2.1 Informal digital learning of English (IDLE) and GenAI

With the continuing adoption of technology in education, IDLE has become a well-established field in its own right (Lee & Dressman, Reference Lee and Dressman2018; Lee et al., Reference Lee, Yeung and Osburn2024; Soyoof, Reynolds, Vazquez-Calvo & McLay, Reference Soyoof, Reynolds, Vazquez-Calvo and McLay2023). IDLE investigates self-directed learning with the help of technology, with AI-mediated informal digital learning of English (AI-IDLE) focusing on “how language learners negotiate access to AI tools for self-directed learning of English outside the classroom” (Liu et al., Reference Liu, Zou, Soyoof and Chiu2025: 2). In contrast, traditional computer-assisted language learning (CALL) studies more commonly focus on structured, teacher-led activities within formal educational settings (Reinders, Lai & Sundqvist, Reference Reinders, Lai and Sundqvist2022; Sundqvist & Sylvén, Reference Sundqvist and Sylvén2014). Recent studies have investigated EFL learners’ engagement in AI-IDLE (Liu, Darvin & Ma, Reference Liu, Darvin and Ma2024a, Reference Liu, Darvin and Ma2024b; Liu et al., Reference Liu, Zou, Soyoof and Chiu2025), showing a number of benefits, such as promoting communication behaviors and willingness (Zou, Azari Noughabi, Sabouri & Zhou, Reference Zou, Azari Noughabi, Sabouri and Zhou2025a), improving FLE (Liu et al., Reference Liu, Zou, Soyoof and Chiu2025), and increasing their learning motivation (Liu et al., Reference Liu, Darvin and Ma2024a, Reference Liu, Darvin and Ma2024b; Zou, Liu, Li & Chen, Reference Zou, Liu, Li and Chen2025c ; Zou et al., Reference Zou, Soyoof, Chiu and He2025h). These studies have evidenced that AI-IDLE has played a significant role in L2 education in a digital realm where L2 learners can mobilize their learning psychological factors and negotiate other contextual factors to leverage AI technologies and obtain benefits from self-guided autonomous language learning practices (Guan, Zhang & Gu, Reference Guan, Zhang and Gu2025). It is important to note that AI-IDLE and GenAI-IDLE are used interchangeably because we regard GenAI as an element of AI.

In the context of the Global South, IDLE and (Gen)AI-IDLE have been adopted to address the limitations of formal education systems and provide learners with more accessible, engaging, and autonomous learning opportunities (Zadorozhnyy & Lee, Reference Zadorozhnyy and Lee2025; Zou, Soyoof & Amjad, Reference Zou, Soyoof and Amjad2025e). A number of factors have been shown to play a particularly important role in the adoption and successful application of IDLE. Some of these act as prerequisites, such as access to technological resources (hardware, software, stable internet access). Within this study, particular emphasis is placed on the ways in which L2 learners mobilize GenAI technologies to enact IDLE practices beyond the classroom.

2.2 AI literacy

As new technologies play an ever-greater role in education, learners require an increasingly sophisticated toolkit to navigate their impact. Digital literacy is conceptualized as “the practices of communicating, relating, thinking and being associated with digital media” (Jones & Hafner, Reference Jones and Hafner2021: 12), as well as “the wide range of technology-mediated social practices that learners enact by assembling and interweaving their linguistic, semiotic, and technical resources” (Liu, Reference Liu, McCallum and Tafazoli2025: 1). Considering new technologies (e.g. generative AI) continue to grow and reframe the digital realm (Liu et al., Reference Liu, Darvin and Ma2024a), digital literacy should also advance to include AI literacy to prepare students to become responsible digital citizens in the digital era (Ng et al., Reference Ng, Wu, Leung, Chiu and Chu2024). AI literacy encompasses (1) knowing and understanding AI, (2) using and applying AI, (3) evaluating and corroborating AI-generated content, and (4) incorporating AI-generated content ethically and effectively (Ng et al., Reference Ng, Wu, Leung, Chiu and Chu2024). Studies have shown that incorporating AI in language learning can boost learners’ self-efficacy and motivation (Wang & Wang, Reference Wang and Wang2025), strengthen their learning enjoyment (Liu et al., Reference Liu, Zou, Soyoof and Chiu2025; Zou et al., Reference Zou, Soyoof and Amjad2025e), foster communication behaviors (Liu & Fan, Reference Liu and Fan2025), facilitate linguistic performance (Guan et al., Reference Guan, Zhang and Gu2025), and increase engagement (Zou, Soyoof, Teng & Chen, Reference Zou, Soyoof, Teng and Chen2025f).

2.3 Foreign language enjoyment (FLE) and foreign language boredom (FLB)

The control-value theory (CVT), pivotal in educational psychology research, has seen an increasingly important role in second language acquisition (SLA) studies. CVT focuses on the complicated associations between various academic emotions, their antecedents, and their consequences (Pekrun, Reference Pekrun2006). In light of the three-dimensional taxonomy of CVT, FLE is recognized as a positive, physiologically and psychologically activating, and activity-oriented academic emotion (Dewaele & MacIntyre, Reference Dewaele and MacIntyre2014), while FLB is identified as a negative, deactivation-focused, and activity-related academic emotion (Pekrun et al., Reference Pekrun, Goetz, Daniels, Stupnisky and Perry2010). FLE as a positive emotion with a high level of activation can positively impact language learning, such as by improving learners’ linguistic performance (Dewaele, Botes & Greiff, Reference Dewaele, Botes and Greiff2023), fostering communication behaviors (Lee et al., Reference Lee, Yeung and Osburn2024), enhancing academic engagement (Zou et al., Reference Zou, Azari Noughabi and Peng2025b), and mitigating anxiety (Dewaele et al., Reference Dewaele, Botes and Greiff2023). FLE has been shown to affect informal digital language learning behaviors (Liu et al., Reference Liu, Darvin and Ma2024b; Zou et al., Reference Zou, Soyoof and Amjad2025e, Reference Zou, Soyoof, Teng and Chen2025f) and thus requires further exploration of its antecedents and outcomes.

In contrast, FLB, as a negative emotion characterized by low arousal, adversely affects academic achievement, as it tends to diminish learners’ motivation (Li, Feng, Zhao & Dewaele, Reference Li, Feng, Zhao and Dewaele2024), leading them to rely on superficial learning strategies (Zou, Azari Noughabi & Peng, Reference Zou, Azari Noughabi and Peng2025b) and reducing their enthusiasm and productivity in learning (Wang, Peng & Patterson, Reference Wang, Peng and Patterson2021). A body of literature has corroborated that individuals who experience boredom are more likely to have diverse outcomes: behavioral avoidance (e.g. avoiding communicating with others in English), affective discomfort (e.g. the feeling of monotony), and cognitive dilapidation (e.g. the awareness of time moving sluggishly) (Li et al., Reference Li, Feng, Zhao and Dewaele2024). Specifically, IDLE may trigger learners’ boredom because learners may lack teachers’ support or peers’ discussion in autonomous and self-directed learning environments (Zou et al., Reference Zou, Azari Noughabi and Peng2025b). Research is very limited in exploring the antecedents and consequences of FLB in an informal digital language learning context (Dewaele et al., Reference Dewaele, Botes and Greiff2023). As such, this study aims to bridge this gap by investigating the roles of FLE and FLB with the affordance of GenAI technologies for language learning purposes beyond the classroom.

While research investigating the association among GenAI-IDLE, AI literacy, and L2 emotions remains limited, a body of studies has displayed (AI-mediated) IDLE’s impact on L2 English learners’ emotions (Lee & Drajati, Reference Lee and Drajati2019; Liu et al., Reference Liu, Zou, Soyoof and Chiu2025) and presented AI literacy’s positive influence on FLE (Liu & Fan, Reference Liu and Fan2025). To be specific, the relationship between AI-IDLE and FLE has been corroborated by several studies. Liu and his colleagues found that students who engage in AI-IDLE practices frequently are more likely to present a clearer ideal L2 self-image and a higher level of FLE (n = 299). Wu et al. (Reference Wu, Hapsari and Huang2025) employed quasi-experimental research and found that students’ participation in AI-mediated language learning activities can saliently mitigate their negative emotions, strengthen their positive emotions, and enhance their speaking competence. Moreover, AI literacy may serve as an important predictor of L2 emotions (Liu & Fan, Reference Liu and Fan2025). Specifically, Liu and Fan (Reference Liu and Fan2025) found that AI literacy significantly and positively influences students’ FLE (n = 637). Although these studies mainly focus on EFL formal and structured classrooms, these insights may translate into informal language learning environments outside the classroom with the affordance of GenAI technologies. Given the tight relationship between GenAI-IDLE, AI literacy, and L2 emotions, this study proposes the following five hypotheses:

H1. GenAI-IDLE participation positively predicts AI literacy.

H2. GenAI-IDLE participation positively predicts FLE.

H3. GenAI-IDLE participation negatively predicts FLB.

H4. AI literacy positively predicts FLE.

H5. AI literacy negatively predicts FLB.

2.4 L2 willingness to communicate (WTC) in L2

WTC refers to “a willingness to engage in discourse at a given moment with a particular individual or group utilising L2” (MacIntyre, Clément, Dörnyei & Noels, Reference MacIntyre, Clément, Dörnyei and Noels1998: 547). L2 learners with a higher level of WTC interact more frequently with others in English (Zou et al., Reference Zou, Azari Noughabi, Sabouri and Zhou2025a), create more opportunities to immerse themselves in an authentic L2 usage environment (Dewaele & Dewaele, Reference Dewaele and Dewaele2018), and become more confident and autonomous over time (Zou, Teng, Soyoof & He, Reference Zou, Teng, Soyoof and He2025d). In addition, there have been some studies that have explored the antecedents of WTC in the IDLE environment (Lee & Chiu, Reference Lee and Chiu2023; Lee & Drajati, Reference Lee and Drajati2019). For instance, Lee and Chiu (Reference Lee and Chiu2023) found that Korean EFL students who have a higher level of self-perceived English ability are more likely to reduce offline communication anxiety, thereby yielding higher L2 WTC (n = 1,269). L2 students with a stronger L2 motivation are inclined to display weak anxiety, which in turn elicits a higher level of L2 WTC in digital learning contexts. Thus, it is important to explore the antecedents of L2 WTC, especially some psychological factors (e.g. L2 emotions) and informal digital learning practices. This study combines EFL learners’ AI-IDLE practices, AI literacy, their L2 emotions, and L2 WTC to understand the interactive mechanism of these variables, thereby enriching the theoretical literature.

While previous studies did not validate the predictive power of GenAI-IDLE on L2 WTC, several studies have examined the positive impact of IDLE on WTC in the Global South, including in Iranian, Indonesian, Thai, and Kazakhstani EFL contexts (Zadorozhnyy & Lee, Reference Zadorozhnyy and Lee2025). For example, Zadorozhnyy and Lee (Reference Zadorozhnyy and Lee2025) found that while IDLE did not directly influence L2 WTC in Kazakhstan, self-efficacy beliefs functioned as a full mediator between IDLE and L2 WTC. IDLE has the potential to influence L2 WTC, which is mediated by cognitive or emotional variables. In the same vein, Lee and Drajati (Reference Lee and Drajati2019) found that Indonesian EFL students’ productive IDLE had a significant impact on students’ L2 WTC (n = 183), which echoes the study by Reinders and Wattana (Reference Reinders and Wattana2015) on the effects of Thai university students’ digital game-based language learning behaviors on L2 WTC. In addition, given the positive impact of IDLE on L2 WTC in previous literature (e.g. Lee & Drajati, Reference Lee and Drajati2019; Lee et al., Reference Lee, Yeung and Osburn2024) and how these affective factors can serve as catalysts of L2 WTC within and beyond the formal instruction environments (Lai, Zhu & Gong, Reference Lai, Zhu and Gong2015; Lee et al., Reference Lee, Yeung and Osburn2024), we pose that:

H6. GenAI-IDLE participation positively predicts WTC.

While research on the predictive role of AI literacy in L2 emotions and L2 WTC respectively is very limited, a few studies have explored how L2 English students’ attitudes toward AI affect L2 WTC (Zhi & Wang, Reference Zhi and Wang2024). Zhi and Wang (Reference Zhi and Wang2024) found that Chinese English major students’ attitudes toward AI made a positive and significant impact on L2 WTC (n = 1,090), which echoed Liu and Fan (Reference Liu and Fan2025) by confirming that EFL students with a higher level of AI literacy are more likely to leverage AI technologies to communicate with others in English. Moreover, a large body of literature has examined the predictive roles of L2 emotion in L2 WTC (e.g. Dewaele et al., Reference Dewaele, Botes and Greiff2023; MacIntyre et al., Reference MacIntyre, Clément, Dörnyei and Noels1998), offering empirical evidence for postulating hypotheses regarding the emotional pattern underlying the involvement of AI-mediated digital activities and L2 WTC. For instance, Alrabai (Reference Alrabai2024) found that Saudi EFL students’ FLB fails to directly predict L2 WTC, while FLE significantly and directly predicts L2 WTC (n = 328). However, Wang et al. (Reference Wang, Peng and Patterson2021) found that Chinese EFL learners’ enjoyment and boredom both positively influence L2 WTC within and beyond the classroom (n = 811). To our best knowledge, regarding the interaction between FLB and L2 WTC in digital language learning environments, it is worth further exploring it, especially in AI-mediated digital contexts, because the findings of prior studies were inconsistent (Alrabai, Reference Alrabai2024; Wang et al., Reference Wang, Peng and Patterson2021):

H7. AI literacy positively predicts WTC.

H8. FLE positively predicts WTC.

H9. FLB negatively predicts WTC.

Taken together, the nine hypotheses are vividly demonstrated in Figure 1.

Figure 1. The hypothesized structural model.

2.5 Theoretical framework

As digital technologies have become pervasive and increasingly sophisticated, it is recognized that individuals seek digital experiences that facilitate, rather than impede, the fulfillment of their goals and psychological needs. In this study, we drew upon self-determination theory (SDT) as conceptualized within the framework of technology design by Peters, Calvo and Ryan (Reference Peters, Calvo and Ryan2018). SDT, one of the pioneering theories to integrate psychological constructs into the domain of technology, provides a foundational perspective on motivation, well-being, and autonomy, thereby offering valuable insights into the dynamics of human–computer interaction. The “Motivation, Engagement, and Thriving in User Experience” (METUX) model, which operationalizes SDT in the context of technology use, delineates six distinct “spheres of technology experience”: adoption, interface, task, behavior, life, and society. The “adoption” sphere pertains to users’ autonomous motivation to engage with a given technology. The “interface” sphere addresses how user interaction with technological interfaces (e.g. navigational controls) affects their psychological states. The “task” sphere focuses on the extent to which technologies fulfill users’ psychological needs during task engagement. The “behavior” sphere involves users’ need for satisfaction in relation to the functional affordances of the technology. The “life” sphere encompasses the broader psychological impacts that extend beyond immediate usage.

In our study, the variables are mapped to the METUX spheres according to their theoretical alignment with SDT’s key psychological needs:

1. 1. GenAI-IDLE participation is mapped onto the “Behavior” sphere as it demonstrates learners’ autonomous engagement with AI technologies for informal language learning beyond the classroom. This mapping reflects how learners exercise autonomy in selecting and using AI tools that align with their personal learning goals. The theoretical pathway here indicates that when AI tools satisfy learners’ psychological needs during informal learning activities, this promotes sustained engagement and self-directed learning behaviors.

2. 2. AI Literacy is mapped onto the “Adoption” sphere. The adoption sphere pertains to users’ autonomous motivation to engage with a given technology. We position AI literacy within this sphere because learners’ competencies in critically evaluating and ethically using AI tools directly influence their initial decision to adopt these technologies. This theoretical pathway suggests that AI literacy serves as a prerequisite for autonomous adoption, as learners must understand AI capabilities and limitations before meaningfully engaging with these tools.

3. 3. L2 Emotions (FLE and FLB) are mapped onto the “Life” sphere. The life sphere encompasses the broader psychological impacts that extend beyond immediate usage. We associate L2 emotions with this sphere because FLE and FLB display enduring emotional states that influence learners’ overall quality of life. The theoretical pathway suggests that positive experiences with AI-enhanced language learning can generate enjoyment that transcends specific learning activities, while negative experiences may lead to boredom that affects learners’ WTC.

4. 4. WTC is mapped onto the “Task” sphere, which focuses on the extent to which technologies fulfill users’ psychological needs during task engagement. WTC is positioned within this sphere because it directly relates to learners’ readiness to engage in communicative tasks using the target language. The theoretical pathway indicates that when AI tools effectively support communicative practice (satisfying competence and relatedness needs), learners experience increased WTC, which reinforces their engagement with language learning tasks.

In summary, the integration of SDT with technology design theory aligns closely with the constructs examined in this study (see Figure 2) and allows us to propose our research questions:

Figure 2. Variables mapped to the Motivation, Engagement, and Thriving in User Experience (METUX) spheres.

RQ1. In what ways, if any, does GenAI-IDLE affect Pakistani L2 English learners’ AI literacy and L2 emotions?

RQ2. In what ways, if any, does GenAI-IDLE interact with AI literacy and L2 emotions to impact Pakistani L2 English learners’ WTC?

3.1 Research context and participants

The study was conducted in tertiary-level EFL institutions across Pakistan. Participants were recruited through convenience sampling (Creswell & Creswell, Reference Creswell and Creswell2017) from institutions located in different regions of the country from October 2024 to March 2025 (after the public release of ChatGPT). Invitations to participate were distributed via social media platforms (WhatsApp, LinkedIn, and Facebook), and some teachers assisted in circulating the questionnaire QR code among their students. The sole eligibility criterion was current enrolment at a Pakistani tertiary institution. It should be noted that English is a national language in Pakistan. In Pakistani institutes, teachers use English to teach diverse subjects. To ensure that respondents had relevant GenAI usage experience for English learning purposes, the demographic section included a screening question: “Do you use GenAI to learn English?” Of the 501 questionnaires initially collected, 359 were retained for analysis – those indicating “Yes.”

Due to space limitations, demographic details are shown in Table 1. Ethics approval was obtained from the third author’s institutional review board, and all participants gave their informed consent to voluntarily participate in this study.

Table 1. Background information of participants

3.2 Data collection and analysis

The questionnaire was administered in English using Google Forms (see Appendix in the supplementary material for the questionnaire), achieving cross-cultural validity. It began with a mandatory consent question, before requesting some demographic information, followed by a number of psychometric measurement scales. The scales segment encompassed 37 items assessing five key constructs: GenAI-IDLE, AI Literacy, FLE, FLB, and WTC. Participants evaluated each item utilising a 6-point Likert scale, with responses ranging from 1 (strongly disagree) to 6 (strongly agree) (with the absence of a midpoint in order to reduce central tendency bias):

1. 1. GenAI-IDLE (8 items): We used the 8-item instrument developed and validated by Liu et al. (Reference Liu, Darvin and Ma2024a, Reference Liu, Darvin and Ma2024b, Reference Liu, Zou, Soyoof and Chiu2025) to enquire about respondents’ use of AI in their independent language learning. The AI-IDLE scale exhibited robust internal consistency in our study, with a Cronbach’s alpha coefficient of .90. An example item reads, “I use AI technologies to simulate real-life English language use situations beyond the classroom.”

2. 2. AI Literacy (6 items): We used the scale initially constructed by Chai, Wang and Xu (Reference Chai, Wang and Xu2020) and subsequently validated by Liu and Fan (Reference Liu and Fan2025) within the Chinese EFL context. The reliability of the comprehensive AI literacy scale yielded a Cronbach’s alpha coefficient of .89, indicating robust internal consistency. An example item reads, “I understand how AI technology optimizes the translation output for online translation.”

3. 3. FLE (7 items): We adapted Liu, Zhao and Yang’s (Reference Liu, Zhao and Yang2024d) FLE scale (7 items) within the context of informal digital language learning and incorporated Liu et al.’s (Reference Liu, Darvin and Ma2024b) FLE scale, validated in AI-mediated contexts. Because Liu et al.’s FLE scale was related to online informal engagement with a target language other than English (LOTE), we modified the expressions of their FLE scale to suit the informal engagement within an EFL context. For example, “Learning the target LOTE is fun” was modified into the item “Learning English is fun.” The items on the FLE scale were checked by three experienced experts in applied linguistics and CALL to ensure the content validity of the scale. The reliability of the comprehensive FLE scale yielded a Cronbach’s alpha coefficient of .90, exemplifying excellent internal consistency.

4. 4. FLB (11 items): We used the Foreign Language Boredom Scale developed and validated by Li et al. (Reference Li, Feng, Zhao and Dewaele2024). This scale was validated in the IDLE context by Zou et al. (Reference Zou, Azari Noughabi and Peng2025b). After the content validity of the FLB scale was checked by three experienced experts, 11 items were retained. A Cronbach’s alpha coefficient of .94 indicates excellent internal consistency. An example item reads, “Studying (English) is dull in general.”

5. 5. WTC (5 items): We used the 5-item WTC in the informal digital language learning scale developed and validated by Lee and Lu (Reference Lee and Lu2023). A Cronbach’s alpha coefficient of .90 exemplifies excellent internal consistency. An example item reads, “When I have an opportunity to explain my own culture online in English to other English speakers, I am willing to communicate in English.”

Adhering to the five-step structural equation modeling (SEM) procedure recommended by Collier (Reference Collier2020), data were analyzed using SPSS 29.0 and AMOS 29.0. After data cleaning and screening (e.g. removal of incomplete or invalid responses), normality was assessed through descriptive statistics (means, standard deviations, skewness, and kurtosis). Reliability was examined via Cronbach’s α, and construct validity was evaluated through confirmatory factor analysis. A structural model was then specified to test the relationships among GenAI-IDLE participation, AI literacy, FLE, FLB, and WTC. SEM enables us to model complex relationships between observed and latent variables and direct and indirect (i.e. mediating roles) simultaneously. It should be noted that it can only establish statistical associations between variables, not causal relationships.

4.1 Descriptive statistics

Table 2 showed participants’ self-perception of their engagement in GenAI-IDLE, and their levels of AI Literacy, FLE, FLB, and WTC. The mean values of all constructs ranged from 3.38 to 4.35 (out of a maximum of 6), signifying that respondents exhibited a relatively high level of AI-IDLE, AI Literacy, FLE, FLB, and WTC. FLE showed the highest mean value (4.37), while FLB demonstrated the lowest mean value (3.38). The standard deviations of the items varied from 1.31 to 1.58, indicating an acceptable degree of data dispersion. The kurtosis and skewness values of each item fall within the ranges of < |8| and < |3|, respectively (see Table 2), suggesting that the data set did not substantially deviate from normality (Kline, Reference Kline2023).

Table 2. Descriptive statistics

4.2 Reliability and validity checking

The Cronbach’s α values of the five variables were calculated and found to be .91 (GenAI-IDLE), .89 (AI Literacy), .90 (FLE), .94 (FLB), and .90 (WTC), which all exceeded the benchmark of .70 (see Table 2), thus suggesting excellent internal consistency (Kline, Reference Kline2023). The high internal consistency adds confidence to subsequent SEM outcomes.

To assess the validity of the constructs, both convergent and discriminant validity were evaluated. Following the recommendation by Collier (Reference Collier2020), composite reliability (CR) and average variance extracted (AVE) were employed to examine convergent validity. The high AVE indicates the high proportion of variance explained by each construct. As illustrated in Table 3, the CRs of the five factors surpassed the threshold value of .70, while the AVEs exceeded the cut-off value of .50 (Hair, Black, Babin & Anderson, Reference Hair, Black, Babin and Anderson2009). To ascertain discriminant validity, the square root of AVE for each construct was compared with their respective inter-factor correlation coefficients, which is a conventional analysis method (i.e. Fornell–Larcker criterion). Table 3 shows that the square roots of AVE were greater than their corresponding correlation coefficients, indicating that discriminant validity was established.

Table 3. Convergent and discriminant validity of GenAI-IDLE, AI Literacy, FLE, FLB, and WTC

Note. The values in bold along the diagonal represent the square root of the average variance extracted (AVE). CR = composite reliability.

In a measurement model constructed using AMOS, five constructs and their corresponding observed variables were employed to enhance the construct validity of the adapted questionnaire. To evaluate the degree of model fit between the measurement model and the data set, seven indices were utilized: the ratio of χ² to degrees of freedom (χ²/df < .5), comparative fit index (CFI > .90), incremental fit index (IFI > .90), Tucker–Lewis index (TLI > .90), parsimony goodness-of-fit index (PGFI > .50), root-mean-square error of approximation (RMSEA < .10), and standardized root-mean-square residual (SRMR < .08) (Kline, Reference Kline2023). All model indices fall within the recommended ranges, indicating a satisfactory model fit (Table 4). Hence, the questionnaire data exhibited excellent construct validity.

Table 4. Model fit indices of GenAI-IDLE, AI Literacy, FLE, FLB, and WTC

Note. CFI = comparative fit index; IFI = incremental fit index; TLI = Tucker–Lewis index; PGFI = parsimony goodness-of-fit index; RMSEA = root-mean-square error of approximation; SRMR = standardized root-mean-square residual.

4.3 The structural equation model and mediation model measurement

Employing AMOS 29, we conducted an SEM analysis. Good fit indices met all thresholds recommended by Kline (Reference Kline2023) (Table 5). Learners’ GenAI-IDLE participation significantly and positively predicted AI Literacy (H1: β = .71, p < .001, t-value = 11.08, r = .71). Learners’ GenAI-IDLE involvement made a positive and significant impact on FLE in a direct way (H2: β = 29, p < .001, t-value = 3.89, r = .60), while GenAI-IDLE participation did not predict FLB (H3: β = .13, p = .129, t-value = 1.52, r = .13). AI Literacy exerted a significant and positive role in FLE (H4: β = .43, p < .001, t-value = 5.50, r = .64), while AI Literacy did not predict FLB (H5: β = −.01, p = .884, t-value = −.15, r = .09). GenAI-IDLE did not predict WTC (H6: β = .15, p = .063, t-value = 1.86, r = .44). AI Literacy did not predict WTC (H7: β = −.04, p = .673, t-value = −.42, r = .41). FLE had a significant and positive influence on WTC (H8: β = .53, p < .001, t-value = 6.58, r = .60), while FLB did not have a direct influence on WTC (H9: β −.04, p = .414, t-value = −.82, r = −.03). As illustrated by the R² values in Figure 3, GenAI-IDLE and AI Literacy collectively accounted for 45% of the variance (moderate to high R² values) in FLE and 2% of the variance in FLB. These four variables (GenAI-IDLE, AI Literacy, FLE, FLB) explained 37% of the variance (moderate to high R² values) in WTC. The R² values and correlation coefficients (r) revealed weak effect sizes for three hypothesized relations and large effect sizes for the remaining six hypothesized relationships (Plonsky & Oswald, Reference Plonsky and Oswald2014), indicating the structural model’s robust explanatory power.

Table 5. Hypotheses testing results of direct and indirect paths

Note. r < .25 = small effect size, .25 < r < .5 = medium effect size, while r > .5 = large effect size (Plonsky & Ghanbar, Reference Plonsky and Ghanbar2018). CI = confidence interval. ***p < .001.

Figure 3. The structural equation model of GenAI-IDLE, AI Literacy, FLE, FLB, and WTC Note. n.s. = non-significance; R²: FLE = 45%; FLB = 2%; WTC = 37%. ***p < .001.

To examine the mediating roles of AI Literacy, FLE, and FLB in the association between GenAI-IDLE and WTC, a bootstrapping test with 2,000 samples was performed at a 95% confidence level using AMOS. It was found that FLE played a fully mediating role between GenAI-IDLE and WTC (95% CI [.07, .31], p = .001). AI Literacy and FLE played a full chain-mediating role between GenAI-IDLE and WTC (95% CI [.12, .58], p = .001) because the lower and upper bounds of the two mediating paths were above zero and their indirect effects were both significant, which provided empirical support for a dual pathway model linking behavior, cognition, and affect (Collier, Reference Collier2020). However, AI Literacy did not exert a mediating effect on the relationship between GenAI-IDLE and WTC (95% CI [−.22, .14], p = .736). AI Literacy and FLB did not play a chain-mediating role in the link between GenAI-IDLE and WTC (95% CI [−.01, .02], p = .891) (Collier, Reference Collier2020). FLB did not play a mediating role in the association between GenAI-IDLE and WTC (95% CI [−.04, .01], p = .531).