learning behavior through internal motivation. This is especially critical in collectivist educational contexts like China, where motivational dynamics are socially embedded and culturally nuanced.

By integrating Self-Determination Theory with the Unified Theory of Acceptance and Use of Technology, this study addresses a theoretical and empirical gap, offering a more comprehensive explanation of how external technological factors influence internal learner motivation and behavioral outcomes in EFL contexts.

This research contributes to a deeper understanding of technology-enhanced language learning and provides actionable insights for learner-centered AI integration in formal English education.

To address these gaps, the present study is guided by the following research questions:

  1. (1)

    How does AI acceptance influence learners’ motivation in the EFL context?

  2. (2)

    How does motivation mediate the relationship between AI acceptance and behavioral engagement?

  3. (3)

    What is the total effect of AI acceptance on learners’ behavioral engagement?

Literature review

The rapid integration of artificial intelligence technologies into educational contexts, especially English as a Foreign Language (EFL) instruction, has sparked growing research interest in how learners perceive, use, and benefit from these tools. Existing literature has increasingly focused on three interrelated constructs—AI acceptance, motivation, and behavioral engagement—as key factors influencing learning outcomes. This review draws on the Unified Theory of Acceptance and Use of Technology (UTAUT) to understand learners’ acceptance of AI tools, and Self-Determination Theory (SDT) to conceptualize motivation and engagement. Although prior research has examined these components independently, few studies have explored their interrelationships in a unified framework, particularly in collectivist educational settings like China. By synthesizing these theoretical and empirical strands, the present study aims to fill this gap by investigating how AI acceptance influences learners’ behavioral engagement, with motivation acting as a key mediator.

AI acceptance in EFL learning

The integration of artificial intelligence (AI) technologies into English as a foreign language (EFL) teaching is transforming both learning processes and outcomes. This transformation aligns with the Unified Theory of Acceptance and Use of Technology (UTAUT), which argues that technology adoption in educational settings depends largely on learners’ acceptance of the system11. According to the UTAUT, four key predictors influence AI acceptance in EFL education:

  1. (1)

    Performance expectancy (PE): Learners believe that AI enhances language skills, particularly vocabulary acquisition and listening comprehension.

  2. (2)

    Effort expectancy (EE): The perceived ease of AI use affects learners’ willingness to adopt AI tools, favoring user-friendly platforms.

  3. (3)

    Social influence (SI): Peer and instructor endorsements significantly impact technology adoption, especially in collectivist learning environments.

  4. (4)

    Facilitating conditions (FCs): Successful AI implementation depends on adequate technical and institutional support.

Empirical research has provided valuable insights into the acceptance of AI tools in EFL learning. For example, Liu and Ma18 examined how learners use ChatGPT for informal English practice. Their study, based on the Technology Acceptance Model (TAM), found that perceived usefulness was a strong predictor of behavioral intention, which in turn predicted actual AI use for language learning. Similarly, studies by Khan19 and Liang et al.20 showed that learners are more inclined to use AI tools that provide individualized feedback and support learning autonomy. These findings emphasize the importance of learner-centered AI design in driving meaningful adoption and usage.

Behavioral engagement in EFL learning

Behavioral engagement refers to the extent to which learners actively participate in academic activities and is characterized by persistence, concentration, questioning, and participation in class discussions21. In the EFL context, this includes behaviors like completing assignments, asking questions, engaging with language input, and maintaining concentration during language tasks22. Behavioral engagement is often distinguished from cognitive and affective engagement, which involve mental effort and emotional investment, respectively23. While all three are essential, this study focuses specifically on behavioral engagement due to its observable nature and strong association with academic achievement24. Measuring behavioral engagement enables the study to directly assess whether learners’ AI acceptance translates into active involvement in language learning tasks.

From the perspective of Social Cognitive Theory25 learner behavior is influenced by the interaction between personal beliefs and environmental factors. AI tools—such as adaptive feedback systems or conversational agents—can enhance behavioral engagement by reducing anxiety, offering customized support, and promoting a sense of competence26,27. These platforms personalize task difficulty and encourage risk-taking without public embarrassment, thereby supporting learners’ autonomous motivation.

Recent empirical studies reinforce the relevance of behavioral engagement in EFL learning contexts. Alshammari and Alrashidi28 found that among behavioral, cognitive, and affective components of engagement, behavioral engagement had a significant and positive impact on students’ learning achievement in online EFL courses. Similarly, Ge and Wang29 demonstrated that behavioral engagement mediates the relationship between learners’ value beliefs and academic outcomes, showing that students who perceive learning English as meaningful are more likely to actively participate in class and achieve better results. Wu30 drawing on Control-Value Theory, further revealed that academic buoyancy indirectly boosts behavioral engagement via the mediating role of hope, emphasizing the role of emotional resilience in sustaining classroom effort.

Together, these studies emphasize behavioral engagement as a critical indicator of successful EFL learning. They also validate our decision to focus on this construct as the most tangible and educationally actionable outcome of learners’ interaction with AI technologies.

The mediating effect of motivation

Motivation is vital to learner engagement, ultimately being a determinant of the workability of the language learning process. Self-determination theory (SDT), suggested by Deci and Ryan31,32 is useful for understanding motivation in education through intrinsic and extrinsic levels of motivation. Intrinsic forms of motivation occur when individuals partake in inherent enjoyment and satisfaction from the activity themselves33 and extrinsic forms of motivation take into account the rewards linked to an action that is external and societal expectations34. There is also a component of the definition that describes the absence of motivation as amotivation, which refers to an absence of intention and a feeling of helplessness, which is related to poor educational success35.

Concerning EFL learning, particularly within collectivist cultures that are found in many Asian countries, including China, motivation is not only considered an individual trait but can also be contextualized within the broader social context36. In collectivist cultures, the focus on goals for groups and social harmony can shape both extrinsic and intrinsic motivation for language learning37. In these contexts, social expectations and collective values may either enhance or constrain even an individual learner’s motivation38.

AI technologies in education have great potential to interact with these motivational aspects by customizing the individual learning experience while responding to individual learner needs and collective educational goals39. AI tools promote intrinsic motivation by presenting customized challenges that meet the individual learner’s personal interests and proficiency level, in turn creating a sense of autonomy and competence40. Similarly, educational and learning AI platforms can respond to extrinsic states of motivation through gamification components and social learning environments that may encourage motivation within collectivist cultures that reward peer recognition and accomplishment as a group41.

Furthermore, AI-enhanced learning experiences can reduce the absence of motivation (amotivation) by creating activities that feel more meaningful and relevant to the context, supporting learners’ sense of meaning for learning and reducing and perhaps challenging feelings of helplessness and irrelevance often found in traditional education42. These enhancements can also support learning pathways while providing immediate feedback, both of which enhance learners’ feelings of competence, including engagement.

The interaction between AI acceptance and motivation is important. As students accept and begin to adopt the use of AI tools within their practices, they begin to feel significantly more engaged in their learning, promoting higher levels of motivation43. This engagement is not merely additive and is based on physical behaviors; rather, it is more connected to increasing cognitive and affective components because of the supportive features of AI44.

Taking all of this literature into consideration, it is evident that there is a great deal of understanding from the current studies on the use of AI to promote motivation, whereas there have been substantially fewer studies framing the motivational impacts of cultural contexts to mediate AI acceptance and language learning effectiveness. This study attempts to address these gaps by examining the level at which AI acceptance enhances motivation and the effect on changes in motivation, behavioral engagement, and other language development outcomes that take place within EFL instruction through practice, particularly in the sociocultural structures that exist within these collectivist communities.

Theoretical framework

This study is anchored in Self-Determination Theory31 and follows a causal logic of “external environment → internal motivation → behavioral performance.” Within this SDT framework, AI acceptance is conceptualized as an external, social‐contextual input that shapes learners’ perceived usefulness, ease of use, emotional response, and behavioral intention toward AI tools. These perceptions are theoretically grounded in the Unified Theory of Acceptance and Use of Technology11 which has been widely applied to model user acceptance across educational contexts. UTAUT thus complements SDT by providing a structured explanation of how learners’ attitudes toward technology influence the satisfaction of their basic psychological needs—autonomy, competence, and relatedness. When these needs are supported, learners are more likely to internalize academic goals and exhibit autonomous forms of motivation45.

Specifically, UTAUT clarifies how external perceptions—such as performance expectancy and effort expectancy—affect learners’ intention to adopt AI tools. When perceived positively, these tools can enhance psychological needs satisfaction, especially competence and autonomy. This supports SDT’s proposition that contextual enablers influence motivation types along a continuum from extrinsic to intrinsic. Therefore, the two theories together inform the hypothesized pathway: AI acceptance → motivation → behavioral engagement.

Motivation occupies the central mediating role in the model. Drawing on Gonzales and Lopez’s46 six-factor scale, this study adopts their dimensions and simplifies it into “Career Economic, Global Citizen, Communicate Affiliate, Self Satisfaction, Self Efficacy, and Culture Integration”. These dimensions reflect learners’ internalized or partially internalized reasons for language learning, which align with SDT’s higher end of the motivational continuum. By situating motivation within this framework, the study emphasizes the importance of learner autonomy, personal relevance, and psychological fulfillment.

Behavioral engagement is the outcome of interest, defined as learners’ sustained attention, active participation, focused effort, and persistence in academic tasks. SDT literature has consistently shown that engagement driven by autonomous motivation tends to be more self-regulated, energetic, and enduring47. Recent research has further confirmed this link between motivation and learning engagement. For example, Liu et al.48 and Zhao et al.49 found that motivational factors and affective states significantly influenced language learning behaviors and self-perceived proficiency in LOTE and L2 contexts. Given its close link to long-term academic achievement, behavioral engagement is a theoretically valid and practically meaningful indicator of learning success, particularly in EFL contexts where sustained effort is critical.

Accordingly, this study posits that AI acceptance can function as a motivational catalyst, indirectly and directly enhancing behavioral engagement. Empirically testing this model provides deeper insight into learner-centered AI integration in foreign language education.

Research hypotheses

Grounded in SDT, this study explores the mechanism through which AI acceptance influences learners’ behavioral engagement, with a focus on the mediating role of motivation. Based on theoretical reasoning and existing empirical literature, the following hypotheses are proposed:

H1: AI acceptance positively predicts learners’ motivation.

When learners perceive AI tools as useful and supportive, they are more likely to develop internalized learning goals. This is in line with SDT’s view that autonomy-supportive environments promote autonomous motivation45.

H2: Motivation positively predicts learners’ behavioral engagement.

Learners with more autonomous forms of motivation tend to show greater effort and persistence in class, as supported by SDT and prior research on language learning motivation47,50.

H3: Motivation mediates the relationship between AI acceptance and behavioral engagement.

SDT suggests that contextual variables influence behavior indirectly through motivational processes45. Thus, motivation is hypothesized to be the key mediator between AI acceptance and engagement.

H4: AI acceptance has a significant direct effect on behavioral engagement.

Beyond its indirect role through motivation, learners’ favorable perceptions of AI may also directly enhance engagement, as supported by studies applying the UTAUT and TAM frameworks53.

H5: The mediating effects of motivational dimensions vary in strength and significance.

Given the multidimensional nature of motivation, some factors (e.g., media/peer-driven or job-related) may exert a stronger mediating influence than others46,52.

The proposed model, including the hypothesized pathways among AI acceptance, motivation, and engagement, is depicted in Fig. 1.

Fig. 1
figure 1

Hypothesized structural model.

Full size image

Methodology

This study employed a quantitative, cross-sectional survey design to examine relationships among AI acceptance, learning motivation, and behavioral engagement at a single point in time53,54. Such a design minimizes temporal confounds and is well suited for assessing immediate perceptions of AI use in educational contexts.

Participants and sampling

Data were collected via an online questionnaire hosted on the Wen Juanxing platform in Spring 2025. Invitations were distributed through classroom announcements and email lists, targeting undergraduates and graduate students from Wuhan University of Engineering Science and partner institutions in Wuhan, Hubei Province. A total of 803 valid responses were obtained.

As reported in the Results section, the majority of respondents were undergraduates aged 18–25, with a balanced gender distribution. All participants were Chinese EFL learners enrolled in formal English courses. They had varying levels of English proficiency (as measured by CET-4/CET-6 records), and all of them reported using AI tools (e.g., ChatGPT, Doubao, and Deepseek) to assist their language learning.

This sampling approach reflects the typical demographic profile of Chinese university EFL learners, ensuring contextual relevance and interpretability of the findings.

Instruments

The Motivation Scale, adapted from the self-determination theory46 consists of 40 items designed to capture six distinct subdimensions: Career Economic, Global Citizen, Communicate Affiliate, Self Satisfaction, Self Efficacy, and Culture Integration. Each item is rated on a 6-point Likert scale, ranging from 1 = “strongly disagree” to 6 = “strongly agree.”

The AI Acceptance Scale is based on the UTAUT framework11 and the Generative AI Acceptance Scale55. This 20-item instrument uses a 7-point Likert format (1 = “strongly disagree” to 7 = “strongly agree”) to assess perceived usefulness, ease of use, and intention to use AI tools in learning.

The Behavioral Engagement Scale was developed by Guo et al.24. This 38-item scale uses a 7-point format (1 = “very low” to 7 = “very high”) to measure effort, persistence, and participation consistent with established engagement instruments in educational psychology. The scale is specifically designed to assess behavioral engagement in the context of foreign language classrooms.

All scales used in this study demonstrated excellent psychometric properties. Cronbach’s α values ranged from 0.9069 to 0.9894, and Composite Reliability (CR) ranged from 0.9283 to 0.9899, indicating high internal consistency. Average Variance Extracted (AVE) values for all constructs exceeded 0.63, supporting convergent validity. Furthermore, discriminant validity was confirmed using the Fornell–Larcker criterion, with each construct’s AVE square root exceeding its inter-construct correlations.

Data analysis

To ensure data quality, responses were screened based on two exclusion criteria: (a) identical answers across all items, and (b) completion time under 300 s, given the length of the questionnaire. After removing inattentive or invalid cases, a final sample of 654 valid responses was retained for analysis.

For structural analyses, item responses were modeled using both WLSMV (as ordered categorical)56 and ML (as continuous) estimators, following recommendations for Likert-type data with 6–7 response categories. This dual approach enabled robustness checks under different distributional assumptions, enhancing the validity of model interpretation. To enable descriptive comparisons across items measured on different 6- and 7-point scales, both item and composite scores were linearly transformed to a 0–100% of maximum possible score (POMP)57. This transformation preserves the original distributions, correlations, and model estimates while facilitating interpretability and comparability across constructs58,59. Prior to SEM analysis, skewness, and kurtosis values were inspected for all percent-scaled composite constructs, and all values fell within ± 2, indicating acceptable univariate normality. These preliminary checks ensured the validity of subsequent SEM analysis.

All analyses were carried out in R (using the Lavaan package) and Python (with Pandas, Scipy, and Graphviz), following a four-stage procedure. R was selected over commercial software such as Amos or Mplus due to its flexibility, transparency, and full support for customizable bootstrapped mediation procedures. Unlike Amos, which is GUI-based and limited in scripting capability, R allows detailed syntax control, reproducibility, and integration with data preprocessing and visualization tools. Compared to Mplus, R offers comparable statistical performance while remaining open-source and more accessible for iterative model development and reporting. Additionally, Lavaan supports both WLSMV and ML estimation, and enables robust percentile bootstrapping essential for mediation analysis.

First, confirmatory factor analyses (CFAs) were conducted on ordered item responses, comparing a single-factor model, a three‐factor baseline model, and a six‐factor extended model; the fit was judged by χ2/df, CFI, TLI, RMSEA, and SRMR, and nested models were compared with Satorra–Bentler χ2 difference tests. Second, scale reliability and validity were examined: internal consistency was assessed via Cronbach’s α and composite reliability, convergent validity via average variance extracted and standardized loadings, and discriminant validity via the Fornell–Larcker criterion. Third, descriptive statistics (including mean, median, mode, standard deviation, variance, range, interquartile range, skewness, and kurtosis) were computed on the POMP, and Pearson correlations among the main constructs were visualized in a heatmap. Finally, a second-order structural model—specifying Motivation as a higher-order factor—was fitted with Weighted Least Squares Mean and Variance Adjusted Estimator to estimate global fit and structural paths, and mediation effects were tested under Maximum Likelihood (ML) estimation with 5,000-draw bootstrap to obtain percentile confidence intervals for both the overall indirect effect and the specific indirect effects through each motivational subdimension. All hypothesis tests were two-tailed with α = 0.05.

Ethical considerations

This study adhered to established ethical standards in educational research. Prior to data collection, the research protocol was reviewed and approved by the Ethics Committee of the School of Humanities at Wuhan University of Engineering Science (Approval Number: WUOES20241202). All participants were provided with a clear explanation of the study’s purpose, procedures, and their rights before giving informed consent.

Participation in the study was entirely voluntary and anonymous. Respondents were assured that their data would be used solely for academic purposes and that their responses would remain confidential. Participants had the right to withdraw from the survey at any time without penalty. No personally identifiable information was collected at any point in the data collection process.

In line with data integrity standards, only valid and attentive responses were retained for analysis after applying exclusion criteria. The study was conducted in accordance with the ethical guidelines set forth by the American Psychological Association (APA) and relevant institutional policies.

Results

This section presents the empirical findings of the study in four stages. First, the measurement model is evaluated using confirmatory factor analysis (CFA) to assess construct validity and reliability. Second, descriptive statistics and correlation analyses are conducted to provide an overview of variable distributions and interrelationships. Third, the structural equation model (SEM) is estimated using the WLSMV estimator, testing both direct and indirect pathways among AI acceptance, motivation, and engagement. Finally, mediation analyses assess the significance and strength of specific motivational subdimensions in explaining the observed relationships.

Measurement model

A series of nested confirmatory factor analyses (CFA) was conducted on the original 6- and 7-point items using the WLSMV estimator and treating all items as ordered. Three models were compared: a single-factor model (common-method test), a three-factor baseline model (AI Acceptance, Motivation, Engagement), and a six-factor extended model (six motivational subdimensions plus AI Acceptance and Engagement).

As shown in Table 1, the single-factor model showed extremely poor fit, with a CFI/TLI < 0.96 and an RMSEA of 0.2483—far above the 0.08 threshold—indicating strong evidence against a one-factor solution. The three-factor model yielded markedly improved fit, though the RMSEA (0.0881) was slightly above the conventional cutoff. Finally, the six-factor model exhibited excellent global fit (CFI = 0.9966; RMSEA = 0.0660; SRMR = 0.0416), meeting all recommended criteria (χ2/df < 5; CFI/TLI > 0.95; RMSEA < 0.08). Model fit was evaluated based on commonly accepted cut-off criteria: CFI and TLI values greater than 0.95, RMSEA values less than 0.08, and SRMR values less than 0.08 were considered indicators of good model fit. All indices in the six-factor model satisfied these thresholds, supporting the adequacy and discriminant structure of the measurement model.

Table 1 Model fit indices for competing CFA models.
Full size table

Nested chi-square difference tests with Satorra–Bentler correction (Table 2) confirmed that the three-factor model fit significantly better than the single-factor model (Δχ2(3) = 1 143.00, p < .001) and that the six-factor model fit significantly better than the three-factor model (Δχ2(25) = 375.75, p < .001). Together, these results demonstrate that (a) a single common factor does not account for the covariance among all items, (b) separating AI Acceptance, Motivation, and Engagement yields a substantially better fit, and (c) modeling the six motivational subdimensions as distinct factors further improves model fit. Accordingly, the six-factor measurement model was retained for all subsequent reliability, validity, and structural analyses.

Table 2 Nested χ2-difference tests.
Full size table

Descriptive statistics and correlation analysis

Table 3 reports the descriptive statistics for all primary constructs, each of which was transformed to a uniform 0–100%-of-maximum scale prior to analysis. Mean scores ranged from 62.58% (Engagement) to 74.90% (Career Economic motivation), suggesting moderately high levels across constructs. Standard deviations ranged from 18.34 to 21.77%, indicating adequate dispersion.

All skewness values fell between − 0.45 and 0.12, suggesting approximately symmetric distributions. Kurtosis values ranged from 2.34 to 2.79, close to the normal reference value of 3. The full range (up to 100%) and IQRs above 25% further confirm that the transformed scores retained variability.

Table 3 Descriptive statistics of main constructs (% of maximum Score).
Full size table

All distributions showed acceptable symmetry and spread, indicating no violations of normality assumptions for subsequent parametric analyses.

Figure 2 displays the demographic characteristics of the sample. The majority of participants were aged 18–20 years (73.5%), followed by 21–23 years (21.7%), with smaller proportions aged 24–26 years (2.1%) or 27 years and above (2.6%). Females comprised 68.0% of the sample, males 28.6%, and 3.4% preferred not to say. Undergraduates dominated, with 39.8% freshmen, 30.0% sophomores, 18.3% juniors, 7.3% seniors, and 4.6% graduate students. Institutional affiliation was split between Wuhan University of Engineering Science (43.9%), other institutions (46.8%), and Wenhua College (9.3%).

Fig. 2
figure 2

Demographic characteristics of the age, gender, academic year and institutional affiliation distributions.

Full size image

The demographic profile suggests a relatively homogeneous population of young, undergraduate EFL learners—ideal for investigating motivational mechanisms in AI-assisted learning.

Figure 3 presents the Pearson correlation matrix computed on the scaled variables. AI acceptance was positively correlated with behavioral engagement (r = .56, p < .001) and with each of the six motivational subdimensions (r = .49–0.57, all p < .001).

Likewise, behavioral engagement showed strong positive correlations with the motivational dimensions (r = .52–0.64). The motivational constructs themselves were also highly intercorrelated (r = .66–0.87), suggesting conceptual coherence and reinforcing the multidimensional nature of motivation in this model.

Fig. 3
figure 3

Heatmap of Pearson correlations among main constructs.

Full size image

These correlational findings align with the hypothesized model structure and validate the premise that motivation mediates the relationship between AI acceptance and engagement.

Structural model analysis

The second-order structural equation model was estimated using the WLSMV estimator, which is appropriate for ordered-categorical item responses. As shown in Table 4, the model demonstrated excellent global fit, with all indices meeting or exceeding recommended thresholds (χ2/df < 5; CFI/TLI > 0.95; RMSEA/SRMR < 0.08). These results indicate that the proposed second-order model, which positions Motivation as a latent construct encompassing six subdimensions, adequately represents the observed data.

To further justify the estimator choice, we compared model fit indices under both WLSMV and ML estimation. As shown in Table 5, WLSMV yielded superior fit across all indices. In contrast, ML estimation—which assumes continuous and normally distributed variables—produced substantially lower CFI (0.853) and TLI (0.849), falling below the conventional 0.90 threshold. This discrepancy reflects the mismatch between ML’s distributional assumptions and the ordinal nature of our Likert-type responses, reinforcing the decision to adopt WLSMV for structural path interpretation.

Table 4 SEM global fit indices (WLSMV).
Full size table

The model fit indices collectively support the appropriateness of the proposed second-order structure for examining direct and indirect effects among AI acceptance, motivation, and engagement.

Table 5 Fit indices comparison for WLSMV vs. ML Estimation.
Full size table

The standardized path coefficients derived from WLSMV estimation are presented in Table 6. All hypothesized direct paths (H1, H2, and H4) were statistically significant and in the expected direction. Specifically, AI acceptance significantly predicted motivation (β = 0.628, p < .001), which in turn significantly predicted engagement (β = 0.517, p < .001). Additionally, AI acceptance retained a significant direct effect on engagement (β = 0.253, p < .001), indicating the presence of both direct and mediated influences.

Table 6 Standardized structural path coefficients (WLSMV).
Full size table

These results provide robust support for the hypothesized model, confirming both the mediating role of motivation and a residual direct effect of AI acceptance on learner engagement.

Figure 4 visually illustrates the structural model. As shown, all hypothesized relationships (H1, H2, H4) were statistically supported.

Fig. 4
figure 4

Second-order SEM path diagram of AI acceptance, six motivation subdimensions, and engagement. Note. Motivation is modeled as a second-order latent construct composed of six reflective subdimensions, which are not shown here for visual clarity.

Full size image

Mediation analyses were conducted using maximum likelihood estimation with 5,000 bootstrap resamples, as the WLSMV estimator in Lavaan does not support bootstrap confidence intervals. The total indirect effect of AI acceptance on engagement through overall motivation (H3) was significant (Estimate = 0.338, 95% CI [0.259, 0.420], p < .001), indicating that motivation partially mediates this relationship (see Table 7).

Among the six subdimensions of motivation, three showed significant mediating effects (H5):

  1. (1)

    Self-satisfaction (std.all = 0.189, p < .001).

  2. (2)

    Self-efficacy (std.all = 0.146, p = .001).

  3. (3)

    Career economic motivation (std.all = 0.084, p = .013).

In contrast, the indirect paths via Global citizen, Communicate affiliate, and Culture integration were nonsignificant, suggesting that self-oriented motivational constructs play a more central mediating role in translating AI acceptance into engagement.

Table 7 Parallel mediation effects (ML + 5 000-draw Bootstrap) H3 & H5.
Full size table

Table 8 provides a concise summary of the hypotheses tested in the study, including path coefficients, significance levels, and mediation effects. All core hypotheses were supported, with partial support for H5 indicating differences in the mediating strength of motivational subdimensions.

Table 8 Summary of hypothesis testing.
Full size table

These findings confirm that motivation mediates the effect of AI acceptance on engagement, particularly through self-related motivational constructs. This emphasizes the critical role of learners’ internalized goals in determining how technology influences engagement behavior.

Discussion

To clarify the theoretical contributions and empirical support, the results confirmed that all core hypotheses were supported, aligning with the hypothesized mediation model. Notably, Self-Satisfaction, Self-Efficacy, and Career Economic motivation emerged as significant mediators, whereas Global Citizen, Communicate Affiliation, and Cultural Integration did not yield significant indirect effects. These differentiated outcomes provide a nuanced understanding of how specific motivational dimensions operate within AI-integrated learning environments.

This study explored how Chinese EFL learners’ acceptance of AI tools influences their behavioral engagement, with a particular emphasis on the mediating role of motivation. The SEM results support a partial mediation model: AI acceptance positively predicted both motivation and engagement, while motivation acted as a significant mediator in this relationship.

These findings strengthen the theoretical linkage between the Unified Theory of Acceptance and Use of Technology11 and Self-Determination Theory45. Specifically, AI acceptance—representing learners’ perceived usefulness and ease of use—was found to enhance internal motivational states, which are central to SDT, and these motivational states in turn promoted behavioral engagement. This integrative framework offers a more comprehensive explanation of how external technological factors interact with internal psychological needs40,50.

The standardized indirect effect (0.338, 95% CI [0.259, 0.420]) underscores that learners who accept AI tools are more motivated, and this motivation serves as a key driver for active classroom participation. Self-Satisfaction and Self-Efficacy were the strongest motivational predictors of engagement. This suggests that when AI tools support a learner’s sense of competence and personal growth, the likelihood of active behavioral engagement increases.

In contrast, motivational dimensions such as Global Citizen and Cultural Integration did not significantly mediate the relationship. This indicates that many current AI tools may be limited in their ability to address culturally oriented or globally framed motivations, likely due to their localized or generic design.

The analysis of six motivational subdimensions provided a deeper understanding of how different motivational factors interact with AI-supported learning environments. Among these, Self-Satisfaction and Self-Efficacy emerged as the most significant mediators of engagement, indicating that learners who feel competent and personally fulfilled by using AI tools are more likely to engage in tasks. This is consistent with earlier studies (e.g., Gupta et al.42), which emphasized that learner autonomy and personalized interventions promote intrinsic motivation and active participation. Career Economic motivation was also a significant mediator (ind_CE = 0.083, p = .013), suggesting that learners who view AI tools as helpful in achieving their career goals are more likely to engage actively. However, dimensions related to Cultural Integration did not demonstrate a significant effect, which could be attributed to the limited global content and cultural relevance of the AI tools used in this study.

From a practical standpoint, these findings provide several actionable insights for educators, policymakers, and AI tool developers:

  1. (1)

    EFL educators should adopt AI tools that promote autonomy and competence—psychological needs shown to mediate AI’s effect on engagement.

  2. (2)

    AI developers are encouraged to embed adaptive feedback, personalized learning paths, and scaffolding to align with learners’ motivational profiles.

  3. (3)

    Teacher training programs should go beyond technical training and emphasize motivational pedagogy, equipping instructors to use AI meaningfully in classroom settings.

These points reflect the theoretical tenets of SDT and UTAUT, emphasizing that both internal and external factors are essential for sustaining engagement.

Several limitations should be noted. The cross-sectional nature of the study limits causal inference. Future longitudinal or experimental research could better capture how motivation and engagement evolve with extended AI use. Additionally, reliance on self-reported data introduces the possibility of bias; future studies might integrate behavioral or observational measures. Moreover, while this study focused on Chinese EFL learners, future cross-cultural comparisons could reveal how sociocultural contexts shape learners’ AI acceptance and motivational profiles.

Finally, the absence of significant effects from some motivational subdimensions suggests the need for AI tools that better address culturally and globally oriented learner goals. Future research could explore how global collaboration features or intercultural content influence motivation and engagement.

Conclusion

In conclusion, this study reveals that AI acceptance significantly enhances EFL learners’ behavioral engagement, both directly and through increased motivation, particularly in the domains of self-efficacy and self-satisfaction. By integrating UTAUT and Self-Determination Theory into a unified model, this research contributes both theoretically and practically to the field of learner-centered AI use. These findings emphasize that psychological factors—especially learners’ internalized goals and perceived competence—play a central role in determining the effectiveness of AI-supported learning environments. Educators and developers should therefore consider not only the technical functionality of AI tools, but also their alignment with learners’ motivational needs. Moreover, the differential effects across motivational subdimensions suggest that AI tools should be customized to address diverse learner profiles, especially in culturally diverse contexts.

From a pedagogical and policy perspective, the findings suggest that educational institutions should provide structured support for AI integration, including teacher training programs focused on motivational pedagogy and learner autonomy. Policymakers may also consider developing guidelines that promote the ethical and learner-centered use of AI technologies in language education. Such measures can help ensure that AI tools are not only technically effective but also pedagogically meaningful across diverse learning settings.

Future research should explore how sociocultural variables and AI design features jointly shape motivational processes and learning outcomes, ideally through longitudinal or cross-context comparisons. As AI technologies continue to evolve, their responsible and psychologically informed integration will be essential for maximizing their educational impact.