Anxious but Active: The Additive Effects of Generative AI and Teacher Support on L2 Willingness to Communicate

Anxious but Active: The Additive Effects of Generative AI and Teacher Support on L2 Willingness to Communicate | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Anxious but Active: The Additive Effects of Generative AI and Teacher Support on L2 Willingness to Communicate CHEN XIN, WEIHE ZHONG This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9578013/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 5 You are reading this latest preprint version Abstract Background Even though generative artificial intelligence is transforming language teaching worldwide, little research has investigated how Perceived Teacher Support (PTS) and AI Partner Support (AIPS) jointly influence learners' Second Language Willingness to Communicate (L2 WTC) and affective states. Methods A cross-sectional survey was administered to 249 EFL learners at a STEM-focused university in China. Hierarchical regression, confirmatory factor analysis, exploratory moderation analysis, and multi-group analysis were conducted to test the proposed hypotheses. Results Findings are reported at two levels. At the confirmatory level, the additive positive effect of PTS and AIPS on L2 WTC was consistently supported across all analytical stages (Δ R² = .317), establishing the dual-pathway human-AI support structure as a robust and replicable predictor of communicative intention within STEM learning ecologies. At the exploratory level, ELA did not emerge as a significant inhibitor of L2 WTC, yet yielded a positive bivariate correlation ( r = .298). This dual-level anomaly departs from classical SLA predictions. An exploratory moderation analysis provided initial evidence against functional substitution as the primary driver, lending indirect support to a suppressor variable interpretation; longitudinal designs are needed to adjudicate among competing accounts. Exploratory multi-group analysis (MGA) detected a directional cross-disciplinary signal that did not reach conventional significance (bootstrap p = .088; post-hoc power = 62%), and is therefore treated as hypothesis-generating rather than confirmatory evidence of discipline-specific differences. Conclusions These findings represent preliminary empirical evidence whose generalizability requires replication across more diverse institutional and demographic contexts. AI partner support Perceived Teacher Support L2 willingness to communicate English language anxiety Emotion-behavior decoupling STEM learners Figures Figure 1 Figure 2 1. Background 1.1 L2 Willingness to Communicate in STEM Contexts L2 WTC is not a fixed personality trait. MacIntyre et al. [ 1 ] define it as a learner being ready to use L2 to communicate with a specific person at a specific moment. This is a state-level construct shaped by situation and context [ 2 ][ 3 ]. As an immediate psychological link before behavior execution, it is a direct predictor of L2 oral communication, playing a mediating role between learners' emotional states and language expression activities [ 4 ]. The heuristic pyramid model by MacIntyre et al.[ 1 ] portrays this dynamic fluctuation. Subsequent empirical studies have clarified its key influencing factors. Anxiety weakens the positive effect of motivation [ 5 ]. Social support and anxiety relief are regarded as important promoters of WTC [ 6 ]. What has not been fully explored is how these relationships will change as the digital learning ecology reshapes the communication environment. This question has not been answered by the core predictive variables identified by Wei and Xu [ 4 ]. The classical second language acquisition (SLA) theory regards ELA as the primary affective barrier to L2 WTC[ 1 ][ 7 ]. However, this linear hypothesis is difficult to account for the actual situation in the STEM academic context. STEM students are largely driven by ought-to L2 self [ 8 ] and face the rigid need to communicate for academic survival - including subject defense, transnational laboratory cooperation and technical literature reading [ 9 ]. This kind of communication pressure still exists structurally in the contemporary STEM environment: empirical evidence from Chinese EMI settings shows that STEM students continue to face language-related challenges, including oral report anxiety and insufficient English proficiency when dealing with high-risk academic tasks [ 10 ][ 11 ]. This necessity existed before the AI era, but its core mechanism is still effective: STEM students must maintain L2 output even under strong ELA. L2 WTC is a direct psychological precursor of communication behavior [ 1 ]. Therefore, learners can still maintain communication output under sustained anxiety, precisely because they retain a correspondingly high willingness to communicate. This paradox manifests not only at the level of external behavior but also at the level of communication intention. According to Self-Determination Theory (SDT) [ 12 ], this pattern aligns with identified regulation: learners internalize the instrumental value of English as academic survival needs and sustain communication output under persistent emotional pressure. In STEM situations, L2 WTC is driven by real necessity rather than intrinsic pleasure; this instrumental identification adjustment overwhelms the classic inhibition effect of anxiety. With the rapid development of generative AI, conversational AI now provides a non-evaluative environment for oral practice [ 13 ][ 14 ]. However, most of the existing literature examines traditional interpersonal support and AI-driven support in isolation (such as [ 15 ][ 16 ]). In high-risk STEM situations, it is not clear how humans and AI systems complement each other to maintain communication. This study responds to this research gap through the Human-AI Symbiosis framework and explores how PTS and AIPS work together under high-intensity academic pressure. 1.2 English Language Anxiety and Emotion-Behavior Decoupling ELA is a situation-specific construct, which is different from general trait anxiety [ 7 ]. Teimouri et al. [ 17 ] found that there is a significant negative correlation between anxiety and learning achievement; Dewaele [ 18 ] found that high ELA is associated with L2 WTC reduction.These findings are the current mainstream views. However, Alpert and Haber [ 19 ] have long pointed out that anxiety is not always an obstacle. Its facilitative aspect can maintain rather than inhibit performance. Therefore, in a task-oriented STEM context, moderate ELA may play a role in this way [ 20 ]. Specifically, under the well-designed task situation, anxious learners sometimes enter a positive flow state [ 21 ]. The standard interpretation regards ELA as an affective filter variable between environmental support and L2 WTC [ 22 ]. STEM groups may not follow this pattern. Communicative imperative [ 9 ] ignores the emotional state and puts forward continuous second language output requirements for learners, which may make them enter the "emotional-behavioral separation" mode. This is different from the situation described in the classical theory of resilience [ 23 ][ 24 ]. Learners do not internally regulate negative emotions, but bypass emotional distress from the structural level in the digital ecological environment and transfer the emotional burden to the external scaffolding instead of dealing with it from the inside. 1.3 The Dual-Support Ecology 1.3.1 The Double-Edged Sword of Perceived Teacher Support (PTS) Perceived Social Support (PSS) reflects learners' subjective assessment of available support [ 25 ]. This study converts the authority-based dimensional operation in PSS into Perceived Teacher Support (PTS). House [ 26 ] divides it into four functional forms: informational, instrumental, emotional and appraisal. These forms of support are distributed from different sources such as peers, families and teachers. This study does not examine all functional forms, but adopts a source-based conceptualization method to separate the evaluation dynamic mechanism unique to authority relations. Its theoretical basis is that in the context of Chinese culture (CHC), the support provided by authoritative figures is essentially related to the expectation of "face" and social evaluation [ 27 ]. This dynamic mechanism is determined by the source of support rather than its functional form, and it cannot be effectively portrayed by functional classification alone. Therefore, this study operationalizes the interpersonal-oriented PSS as PTS, focusing on the support of authoritative sources (i.e. teachers and institutional support). PTS is a recognized positive predictive factor of L2 WTC [ 15 ], but its vertical structure implies social evaluation attributes [ 28 ]. Under the background of CHC, this has formed a double-edged dynamic: the same support provides academic support while conveying evaluation expectations, activating "face" concerns and rigidity of standards [ 27 ]. The resulting "forced compliance" [ 29 ] maintains communication behavior through external regulation at the cost of an increase in ELA. This association is not purely theoretically derived. In the situation of CHC and Chinese English as a foreign language (EFL), teachers' strictness and authoritative evaluation have been proven to be a direct trigger for anxiety[ 18 ], and students report a stronger tendency to remain silent under strict teacher expectations. The expected PTS-ELA correlation in this study is based on this empirical pattern. 1.3.2 AI Partner Support (AIPS) as a Parallel Safety Valve Generative artificial intelligence (GenAI) (such as ChatGPT and Doubao) provides a unique form of support, which is reflected in AIPS. There is no "social self" in AIPS[ 30 ], which is baseline-agnostic and free of evaluative gaze. This non-evaluative and high error tolerance environment has been proven to significantly reduce ELA levels [ 14 ][ 16 ], including in the context of high-risk tasks [ 31 ]. Artificial intelligence-assisted interaction also yields linguistic gains and motivational effects comparable to face-to-face peer learning [ 32 ][ 33 ]. AIPS does not change the evaluative nature of teacher support. On the contrary, it provides an independent space free from fear of negative evaluation, and becomes an emotional safety valve for students under the pressure of academic anxiety [ 13 ] [ 33 ]. This makes AIPS an independent parallel path rather than a moderating variable [ 34 ]: it works in parallel with interpersonal support without changing its nature, while absorbing the pressure generated by authoritative channels. 1.4 Research Gaps and Hypotheses This study stems from two theoretical gaps. One involves the behavioral driving mechanism: the SLA research paradigm regards ELA as a behavioral inhibitor, but whether the STEM group operating in the digital ecology really shows the phenomenon of "emotion-behavior decoupling" has not been directly tested. The second involves dual support dynamics: the existing model's evaluation of PTS and AIPS is independent of each other. How authority-based PTS and non-evaluative AIPS interact in the same ecology is still an open question. It is not clear whether the two will compound or cancel each other's emotional effects. Based on the human-AI Symbiosis framework and the psychological characteristics of STEM learners in the CHC, this study puts forward the following hypotheses: H1 (corresponding research question RQ1): PTS and AIPS both have an additive effect, and significantly positively predict the learners' L2 WTC. H2 (corresponding to the research question RQ2): There is a differentiated predictive relationship between the two support systems and ELA. Due to the implicit evaluation pressure, PTS predicts ELA positively; in contrast, AIPS is not a significant source of pressure (that is, after controlling for PTS, it has no significant positive prediction effect on ELA) and plays the function of a non-evaluative safety valve. H3 (corresponding research question RQ3): Under high-intensity academic pressure, it shows the characteristics of "emotion-behavior decoupling", and ELA has no significant negative predictive effect on the L2 WTC of STEM learners. This study examines the dynamics of human-AI collaboration through an integrated analytical framework of PTS, AIPS, ELA and L2 WTC. Three research questions are addressed: RQ1: How do PTS and AIPS jointly shape the L2 WTC of STEM learners? RQ2: How do PTS and AIPS differentially influence learners' ELA? RQ3: Is ELA a significant behavioral inhibitor of L2 WTC, or will emotion-behavior decoupling emerge among STEM learners under intense academic pressure? This study contributes to SLA theory and teaching practice from three levels. First, about the dual-support learning ecology. This study does not regard AI as a traditional "pressure buffer". Instead, it examines how AI, as an independent and nonjudgmental affective safety valve, provides complementary affective scaffolding to offset the evaluative pressure inherent in authority-based support. Second, about the affective filter mechanism. In this dual-support ecology, ELA does not constitute a significant behavioral inhibitor of L2 WTC, which is inconsistent with the prediction of classical SLA theory. We propose Pragmatic Resilience as a conceptual label to guide subsequent research. Its association with identified regulation will be explained in the discussion section[ 8 ][ 12 ]. Third, about instructional design. By building a Strategic Sandwich (Human-AI-Human) heuristic model, this study provides a testable teaching process for educators and subsequent researchers, and hands over the anxiety-inducing trial-and-error process to AI, in order to optimize learners' communicative performance in high-stakes contexts. To enhance the rigor of the research and explore potential boundary conditions, this study adds an exploratory step.As an exploratory step, this study uses MGA to test the potential heterogeneity between different disciplines. The cross-group results obtained are only used to generate hypotheses and do not constitute confirmatory evidence of disciplinary differences. Note PTS = Perceived Teacher Support; AIPS = AI Partner Support; ELA = English Language Anxiety; L2 WTC = L2 Willingness to Communicate. Solid lines denote hypothesized direct paths. The model illustrates the dual-pathway additive effect on L2 WTC and the emotion-behavior decoupling between ELA and L2 WTC. 2. Methods 2.1 Participants This study adopts convenience sampling to select English as a Foreign Language (EFL) learners from a STEM-focused university in northern China. This research setting enables direct access to the intersection of academic pressure and technological integration. A total of 249 valid questionnaires were retained in this study. The sample includes science and engineering non-English majors (61.8%, n = 154) and English majors (38.2%, n = 95), both from the same institution. This study adopted a same-institution design after careful consideration: by holding constant campus culture, hardware resources, and the university’s AI integration level, the observed structural differences are more likely attributed to disciplinary differences than to confounding variables at the macro level. The gender distribution (66.7% male, 33.3% female) reflects the reality of STEM majors but also limits the generalizability of the findings in female-dominated academic contexts. A supplementary independent-samples t-test showed no significant gender differences in English Language Anxiety (ELA) and AI Partner Support (AIPS) scores (all p > .05), indicating that the observed structural patterns were not mainly driven by gender composition. According to structural equation modeling (SEM) criteria, the sample size of this study is adequate:Kline [ 35 ] suggests a minimum sample size of 200 for complex path models; with fewer than 30 observed variables, the sample size ( N = 249) meets the 5:1 observation-to-variable ratio recommended by Hair et al. [ 36 ]. All participants were enrolled in blended courses and actively used generative AI tools for language learning. 2.2 Measures All constructs are measured using a 5-point Likert scale. This study conducted confirmatory factor analysis (CFA) and reliability tests, and then used item purification to optimize model fit. Item purification proceeded in two stages. In the statistical purification stage, the initial CFA excluded AIPS item A1 and L2 WTC item C1, as their factor loadings fell below 0.50 or showed severe cross-loadings. For theoretical purification, we adopted an a priori approach and excluded items D1–D3 and D7–D9 from the PTS scale to isolate the authority-based evaluative dimension central to this study’s research questions. The relevant cultural and psychometric rationale appears in detail in Section 2.2.1 . 2.2.1 Perceived Teacher Support (PTS) This study adapted the Multidimensional Perceived Social Support Scale (MSPSS [ 25 ]), and only the teacher support items were retained (D4–D6). The exclusion of peer and family dimensions is based on two reasons. At the theoretical level, according to the framework of House [ 26 ], the above dimensions reflect the non-evaluative buffering function, which will interfere with the social-evaluative threat that this study is concerned about. At the empirical level, a supplementary confirmatory factor analysis that retains all dimensions confirmed the decision: incorporating peer and family items will significantly reduce the model fit (CFI: .929 → .878; RMSEA: .065 → .078). It warrants explicit acknowledgment that items D4–D6 operationalize teacher caring and academic scaffolding rather than evaluative threat per se. The theoretical linkage between authority-based scaffolding and anxiety activation is inferred through CHC cultural logic rather than directly measured. Within CHC relational frameworks, however, this inferential step is not arbitrary: Hwang [ 27 ] argues that authority-based support is constitutively inseparable from implicit face expectations and social appraisal, such that the provision of scaffolding by an authority figure inherently activates evaluative concerns at the cultural-construct level. Empirical evidence from Chinese EFL classrooms corroborates this cultural logic: teacher strictness and authoritative evaluation are direct triggers of learner anxiety, with students reporting stronger silence tendencies under demanding teacher expectations [ 18 ].Therefore, the distinction between the measured content (scaffolding) and the content of theoretical construction (evaluative pressure) is the boundary condition shaped by the high-power distance cultural logic (CHC) and defined by the operationalization of this study, not the measurement error. This article maintains the distinction between the perceived teacher support (PTS) in measurement and the theoretical version of PTS, and its connotation will be discussed in Section 4.3 . Accordingly, the PTS-ELA relationship should be interpreted as a statistically observed pattern — the evaluation pressure mechanism behind it is inferred, not directly verified. 2.2.2 English Language Anxiety (ELA) This study uses 6 items from the Foreign Language Classroom Anxiety Scale (FLCAS[ 7 ]). The revised scale shows good internal consistency in STEM samples. 2.2.3 L2 Willingness to Communicate (L2 WTC) This study adopts the 5-item scale revised by MacIntyre et al. [ 1 ]. The scale meets the standards of reliability and construct validity in the target sample. 2.2.4 Perceived AI Partner Support (AIPS) Based on the social support framework of House [ 26 ] and the digital learning model of Wei and Xu [ 4 ], this study compiles a new 7-item scale. The items aim to capture the characteristics of quasi-social interaction and non-evaluative support of generative artificial intelligence, which is specifically reflected in the tolerance of language errors and no negative evaluation. The research deliberately adopts anthropomorphic expressions, such as "comforting me like a partner". The scale measures the functional emotional results brought about by AI interaction, not learners' ontological beliefs about AI. The design of this study constitutes an activation mechanism based on functional anthropomorphism theory [ 37 ] [ 38 ]. Discriminant validity evidence (HTMT = .711 < .85) confirms that participants did not confuse AI partner support (AIPS) with perceived teacher support (PTS) at the measurement level. Limitations of the operational definition are discussed in Section 4.3 . A pre-experiment with 56 participants verified the clarity of the items. In the formal survey, exploratory factor analysis (EFA) revealed a KMO value of .898, and the Bartlett test of sphericity was significant (χ² = 919.762, df = 21, p < .001). All items loaded on a single common factor, confirming unidimensionality. The Cronbach's α coefficient of the final 7-item scale is .899. To test the cross-sample stability of the unidimensional AIPS structure, this study implements split-half validation. The total sample ( N = 249) is randomly divided into two subsamples using a uniform random number generator. Sub-sample 1 ( n = 128) is used for exploratory factor analysis. A single factor is extracted, with eigenvalue 4.17, explaining 59.57% of total variance. All item loadings are between .732–.803. Sub-sample 2 ( n = 121) is used for confirmatory factor analysis (CFA) of the unidimensional model. Model fit is acceptable: CFI = .954, TLI = .931, RMR = .039, SRMR = .044. RMSEA (.118) exceeds the conventional threshold, which is consistent with the known characteristic that this index is sensitive to small samples[ 39 ]. Overall, EFA and CFA convergent evidence from independent subsamples support the replicability of the unidimensional AIPS structure. To test potential multidimensional structure, this study tests the competing model — a first-order three-factor CFA model. The model shows a Heywood case (inter-factor correlation r = 1.075), indicating the model is empirically unidentified under current sample constraints. Heywood cases in CFA are usually caused by: small sample size relative to model complexity, the construct being unidimensional, or item redundancy[ 40 ][ 41 ]. Since the competing model only estimates inter-factor correlations across seven items in a sample of 249, empirical underidentification is the most parsimonious interpretation. The emergence of a Heywood case therefore provides indirect evidence consistent with unidimensionality rather than against it. The unidimensional structure was retained as the only identifiable solution, and the pragmatic nature of this decision is addressed in Section 4.3 . 2.3 Data Collection Procedure A pilot study with 56 participants was conducted on November 6, 2025 to test the reliability and validity of the scale. A formal online questionnaire survey was then administered. All participants provided implicit informed consent before participation. To ensure ecological validity, the study only included learners who regularly used generative AI tools to practice English. These tools included text-generating platforms (e.g., ChatGPT, ERNIE Bot) and voice-interaction programs (e.g., Doubao, Duolingo). After data collection, responses were screened and excluded if participants reported no AI usage, submitted incomplete questionnaires, or completed the survey in an unreasonably short time. Finally, 249 valid cases were obtained. 2.4 Statistical Analysis All analyses were conducted using IBM SPSS Statistics 27.0 and AMOS 29.0. The study first computed descriptive statistics, Pearson correlations, and Harman's single-factor test to evaluate common method bias.Then, we conducted CFA to test the validity and reliability of the measurement model.Hierarchical regression was employed to examine the study hypotheses. All variance inflation factor (VIF) values were below 2.0, confirming the independence assumption.We adopted bootstrapping with 5,000 resamples to obtain robust 95% bias-corrected confidence intervals. An exploratory moderation analysis was additionally conducted using PROCESS macro version 5.0 [ 42 ] with 5,000 bootstrap resamples to examine whether AIPS moderated the ELA→L2 WTC relationship. To examine potential sample heterogeneity, this study performed multi-group analysis (MGA) in Mplus 8.11 using the Maximum Likelihood with Robust Standard Errors (MLR) estimator. Structural paths were freely estimated across English majors and STEM learners, and the Wald test of parameter constraints was used to exploratorily examine the hypothesized emotion-behavior decoupling in the STEM group. 3. Results 3.1 Data Suitability and Common Method Bias Test KMO values (.851–.910) and Bartlett's test of sphericity ( p < .001) indicate that the data are suitable for factor analysis[ 43 ].We further checked skewness (≤ .83) and kurtosis (≤ .156) to verify normality, which provides justification for subsequent parametric tests [ 35 ]. During the questionnaire design stage, we adopted several procedural controls to reduce common method bias (CMB).These steps included guaranteeing participant anonymity, counterbalancing item blocks, and spatially separating support items from outcome scales [ 44 ]. Harman's single-factor test shows that the first unrotated factor accounts for 36.81% of the total variance, below the 50% threshold. However, the test has well-documented limitations, so the result has limited diagnostic value. This study attempted to model an Unmeasured Latent Method Construct (ULMC) through confirmatory factor analysis (CFA) but encountered a Heywood case, making it impossible to statistically quantify method variance. CMB therefore remains an unresolved threat to internal validity. One structural feature of the data bears on this concern. The bivariate correlations of PTS and AIPS with ELA were similar in magnitude (PTS: r = .329; AIPS: r = .287). If CMB inflation were uniform across both pathways, the regression model would be expected to treat them comparably. It did not: PTS→ELA reached significance (β = .246, p = .001) while AIPS→ELA did not (β = .141, p = .064). A uniform inflation account has difficulty explaining this divergence, which lends indirect support to the substantive interpretation of the PTS→ELA path. That said, without a valid ULMC estimate, residual method variance cannot be definitively quantified, and all path coefficients should be interpreted with this caveat. 3.2 Measurement Model Assessment Following the item purification procedures outlined in Section 2.2 (removing PTS D1–D3, AIPS A1, and L2 WTC C1), the measurement model achieved acceptable fit (Table 1 ). The GFI (.883) is considered adequate for complex models [ 45 ][ 46 ]. Table 1 Confirmatory Factor Analysis (CFA) Model Fit Indices Fit Indices Study Result Academic Threshold Interpretation CMIN/DF 2.057 < 3 Good RMSEA .065 < .08 Acceptable RMR .045 .8 (Acceptable) Acceptable AGFI .851 > .8 Acceptable adjusted fit CFI .929 > .9 Good TLI .918 > .9 Good PNFI .753 > .5 Good PCFI .802 > .5 Good Note . N = 249. Thresholds follow [ 36 ] . 3.2.1 Reliability and Convergent Validity Internal consistency and convergent validity were strong (Table 2 ). All Cronbach's α and Composite Reliability (CR) values were above .80, while Average Variance Extracted (AVE) scores exceeded the .50 threshold. Additionally, all factor loadings were statistically significant ( p < .001). Table 2 . Reliability and Convergent Validity Results (Cronbach's α + CR + AVE) Scale Items Cronbach's α CR AVE AIPS (A2-A8) 7 .899 .897 .554 ELA (B1-B6) 6 .864 .865 .517 L2 WTC (C2-C5) 4 .826 .828 .548 PTS (D4-D6) 3 .820 .833 .624 Note . N = 249. Cr = Composite Reliability; AVE = Average Variance Extracted. 3.2.2 Discriminant Validity Discriminant validity was assessed using the Fornell-Larcker criterion (Table 3 ). The square roots of the AVE values for each construct, shown on the diagonal, were greater than the correlations between the constructs, indicating that all constructs were statistically distinct. Crucially, the Fornell-Larcker criterion results provide robust statistical evidence against potential measurement confounding regarding the AIPS scale. Despite the functional anthropomorphism deliberately utilized in the AIPS items (e.g., 'comforts me'), the square roots of the AVEs for both PTS (.790) and AIPS (.744) strictly exceeded their inter-construct correlation (r = .615). We computed the Heterotrait-Monotrait Ratio (HTMT) to better establish discriminant validity and formally evaluate the presence of common method bias (CMB).Between PTS and AIPS, the largest HTMT value reached .711, well beneath the .85 cutoff widely adopted in structural equation modeling research [ 35 ] [ 47 ].These results offer additional empirical support for the distinctiveness of all focal constructs at the measurement stage. Table 3 Discriminant Validity Test (Square Root of AVE vs. Correlation Coefficients) Variables PTS L2 WTC ELA AI PS PTS (.790) L2 WTC .494** (.740) ELA .329** .298** (.719) AIPS .615** .501** .287** (.744) Note . N = 249. Diagonal values in bold and parentheses are the square roots of the AVE; off-diagonal values are Pearson correlations. ** p < .01. 3.3 Descriptive Statistics and Correlation Analysis Descriptive statistics and Pearson correlations for all focal variables are displayed in Table 4 .These correlational results offer initial empirical support for the proposed research model. Both PTS and AIPS showed significant positive associations with L2 WTC ( p < .01). Of particular interest, ELA also yielded a significant positive correlation with L2 WTC ( r = .298, p < .01), a pattern whose theoretical implications are elaborated in the note below. Theoretical Anomaly Note. Classical SLA theories yield a clear directional expectation for the current sample: with a mean ELA score of 3.68 and widespread self-reported communication apprehension, ELA should relate negatively and significantly to L2 WTC[ 1 ][ 7 ]. However, the observations deviate from this prediction at two distinct levels. At the bivariate level, there is a significant positive correlation between ELA and L2 WTC ( r = .298, p < .01), which is directly contrary to the prediction of classical theory. After controlling the double support source in the multivariate model, ELA does not have a significant inhibitory effect on L2 WTC (β = .105, p = .063), and the coefficient is still positive. Overall, these findings reflect a structural anomaly that cannot be fully explained by insufficient statistical testing: under the framework of traditional second language acquisition (SLA) theory, even if social support is controlled, samples with this average level of anxiety should still present a negative ELA→WTC path. Therefore, the consistent positive correlation across the analytical level needs to go beyond the simple interpretation of "no significant effect" and carry out in-depth theoretical discussion, which also provides a basis for the analysis of Section 4.1.2 of this study. Table 4 Correlation Matrix of Key Constructs Variable M SD 1 2 3 4 1. PTS 3.99 .77 1 - - - 2. L2 WTC 3.62 .84 .494** 1 - - 3. ELA 3.68 .81 .329** .298** 1 - 4. AIPS 3.89 .75 .615** .501** .287** 1 Note. N = 249. ** p < .01 (two-tailed). 3.4 Hypothesis Testing: Hierarchical Regression Results Hierarchical regression analyses were conducted to sequentially test the proposed hypotheses. The results predicting L2 WTC are presented in Table 5 . Table 5 Hierarchical Regression Analysis Results Variables Model 1(Controls ) Model 2(Dual Support) Model 3(Anxiety Inclusion) Model 4 (Predicting ELA) Dependent Variable L2 WTC L2 WTC L2 WTC ELA Gender β = .032, p=.636 β = − .011 , p=.846 β= − .018, p =.752 β = .064, p =.312 Major β = − .109, p=.105 β= − .164**, p =.003 β= − .154**, p =.006 β= − .098, p = .120 PTS (Centered) — β = .311***, p <.001 β = .285***, p <.001 β = .246***, p = .001 AIPS(Centered) — β = .319***, p <.001 β = .304***, p <.001 β = .141, p = .064 ELA (Anxiety) — — β = .105, p=.063 — R² .015 .332 .342 .137 ΔR² .015 .317 .009 .137 F 1.882 57.923*** 3.497 9.712*** Note. N = 249. Standardized coefficients ( β ) are reported. * p < .05, ** p < .01, *** p < .001. 3.4.1 Regression Model Interpretations Step 1: Robustness Check with Control Variables This study incorporates demographic variables (gender and major) to test whether there is a systematic deviation in a male-dominated STEM sample. The explanatory power of the benchmark model is extremely low ( R² = .015), and gender is not a significant predictor. Step 2: Dual-Path Support Validation After adding PTS and AIPS, the model’s explained variance increased significantly. Both predictors reached significance (β = .311, p < .001 for PTS; β = .319, p < .001 for AIPS), supporting H1. Learners maintained communicative intent through both human and AI support. Step 3: Differentiated effect on anxiety (test H2) Model 4 is tested with ELA as the dependent variable.PTS has a significant positive correlation with anxiety (β = .246, p = .001), which is consistent with the inferential operation definition described in Section 2.2.1 . This direction is in line with theoretical expectations, but in view of the cross-sectional design, the reverse interpretation is still valid: anxious learners may actively seek more teacher support as a compensatory response. In contrast, AIPS has no significant effect on anxiety (β = .141, p = .064). This model supports hypothesis H2 in direction: Authority-based support will aggravate evaluation anxiety, while AI support will not. In order to formally test whether there is a significant difference between the two coefficients, this study uses the non-standardized regression coefficient to conduct the Paternoster z test[ 48 ]. The test result is not significant ( z = 0.943, p = .346), indicating that under the current sample size, the PTS→ELA path ( B = .260, SE = .080) and the AIPS→ELA path ( B = .152, SE = .082) is within the sampling error range. The result recalibrates the interpretation of H2: this study data does not support the formal distinction between the two effects, but supports the directional model consistent with H2. The positive correlation strength of PTS and ELA is higher than that of AIPS, but to achieve the distinction at the coefficient level, repeated verification with a larger sample is needed (recommended sample size N > 500). Nevertheless, this directional difference is consistent with the theoretical distinction between evaluative support and non-evaluative support, and the structural argument of the common method deviation presented in Section 3.1 further supports its substantive interpretation. Step 4: The Non-Inhibitory Pattern of ELA (Model 3) Model 3 incorporates ELA with the dual support variables to predict L2 WTC. ELA did not emerge as a significant negative predictor (β = .105, p = .063). The positive direction of the coefficient warrants an alternative interpretation: the positive bivariate correlation between ELA and WTC ( r = .298) may reflect shared variance from an unmeasured third variable (most plausibly academic survival motivation). After controlling for PTS and AIPS, the path became non-significantly positive, which to some extent reflects a suppression-like pattern in the dual-support ecology. Existing cross-sectional data cannot rule out this explanation. To empirically arbitrate among these competing accounts, an exploratory moderation analysis was conducted and is reported in Section 3.5 . To examine whether the non-significant result reflects a genuine absence of inhibitory effects rather than insufficient statistical power, this study adopted the Two One-Sided Tests (TOST) equivalence procedure [ 49 ]. The equivalence bound was set at Δβ = ±0.20, based on the smallest effect size of practical significance for instructional design in applied linguistics [ 50 ]. Effects below this threshold are practically meaningless for teaching design.The TOST result was significant at this bound ( p = .026). A stricter bound of ± 0.10 was also tested; equivalence was not established ( p = .210), indicating that under a more conservative standard, the effect cannot be judged as negligible.Therefore, the main equivalence conclusion holds at ± 0.20 and is explicitly boundary-dependent. 3.5 Exploratory Moderation Analysis In order to provide an empirical basis for the preliminary distinction of the three competitive interpretations proposed in Section 4.1.2 , this study uses PROCESS Model 1 [ 42 ] for exploratory moderation effect analysis and sets up 5,000 Bootstrap resampling. And centralize the processing of the predicted variables. Take English Language Anxiety (ELA) as the predictive variable, L2 Willingness to Communicate (L2 WTC) as the result variable, and AI Partner Support (AIPS) as the moderator. The interaction term ELA × AIPS is not significant ( B = − .007, SE = .005, p = .171, 95% CI [− .020, .007]), indicating that AIPS does not moderate the relationship between ELA and L2 WTC. The overall model accounted for 28.25% of variance in L2 WTC (R ² = .283, F (3, 245) = 32.15, p < .001). Theoretical interpretation of this result is provided in Section 4.1.2 . 3.6 Multi-Group Analysis: An Exploratory Cross-Discipline Signal This study uses the MLR estimation method to conduct an exploratory multi-group analysis in Mplus 8.11. Full metric invariance between groups is supported (ΔCFI = .002) [ 51 ], indicating that the measurement framework is functionally equivalent in two subsamples. Although the sub-sample size is relatively small, the baseline configural model still shows an acceptable fit (CFI = .865, RMSEA = .070).The interdisciplinary structure path shows a clear direction pattern:For English majors, anxiety has a significant positive predictive effect on willingness to communicate (WTC) (β = .314, p = .025); There is no significant predictive effect on STEM students (β = .029, p = .742). The bootstrapped contrast did not reach the conventional significant level (Δβ = .285, p = .088). The post-hoc power analysis showed that the statistical power was 62%, which was lower than the generally accepted 80% standard. Therefore, the differences between the groups should be regarded as a suggestive trend rather than a solid empirical result that can be strongly inferred. Discipline-based comparative research usually requires 150–200 subjects per group to obtain sufficient statistical power to make effective cross-group inference. 4. Discussion 4.1 Findings We examined how PTS and AIPS jointly shape L2 WTC and ELA in Chinese STEM learners.Our findings offer robust support for the additive positive effect of the dual-support system (H1). For Hypothesis 2, the Paternoster z-test showed that the two coefficients were not statistically different under the current sample size ( z = 0.943, p = .346). Nevertheless, the observed directional pattern was consistent with H2: the positive link between PTS and ELA (β = .246, p = .001) was stronger than the corresponding association for AIPS (β = .141, p = .064). This pattern suggests that PTS and AIPS differ in strength rather than nature in their relations with ELA: both convey evaluative signals, yet PTS conveys a markedly stronger signal.The formal statistical separation of the two paths is regarded as a target for subsequent replication, rather than a confirmed research finding. A pattern consistent with H3 also emerged: ELA did not inhibit L2 WTC, and its positive coefficient across both bivariate and multivariate levels constitutes a dual-level anomaly elaborated in Section 4.1.2 . All paths involving PTS and ELA represent culturally inferred associations, bounded by the CHC heuristic framework established in Section 2.2.1 , and the cross-sectional design does not permit definitive causal inferences. 4.1.1 The Synergistic Drivers: Dual-Path Support on WTC Both PTS and AIPS have become powerful parallel predictors of L2 WTC, which is consistent with H1. These findings extend the classic social support theories [ 26 ][ 34 ] to the AI-assisted learning context. The correlation coefficient between PTS and WTC ( r = .494) shows that even if there is evaluative pressure, authority-based teacher support still retains its scaffolding function in Confucian culture. At the same time, AIPS demonstrates independent predictive power—which suggests that generative AI may not only be used as a practice tool, but also as a structurally distinct support source that works in parallel with human scaffolding. The two forms of support appear to work in tandem rather than overlap: students receive instructional support from human instructors while drawing affective reassurance from AI. 4.1.2 A Directional Anomaly in the Classic Affective Filter: Boundary Conditions and Theoretical Implications The dual-level oddity in Section 3.3 requires comprehensive theoretical consideration.Traditional SLA frameworks [ 1 ][ 7 ] indicate a significant negative ELA→WTC route in our sample, with a mean ELA of 3.68.ELA positively correlates with WTC at the bivariate level ( r = .298, p < .01) but does not inhibit WTC at the multivariate level (β = .105, p = .063).Insufficient statistical power is hard to explain when both pieces of data point in the same unusual direction.Cross-sectional data cannot discriminate between three plausible competing theories. The first is a suppressor explanation. An unmeasured third variable, most likely academic survival motivation driven by the communicative imperative [ 9 ], simultaneously increases both ELA and WTC. English is high-stakes and thus anxiety-inducing for ELA, and high-stakes and thus worth attempting for WTC. This pattern produces the observed positive bivariate correlation while suppressing the inhibitory path in regression. This account requires no new theoretical assumptions and only invokes an unmeasured variable.This is consistent with the communication necessity framework in Section 1.1 : Academic survival motivation drives both English Language Anxiety (ELA) and Willingness to Communicate (WTC). The second explanation is the facilitative anxiety perspective. Under task pressure, STEM learners' ELA may manifest as facilitative rather than debilitating anxiety [ 19 ][ 20 ], that is, activating cognitive resources rather than triggering avoidance behavior. The FLCAS revised scale adopted in this study does not distinguish between facilitative debilitating anxiety subtypes, so this possibility cannot be ruled out. The third explanation is the functional substitution perspective. Learners can hand over the highly anxious trial and error process to non-evaluative AI, so as to structurally remove the connection between ELA and communication avoidance. The view is that with the improvement of AIPS, ELA's inhibition path to WTC will weaken, which is different from the first two explanations. In order to test whether functional substitution can explain the non-suppressive mode of ELA, this study uses PROCESS Model 1[ 42 ] for exploratory moderation analysis, set up 5,000 Bootstrap resampling and centralize the prediction variables. The interaction term ELA × AIPS is not significant ( B = − .007, SE = .005, p = .171, 95% CI [− .020, .007]), indicating that AIPS does not moderate the relationship between ELA and L2 WTC. This null interaction result suggests that the non-inhibitory ELA pattern is unlikely to be driven by functional substitution alone, lending indirect support to the suppressor variable interpretation as the more parsimonious account under the present data. Pragmatic Resilience serves as a conceptual label to guide systematic investigation of this anomaly.The three mechanisms identified are not mutually exclusive: academic survival motivation may establish the baseline, facilitative anxiety may shape task engagement, and AI-enabled functional substitution may act as a secondary mechanism.Distinguishing among these accounts requires longitudinal designs, experimental manipulation of AI use, and direct measures of academic survival motivation. 4.1.3 AI's Nonjudgmental Role: A Safety Valve for Learner Anxiety PTS aids communication, however statistics show its dual nature in CHC situations.Students often regard the intensive support of teachers as the embodiment of academic care and strong authority[ 27 ]. The mechanism of AIPS is completely different: non-evaluative gaze [ 30 ] creates a unique emotional space that has a non-significant impact on anxiety (β = .141, p = .064). This contradiction is not a simple opposition. This study calls it "ontological dualism". Learners can get psychological comfort in the anthropomorphic language of AI without worrying about evaluation. Based on the recent expansion of media equation theory [ 37 ] and the mind perception framework [ 38 ], this study believes that users can anthropomorphize the intelligent body for social and emotional benefits, while maintaining non-human ontology awareness at the metacognitive level. STEM students may adopt the attitude of "willing suspension of disbelief" in this digital learning ecology: using AI's functional anthropomorphism to meet emotional needs, while avoiding the risk of "face" in reality with the help of its machine ontological attributes. This makes AIPS a unique psychological affordance, rather than a simple reproduction of human interaction. 4.1.4 Alternative Explanations: The Bidirectionality of Support and Anxiety The cross-sectional design limits the causal interpretation of the PTS–ELA path.In the Chinese cultural context (CHC), evaluative support may exacerbate anxiety due to “face” concerns [ 27 ];but highly anxious learners may also actively seek more teacher support as a compensatory response driven by communicative imperative.Longitudinal data are needed to clarify these causal directions. 4.2 Practical Implications The structural model supports testable teaching heuristic strategies, such as the Strategic Sandwich model (Human-AI-Human). This paper proposes a design-based research agenda to carry out experimental verification instead of constructing a deterministic causal model (Fig. 2 ). Based on control-value theory [ 20 ], the process will hand over anxiety-inducing practice to AI, while retaining the motivational authority of teachers. Stage 1: Teachers lead input and motivation. Teachers set academic goals and emphasize the value of tasks; evaluative stakes and communicative imperative are established at this stage. Stage 2: AI-assisted buffer. Students use generative AI for low-risk private exercises. The non-evaluative environment promotes trial and error and no "face" cost, so that learners can gain a sense of control before high-stakes performance and improve linguistic accuracy. Stage 3: Output and Performance (Interpersonal Interaction). Classroom display and peer communication enable students to return to the interpersonal environment. Evaluative pressure still exists, but the early AI practice reduces the cognitive load and avoids anxiety-induced avoidance behavior. Anxiety is addressed rather than suppressed. The model does not advocate that AI replaces the affective needs of interpersonal interaction, but proposes that putting AI exercises before interpersonal performance can reduce the threshold of behavioral disruption caused by anxiety. This hypothesis can be tested by the experimental design of random assignment to sequencing conditions. Note The "Strategic Sandwich" model integrates the dual-path ecology into a three-stage sequence: teacher-led goal setting (stage 1) to induce evaluation pressure, AI-assisted drill (stage 2) to provide a nonjudgmental buffer, and interpersonal interaction (stage 3) to implement the necessity of communication. AI acts as an emotional scaffold, not a substitute for real social interaction. 4.3 Limitations and Future Research There are significant interpretation limitations in this research, and the most fundamental constraint is cross-sectional design. There may be a bidirectional relationship between ELA and WTC: continuous positive AI interaction may reduce baseline anxiety, but existing research cannot verify this, and it is necessary to use longitudinal design to track such dynamic reciprocal relationships. There are two factors that affect the generalizability: English majors in a single institution may not represent all language majors in each university, and it is necessary to carry out multi-institutional replication to judge whether the discipline-specific effects observed in this study are universal; the sub-sample size of multi-group analysis ( n = 95; n = 154) results in statistical power insufficient (62%), which is lower than the threshold of 80%. In future studies, each subgroup should ensure 150–200 subjects before drawing discipline-specific conclusions. Sampling criteria bring inherent tension: From a methodological point of view, limiting AIPS evaluation to learners with experience in using AI will limit the generalizability to non-AI-user groups and may lead to a ceiling effect on the baseline AIPS score. Future research should include a control group without AI experience to establish the boundary conditions of the support effect. Future research must solve the two major measurement limitations of the AIPS scale: because its anthropomorphic language is not verified in combination with the respondents' explicit AI ontology awareness, some subjects may confuse AI support with social support; Heywood case appears in the competing CFA model, indicating the unidimensional structure is more practical than the theoretically grounded structure. The psychometric properties of the scale must be confirmed by cross-sample replication and MTMM-based convergent validity testing. The conceptual boundary of PTS is deliberate methodological choice, not omission. Items D4–D6 measure teacher care and academic scaffolding; The evaluative pressure at the core of this study's PTS-ELA theory is inferred based on Chinese cultural logic (CHC) — in this culture, authority-based support is intrinsically related to face expectations and social appraisal [ 27 ]. This inferential gap means that the observed PTS-ELA significant correlation (β = .246, p = .001) should be regarded as a statistically observed pattern, and the evaluative pressure mechanism behind it comes from cultural inference rather than direct verification. In order to bridge this conceptual gap, future research should develop and validate a dedicated “Perceived Academic Evaluative Threat” scale, which should be combined with PTS scaffolding items to directly empirically test the evaluative pressure mechanism proposed by this study. The data of this study only partially distinguish the three competing explanations for the ELA non-inhibitory pattern proposed in Section 4.1.2 . Exploratory moderation analysis provides indirect evidence against the "functional substitution as main effect", but limited by cross-sectional design, the suppressor variable account and the facilitative anxiety account cannot be empirically separated. Future research should directly measure academic survival motivation and distinguish these explanations through experimental manipulation of AI availability. Quantitative designs can portray the structural pattern, but they cannot touch learners' lived experiences. This study constitutes the first stage of an explanatory sequential mixed-methods design; the subsequent qualitative stage will use in-depth interviews to analyze how STEM students cope with teacher evaluation pressure and why AI is constructed as an affectively safe space, thus enriching the boundary conditions of this study. 5. Conclusion Under the human-AI dual support context, this study examines how PTS and AIPS jointly affect L2 WTC and ELA among Chinese STEM learners. At the confirmatory analysis level, the two sources of support independently and additively predict L2 WTC (ΔR² = .317). PTS provides academic scaffolding, while AIPS offers a non-judgmental space that does not amplify the inherent evaluative pressure of teacher feedback. This dual-pathway structure constitutes a replicable empirical pattern with direct implications for instructional design. At the exploratory level, ELA did not act as a significant inhibitor of L2 WTC. Compared with classical SLA predictions, this finding reveals a dual-level anomaly. Pragmatic Resilience serves as a conceptual framework to guide future research on this boundary condition, rather than functioning as a settled theoretical explanation. Interpersonal support and AI support can coexist as functionally distinct resources: the former establishes academic standards, while the latter buffers affective costs. Whether such coexistence can structurally reshape the classic anxiety–communication relationship remains an open empirical question. Abbreviations Full term AIPS AI Partner Support CFA Confirmatory Factor Analysis CHC Chinese Heritage Culture EFA Exploratory Factor Analysis EFL English as a Foreign Language ELA English Language Anxiety L2 WTC Second Language Willingness to Communicate MGA Multi-Group Analysis PTS Perceived Teacher Support SLA Second Language Acquisition STEM Science, Technology, Engineering, and Mathematics Declarations Ethics approval and consent to participate This study was approved by the Ethics Review Committee of the University International College (UIC), Macau University of Science and Technology (MUST) (Ref. No. : UIC/L/26/132; approved November 1, 2025), and was conducted in accordance with the Declaration of Helsinki. All participants received a comprehensive introduction prior to participation; completing and submitting the questionnaire was treated as implied informed consent. Participant anonymity and data confidentiality were strictly maintained throughout, and participants retained the right to withdraw at any time without consequence. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Funding No external funding was received for this study. Author Contribution CX conceived and designed the study, collected and analysed the data, and wrote the manuscript. WZ supervised the study and provided critical revisions. All authors read and approved the final manuscript. Acknowledgement The author used ChatGPT (OpenAI) for grammar checking and language editing during manuscript preparation. All content was reviewed and revised by the author, who takes full responsibility for the final manuscript. Data Availability The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request. References MacIntyre PD, Clément R, Dörnyei Z, Noels KA. Conceptualizing willingness to communicate in a L2: A situational model of L2 confidence and affiliation. Mod Lang J. 1998;82(4):545–62. https://doi.org/10.1111/j.1540-4781.1998.tb05543.x . Mystkowska-Wiertelak A, Pawlak M. Willingness to communicate in instructed second language acquisition: Combining a macro- and micro-perspective. Multilingual Matters; 2017. Yashima T, MacIntyre PD, Ikeda M. Situated willingness to communicate in an L2: Interplay of individual characteristics and context. Lang Teach Res. 2018;22(1):115–37. https://doi.org/10.1177/1362168816657851 . Wei X, Xu Q. Predictors of willingness to communicate in a second language (L2 WTC): Toward an integrated L2 WTC model from the socio-psychological perspective. Foreign Lang Ann. 2022;55(1):258–82. https://doi.org/10.1111/flan.12595 . Yu Q, Xu D, Wan Q. Exploring the relationship between L2 motivational self system, academic emotions, and L2 willingness to communicate among university EFL learners in central and western China. Acta Psychol. 2025;261:105969. https://doi.org/10.1016/j.actpsy.2025.105969 . Katsaris T. The Willingness to Communicate (WTC): Origins, significance, and propositions for the L2/FL classroom. J Appl Lang Linguist . 2019;3(2):31–42. https://als-edu.wixsite.com/jall . Accessed 24 Feb 2026. Horwitz EK, Horwitz MB, Cope J. Foreign language classroom anxiety. Mod Lang J. 1986;70(2):125–32. https://doi.org/10.1111/j.1540-4781.1986.tb05256.x . Dornyei Z. The L2 motivational self system. In: Dornyei Z, Ushioda E, editors. Motivation, language identity and the L2 self. Multilingual Matters; 2009. pp. 9–42. Duff PA. Language socialization into academic discourse communities. Annu Rev Appl Linguist. 2010;30:169–92. https://doi.org/10.1017/S0267190510000048 . Pun J, Fu BX, Cheung KKC. Language challenges and coping strategies in English Medium Instruction (EMI) science classrooms: A critical review of literature. Stud Sci Educ. 2023;60(1):1–32. https://doi.org/10.1080/03057267.2023.2188704 . Xie W, Curle S. Success in English Medium Instruction in China: Significant indicators and implications. Int J Biling Educ Biling. 2022;25(2):585–97. https://doi.org/10.1080/13670050.2019.1703898 . Deci EL, Ryan RM. The What and Why of goal pursuits: Human needs and the self-determination of behavior. Psychol Inq. 2000;11(4):227–68. https://doi.org/10.1207/S15327965PLI1104_01 . Kohnke L, Moorhouse BL, Zou D. ChatGPT for language teaching and learning. RELC J. 2023;54(2):537–53. https://doi.org/10.1177/00336882231162868 . Zhang C, Meng Y, Ma X. Artificial intelligence in EFL speaking: Impact on enjoyment, anxiety, and willingness to communicate. System. 2024;121:103259. https://doi.org/10.1016/j.system.2024.103259 . Kalsoom A, Soomro NH, Pathan ZH. How social support and foreign language anxiety impact willingness to communicate in English in an EFL classroom. Int J Engl Linguist. 2020;10(2):80–91. https://doi.org/10.5539/ijel.v10n2p80 . Ma M, Noordin N, Razali AB. Effects of an AI Chatbot Mobile Application on Foreign Language Anxiety among Chinese EFL Undergraduates. Int J Acad Res Prog Educ Dev . 2024;13(4). https://ijarped.com/index.php/journal/article/view/3301 . Accessed 24 Feb 2026. Teimouri Y, Goetze J, Plonsky L. Second language anxiety and achievement: A meta-analysis. Stud Second Lang Acquis. 2019;41(2):363–87. https://doi.org/10.1017/S0272263118000311 . Dewaele J-M. The effect of classroom emotions, attitudes toward English, and teacher behavior on willingness to communicate among English foreign language learners. J Lang Soc Psychol. 2019;38(4):523–35. https://doi.org/10.1177/0261927X19864996 . Alpert R, Haber RN. Anxiety in academic achievement situations. J Abnorm Soc Psychol. 1960;61(2):207–15. https://doi.org/10.1037/h0045464 . Pekrun R. The control-value theory of achievement emotions: Assumptions, corollaries, and implications for educational research and practice. Educ Psychol Rev. 2006;18(4):315–41. https://doi.org/10.1007/s10648-006-9029-9 . Wu C-P, Lin H-J. Anxiety about speaking a foreign language as a mediator of the relation between motivation and willingness to communicate. Percept Mot Skills. 2014;119(3):785–98. https://doi.org/10.2466/22.PMS.119c32z7 . Almukhaild H, King J. Scaling the apex of engagement: Anxious language learners' positive flow experiences during speaking tasks. System. 2024;124:103373. https://doi.org/10.1016/j.system.2024.103373 . Bonanno GA. Loss, trauma, and human resilience: Have we underestimated the human capacity to thrive after extremely aversive events? Am Psychol. 2004;59(1):20–8. https://doi.org/10.1037/0003-066X.59.1.20 . Luthar SS, Cicchetti D, Becker B. The construct of resilience: A critical evaluation and guidelines for future work. Child Dev. 2000;71(3):543–62. https://doi.org/10.1111/1467-8624.00164 . Zimet GD, Dahlem NW, Zimet SG, Farley GK. The multidimensional scale of perceived social support. J Pers Assess. 1988;52(1):30–41. https://doi.org/10.1207/s15327752jpa5201_2 . House JS. Work stress and social support. Addison-Wesley; 1981. Hwang KK. Face and favor: The Chinese power game. Am J Sociol. 1987;92(4):944–74. https://doi.org/10.1086/228588 . Young DJ. Creating a low-anxiety classroom environment: What does language anxiety research suggest? Mod Lang J. 1991;75(4):426–39. https://doi.org/10.2307/329492 . Noels KA, Pelletier LG, Clement R, Vallerand RJ. Why are you learning a second language? Motivational orientations and self-determination theory. Lang Learn. 2000;50(1):57–85. https://doi.org/10.1111/0023-8333.00111 . Short J, Williams E, Christie B. The social psychology of telecommunications. Wiley; 1976. Abdelsaheed BSM, Alsalamah LA. ChatGPT as a tool for reducing anxiety and enhancing presentation skills: Perspectives of Saudi EFL students. Int J Instr. 2026;19(1):719–48. https://doi.org/10.29333/iji.2026.19137a . Fathi J, Rahimi M, Derakhshan A, Improving. EFL learners' speaking skills and willingness to communicate via artificial intelligence-mediated interactions. System. 2024;121:103254. https://doi.org/10.1016/j.system.2024.103254 . Cohen S, Wills TA. Stress, social support, and the buffering hypothesis. Psychol Bull. 1985;98(2):310–57. https://doi.org/10.1037/0033-2909.98.2.310 . Chen A, Jia J, Li Y, Fu L. Investigating the effect of role-play activity with GenAI agent on EFL students' speaking performance. J Educ Comput Res. 2025;63(1):99–125. https://doi.org/10.1177/07356331241299058 . Kline RB. Principles and practice of structural equation modeling. 5th ed. Guilford; 2023. Hair JF, Black WC, Babin BJ, Anderson RE. Multivariate data analysis. 7th ed. Pearson; 2010. Gambino A, Fox J, Ratan RA. Building a stronger CASA: Extending the computers are social actors paradigm. Hum Mach Commun. 2020;1:71–85. https://doi.org/10.30658/hmc.1.5 . Waytz A, Cacioppo J, Epley N. Who sees human? The stability and importance of individual differences in anthropomorphism. Perspect Psychol Sci. 2010;5(3):219–32. https://doi.org/10.1177/1745691610369336 . Kenny DA, Kaniskan B, McCoach DB. The performance of RMSEA in models with small degrees of freedom. Sociol Methods Res. 2015;44(3):486–507. https://doi.org/10.1177/0049124114543236 . van Driel OP. On various causes of improper solutions in maximum likelihood factor analysis. Psychometrika. 1978;43(2):225–43. https://doi.org/10.1007/BF02293865 . Kolenikov S, Bollen KA. Testing negative error variances: Is a Heywood case a symptom of misspecification? Sociol Methods Res. 2012;41(1):124–67. https://doi.org/10.1177/0049124112442138 . Hayes AF. Introduction to mediation, moderation, and conditional process analysis: A regression-based approach. 3rd ed. Guilford Press; 2022. Kaiser HF, Rice J, Little Jiffy MIV. Educ Psychol Meas. 1974;34(1):111–7. https://doi.org/10.1177/001316447403400115 . Podsakoff PM, MacKenzie SB, Lee JY, Podsakoff NP. Common method biases in behavioral research. J Appl Psychol. 2003;88(5):879–903. https://doi.org/10.1037/0021-9010.88.5.879 . Baumgartner H, Homburg C. Applications of structural equation modeling in marketing and consumer research: A review. Int J Res Mark. 1996;13(2):139–61. https://doi.org/10.1016/0167-8116(95)00038-0 . Doll WJ, Xia W, Torkzadeh G. A confirmatory factor analysis of the end-user computing satisfaction instrument. MIS Q. 1994;18(4):453–61. https://doi.org/10.2307/249524 . Henseler J, Ringle CM, Sarstedt M. A new criterion for assessing discriminant validity in variance-based structural equation modeling. J Acad Mark Sci. 2015;43(1):115–35. https://doi.org/10.1007/s11747-014-0403-8 . Paternoster R, Brame R, Mazerolle P, Piquero A. Using the correct statistical test for the equality of regression coefficients. Criminology. 1998;36(4):859–66. https://doi.org/10.1111/j.1745-9125.1998.tb01268.x . Lakens D, Scheel AM, Isager PM. Equivalence testing for psychological research: A tutorial. Adv Methods Pract Psychol Sci. 2018;1(2):259–69. https://doi.org/10.1177/2515245918770963 . Plonsky L, Oswald FL. How big is big? Interpreting effect sizes in L2 research. Lang Learn. 2014;64(4):878–912. https://doi.org/10.1111/lang.12079 . Cheung GW, Rensvold RB. Evaluating goodness-of-fit indexes for testing measurement invariance. Struct Equ Model. 2002;9(2):233–55. https://doi.org/10.1207/S15328007SEM0902_5 . Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Reviewers invited by journal 06 May, 2026 Editor invited by journal 05 May, 2026 Editor assigned by journal 04 May, 2026 Submission checks completed at journal 04 May, 2026 First submitted to journal 30 Apr, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9578013","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":638633801,"identity":"aec156c4-fa13-4dbe-b1ec-1e023c5f055e","order_by":0,"name":"CHEN XIN","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAu0lEQVRIiWNgGAWjYDACZiBOYLABsRoYGA4QryWNgYGNkVgtEHCYBC0Gx5mfSTzccV7O4H5j44MPZxjk+cUI6DM4zGZskHjmtrHBMcZmwxk3GAxnzk7Ar0WymcHwQWLb7cQNxxjbpHk+MCQY3Caohf3DgcS2cyRo4WfmAdlyAKrlBnFaig0S25KNJY8lAv1yRoKwX9j4j2+T/NlmJ8d3+PDBBx+O2cjzSxPQgg4kSFM+CkbBKBgFowA7AAB1CUN5RjQzWgAAAABJRU5ErkJggg==","orcid":"","institution":"Macau University of Science and Technology","correspondingAuthor":true,"prefix":"","firstName":"CHEN","middleName":"","lastName":"XIN","suffix":""},{"id":638633803,"identity":"4f265499-5d5e-415f-bc33-f7fe2143bcb5","order_by":1,"name":"WEIHE ZHONG","email":"","orcid":"","institution":"Macau Millennium College","correspondingAuthor":false,"prefix":"","firstName":"WEIHE","middleName":"","lastName":"ZHONG","suffix":""}],"badges":[],"createdAt":"2026-04-30 13:53:10","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9578013/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9578013/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":109341534,"identity":"11986a33-90a7-4e4f-ba47-846fb925d324","added_by":"auto","created_at":"2026-05-15 19:03:39","extension":"jpeg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":46503,"visible":true,"origin":"","legend":"\u003cp\u003eTheoretical Framework of Dual-Path Support and Pragmatic Resilience\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eNote.\u003c/em\u003e PTS = Perceived Teacher Support; AIPS = AI Partner Support; ELA = English Language Anxiety; L2 WTC = L2 Willingness to Communicate. Solid lines denote hypothesized direct paths. The model illustrates the dual-pathway additive effect on L2 WTC and the emotion-behavior decoupling between ELA and L2 WTC.\u003c/p\u003e","description":"","filename":"floatimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-9578013/v1/5143d07b2a20204d127d85bc.jpeg"},{"id":109405676,"identity":"ed519bb5-6141-478f-8298-512f4c8f3245","added_by":"auto","created_at":"2026-05-17 13:19:39","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":772772,"visible":true,"origin":"","legend":"\u003cp\u003eThe \"Human-AI-Human\" Symbiotic Sandwich Model\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eNote.\u003c/em\u003e The \"Strategic Sandwich\" model integrates the dual-path ecology into a three-stage sequence: teacher-led goal setting (stage 1) to induce evaluation pressure, AI-assisted drill (stage 2) to provide a nonjudgmental buffer, and interpersonal interaction (stage 3) to implement the necessity of communication. AI acts as an emotional scaffold, not a substitute for real social interaction.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-9578013/v1/30a98d7780659212ecb27d34.png"},{"id":109406121,"identity":"c991c24b-4ce9-4ce9-9fbe-98079e502b9b","added_by":"auto","created_at":"2026-05-17 13:25:01","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1328948,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9578013/v1/37d6e555-898f-4ab3-9b28-e29572f09188.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Anxious but Active: The Additive Effects of Generative AI and Teacher Support on L2 Willingness to Communicate","fulltext":[{"header":"1. Background","content":"\u003cdiv id=\"Sec2\" class=\"Section2\"\u003e \u003ch2\u003e1.1 L2 Willingness to Communicate in STEM Contexts\u003c/h2\u003e \u003cp\u003eL2 WTC is not a fixed personality trait. MacIntyre et al. [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e] define it as a learner being ready to use L2 to communicate with a specific person at a specific moment. This is a state-level construct shaped by situation and context [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e][\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. As an immediate psychological link before behavior execution, it is a direct predictor of L2 oral communication, playing a mediating role between learners' emotional states and language expression activities [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe heuristic pyramid model by MacIntyre et al.[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e] portrays this dynamic fluctuation. Subsequent empirical studies have clarified its key influencing factors. Anxiety weakens the positive effect of motivation [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Social support and anxiety relief are regarded as important promoters of WTC [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. What has not been fully explored is how these relationships will change as the digital learning ecology reshapes the communication environment. This question has not been answered by the core predictive variables identified by Wei and Xu [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe classical second language acquisition (SLA) theory regards ELA as the primary affective barrier to L2 WTC[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e][\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. However, this linear hypothesis is difficult to account for the actual situation in the STEM academic context. STEM students are largely driven by ought-to L2 self [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e] and face the rigid need to communicate for academic survival - including subject defense, transnational laboratory cooperation and technical literature reading [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. This kind of communication pressure still exists structurally in the contemporary STEM environment: empirical evidence from Chinese EMI settings shows that STEM students continue to face language-related challenges, including oral report anxiety and insufficient English proficiency when dealing with high-risk academic tasks [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e][\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. This necessity existed before the AI era, but its core mechanism is still effective: STEM students must maintain L2 output even under strong ELA. L2 WTC is a direct psychological precursor of communication behavior [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Therefore, learners can still maintain communication output under sustained anxiety, precisely because they retain a correspondingly high willingness to communicate. This paradox manifests not only at the level of external behavior but also at the level of communication intention. According to Self-Determination Theory (SDT) [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e], this pattern aligns with identified regulation: learners internalize the instrumental value of English as academic survival needs and sustain communication output under persistent emotional pressure.\u003c/p\u003e \u003cp\u003eIn STEM situations, L2 WTC is driven by real necessity rather than intrinsic pleasure; this instrumental identification adjustment overwhelms the classic inhibition effect of anxiety.\u003c/p\u003e \u003cp\u003eWith the rapid development of generative AI, conversational AI now provides a non-evaluative environment for oral practice [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e][\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. However, most of the existing literature examines traditional interpersonal support and AI-driven support in isolation (such as [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e][\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]). In high-risk STEM situations, it is not clear how humans and AI systems complement each other to maintain communication. This study responds to this research gap through the Human-AI Symbiosis framework and explores how PTS and AIPS work together under high-intensity academic pressure.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e1.2 English Language Anxiety and Emotion-Behavior Decoupling\u003c/h2\u003e \u003cp\u003eELA is a situation-specific construct, which is different from general trait anxiety [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. Teimouri et al. [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e] found that there is a significant negative correlation between anxiety and learning achievement; Dewaele [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e] found that high ELA is associated with L2 WTC reduction.These findings are the current mainstream views. However, Alpert and Haber [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e] have long pointed out that anxiety is not always an obstacle. Its facilitative aspect can maintain rather than inhibit performance. Therefore, in a task-oriented STEM context, moderate ELA may play a role in this way [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. Specifically, under the well-designed task situation, anxious learners sometimes enter a positive flow state [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe standard interpretation regards ELA as an affective filter variable between environmental support and L2 WTC [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. STEM groups may not follow this pattern. Communicative imperative [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e] ignores the emotional state and puts forward continuous second language output requirements for learners, which may make them enter the \"emotional-behavioral separation\" mode. This is different from the situation described in the classical theory of resilience [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e][\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. Learners do not internally regulate negative emotions, but bypass emotional distress from the structural level in the digital ecological environment and transfer the emotional burden to the external scaffolding instead of dealing with it from the inside.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e1.3 The Dual-Support Ecology\u003c/h2\u003e \u003cdiv id=\"Sec5\" class=\"Section3\"\u003e \u003ch2\u003e1.3.1 The Double-Edged Sword of Perceived Teacher Support (PTS)\u003c/h2\u003e \u003cp\u003ePerceived Social Support (PSS) reflects learners' subjective assessment of available support [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. This study converts the authority-based dimensional operation in PSS into Perceived Teacher Support (PTS).\u003c/p\u003e \u003cp\u003eHouse [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e] divides it into four functional forms: informational, instrumental, emotional and appraisal. These forms of support are distributed from different sources such as peers, families and teachers. This study does not examine all functional forms, but adopts a source-based conceptualization method to separate the evaluation dynamic mechanism unique to authority relations.\u003c/p\u003e \u003cp\u003eIts theoretical basis is that in the context of Chinese culture (CHC), the support provided by authoritative figures is essentially related to the expectation of \"face\" and social evaluation [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. This dynamic mechanism is determined by the source of support rather than its functional form, and it cannot be effectively portrayed by functional classification alone. Therefore, this study operationalizes the interpersonal-oriented PSS as PTS, focusing on the support of authoritative sources (i.e. teachers and institutional support).\u003c/p\u003e \u003cp\u003ePTS is a recognized positive predictive factor of L2 WTC [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e], but its vertical structure implies social evaluation attributes [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]. Under the background of CHC, this has formed a double-edged dynamic: the same support provides academic support while conveying evaluation expectations, activating \"face\" concerns and rigidity of standards [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. The resulting \"forced compliance\" [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e] maintains communication behavior through external regulation at the cost of an increase in ELA. This association is not purely theoretically derived. In the situation of CHC and Chinese English as a foreign language (EFL), teachers' strictness and authoritative evaluation have been proven to be a direct trigger for anxiety[\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e], and students report a stronger tendency to remain silent under strict teacher expectations. The expected PTS-ELA correlation in this study is based on this empirical pattern.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section3\"\u003e \u003ch2\u003e1.3.2 AI Partner Support (AIPS) as a Parallel Safety Valve\u003c/h2\u003e \u003cp\u003eGenerative artificial intelligence (GenAI) (such as ChatGPT and Doubao) provides a unique form of support, which is reflected in AIPS. There is no \"social self\" in AIPS[\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e], which is baseline-agnostic and free of evaluative gaze. This non-evaluative and high error tolerance environment has been proven to significantly reduce ELA levels [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e][\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e], including in the context of high-risk tasks [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e]. Artificial intelligence-assisted interaction also yields linguistic gains and motivational effects comparable to face-to-face peer learning [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e][\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eAIPS does not change the evaluative nature of teacher support. On the contrary, it provides an independent space free from fear of negative evaluation, and becomes an emotional safety valve for students under the pressure of academic anxiety [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e] [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]. This makes AIPS an independent parallel path rather than a moderating variable [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]: it works in parallel with interpersonal support without changing its nature, while absorbing the pressure generated by authoritative channels.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e1.4 Research Gaps and Hypotheses\u003c/h2\u003e \u003cp\u003eThis study stems from two theoretical gaps. One involves the behavioral driving mechanism: the SLA research paradigm regards ELA as a behavioral inhibitor, but whether the STEM group operating in the digital ecology really shows the phenomenon of \"emotion-behavior decoupling\" has not been directly tested. The second involves dual support dynamics: the existing model's evaluation of PTS and AIPS is independent of each other. How authority-based PTS and non-evaluative AIPS interact in the same ecology is still an open question. It is not clear whether the two will compound or cancel each other's emotional effects.\u003c/p\u003e \u003cp\u003eBased on the human-AI Symbiosis framework and the psychological characteristics of STEM learners in the CHC, this study puts forward the following hypotheses:\u003c/p\u003e \u003cp\u003eH1 (corresponding research question RQ1): PTS and AIPS both have an additive effect, and significantly positively predict the learners' L2 WTC.\u003c/p\u003e \u003cp\u003eH2 (corresponding to the research question RQ2): There is a differentiated predictive relationship between the two support systems and ELA. Due to the implicit evaluation pressure, PTS predicts ELA positively; in contrast, AIPS is not a significant source of pressure (that is, after controlling for PTS, it has no significant positive prediction effect on ELA) and plays the function of a non-evaluative safety valve.\u003c/p\u003e \u003cp\u003eH3 (corresponding research question RQ3): Under high-intensity academic pressure, it shows the characteristics of \"emotion-behavior decoupling\", and ELA has no significant negative predictive effect on the L2 WTC of STEM learners.\u003c/p\u003e \u003cp\u003eThis study examines the dynamics of human-AI collaboration through an integrated analytical framework of PTS, AIPS, ELA and L2 WTC. Three research questions are addressed:\u003c/p\u003e \u003cp\u003eRQ1: How do PTS and AIPS jointly shape the L2 WTC of STEM learners?\u003c/p\u003e \u003cp\u003eRQ2: How do PTS and AIPS differentially influence learners' ELA?\u003c/p\u003e \u003cp\u003eRQ3: Is ELA a significant behavioral inhibitor of L2 WTC, or will emotion-behavior decoupling emerge among STEM learners under intense academic pressure?\u003c/p\u003e \u003cp\u003eThis study contributes to SLA theory and teaching practice from three levels.\u003c/p\u003e \u003cp\u003eFirst, about the dual-support learning ecology. This study does not regard AI as a traditional \"pressure buffer\". Instead, it examines how AI, as an independent and nonjudgmental affective safety valve, provides complementary affective scaffolding to offset the evaluative pressure inherent in authority-based support.\u003c/p\u003e \u003cp\u003eSecond, about the affective filter mechanism. In this dual-support ecology, ELA does not constitute a significant behavioral inhibitor of L2 WTC, which is inconsistent with the prediction of classical SLA theory. We propose Pragmatic Resilience as a conceptual label to guide subsequent research. Its association with identified regulation will be explained in the discussion section[\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e][\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThird, about instructional design. By building a Strategic Sandwich (Human-AI-Human) heuristic model, this study provides a testable teaching process for educators and subsequent researchers, and hands over the anxiety-inducing trial-and-error process to AI, in order to optimize learners' communicative performance in high-stakes contexts.\u003c/p\u003e \u003cp\u003eTo enhance the rigor of the research and explore potential boundary conditions, this study adds an exploratory step.As an exploratory step, this study uses MGA to test the potential heterogeneity between different disciplines. The cross-group results obtained are only used to generate hypotheses and do not constitute confirmatory evidence of disciplinary differences.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eNote\u003c/strong\u003e \u003cp\u003ePTS\u0026thinsp;=\u0026thinsp;Perceived Teacher Support; AIPS\u0026thinsp;=\u0026thinsp;AI Partner Support; ELA\u0026thinsp;=\u0026thinsp;English Language Anxiety; L2 WTC\u0026thinsp;=\u0026thinsp;L2 Willingness to Communicate. Solid lines denote hypothesized direct paths. The model illustrates the dual-pathway additive effect on L2 WTC and the emotion-behavior decoupling between ELA and L2 WTC.\u003c/p\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"2. Methods","content":"\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Participants\u003c/h2\u003e \u003cp\u003eThis study adopts convenience sampling to select English as a Foreign Language (EFL) learners from a STEM-focused university in northern China. This research setting enables direct access to the intersection of academic pressure and technological integration.\u003c/p\u003e \u003cp\u003eA total of 249 valid questionnaires were retained in this study. The sample includes science and engineering non-English majors (61.8%, \u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;154) and English majors (38.2%, \u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;95), both from the same institution. This study adopted a same-institution design after careful consideration: by holding constant campus culture, hardware resources, and the university\u0026rsquo;s AI integration level, the observed structural differences are more likely attributed to disciplinary differences than to confounding variables at the macro level. The gender distribution (66.7% male, 33.3% female) reflects the reality of STEM majors but also limits the generalizability of the findings in female-dominated academic contexts. A supplementary independent-samples t-test showed no significant gender differences in English Language Anxiety (ELA) and AI Partner Support (AIPS) scores (all \u003cem\u003ep\u003c/em\u003e \u0026gt; .05), indicating that the observed structural patterns were not mainly driven by gender composition.\u003c/p\u003e \u003cp\u003eAccording to structural equation modeling (SEM) criteria, the sample size of this study is adequate:Kline [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e] suggests a minimum sample size of 200 for complex path models; with fewer than 30 observed variables, the sample size (\u003cem\u003eN\u003c/em\u003e\u0026thinsp;=\u0026thinsp;249) meets the 5:1 observation-to-variable ratio recommended by Hair et al. [\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e].\u003c/p\u003e \u003cp\u003e All participants were enrolled in blended courses and actively used generative AI tools for language learning.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Measures\u003c/h2\u003e \u003cp\u003eAll constructs are measured using a 5-point Likert scale. This study conducted confirmatory factor analysis (CFA) and reliability tests, and then used item purification to optimize model fit.\u003c/p\u003e \u003cp\u003eItem purification proceeded in two stages. In the statistical purification stage, the initial CFA excluded AIPS item A1 and L2 WTC item C1, as their factor loadings fell below 0.50 or showed severe cross-loadings. For theoretical purification, we adopted an a priori approach and excluded items D1\u0026ndash;D3 and D7\u0026ndash;D9 from the PTS scale to isolate the authority-based evaluative dimension central to this study\u0026rsquo;s research questions. The relevant cultural and psychometric rationale appears in detail in Section \u003cspan refid=\"Sec11\" class=\"InternalRef\"\u003e2.2.1\u003c/span\u003e.\u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section3\"\u003e \u003ch2\u003e2.2.1 Perceived Teacher Support (PTS)\u003c/h2\u003e \u003cp\u003eThis study adapted the Multidimensional Perceived Social Support Scale (MSPSS [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]), and only the teacher support items were retained (D4\u0026ndash;D6). The exclusion of peer and family dimensions is based on two reasons. At the theoretical level, according to the framework of House [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e], the above dimensions reflect the non-evaluative buffering function, which will interfere with the social-evaluative threat that this study is concerned about. At the empirical level, a supplementary confirmatory factor analysis that retains all dimensions confirmed the decision: incorporating peer and family items will significantly reduce the model fit (CFI: .929 \u0026rarr; .878; RMSEA: .065 \u0026rarr; .078).\u003c/p\u003e \u003cp\u003eIt warrants explicit acknowledgment that items D4\u0026ndash;D6 operationalize teacher caring and academic scaffolding rather than evaluative threat per se. The theoretical linkage between authority-based scaffolding and anxiety activation is inferred through CHC cultural logic rather than directly measured. Within CHC relational frameworks, however, this inferential step is not arbitrary: Hwang [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e] argues that authority-based support is constitutively inseparable from implicit face expectations and social appraisal, such that the provision of scaffolding by an authority figure inherently activates evaluative concerns at the cultural-construct level. Empirical evidence from Chinese EFL classrooms corroborates this cultural logic: teacher strictness and authoritative evaluation are direct triggers of learner anxiety, with students reporting stronger silence tendencies under demanding teacher expectations [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e].Therefore, the distinction between the measured content (scaffolding) and the content of theoretical construction (evaluative pressure) is the boundary condition shaped by the high-power distance cultural logic (CHC) and defined by the operationalization of this study, not the measurement error.\u003c/p\u003e \u003cp\u003eThis article maintains the distinction between the perceived teacher support (PTS) in measurement and the theoretical version of PTS, and its connotation will be discussed in Section \u003cspan refid=\"Sec34\" class=\"InternalRef\"\u003e4.3\u003c/span\u003e. Accordingly, the PTS-ELA relationship should be interpreted as a statistically observed pattern \u0026mdash; the evaluation pressure mechanism behind it is inferred, not directly verified.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section3\"\u003e \u003ch2\u003e2.2.2 English Language Anxiety (ELA)\u003c/h2\u003e \u003cp\u003eThis study uses 6 items from the Foreign Language Classroom Anxiety Scale (FLCAS[\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]). The revised scale shows good internal consistency in STEM samples.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section3\"\u003e \u003ch2\u003e2.2.3 L2 Willingness to Communicate (L2 WTC)\u003c/h2\u003e \u003cp\u003eThis study adopts the 5-item scale revised by MacIntyre et al. [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. The scale meets the standards of reliability and construct validity in the target sample.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section3\"\u003e \u003ch2\u003e2.2.4 Perceived AI Partner Support (AIPS)\u003c/h2\u003e \u003cp\u003eBased on the social support framework of House [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e] and the digital learning model of Wei and Xu [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e], this study compiles a new 7-item scale. The items aim to capture the characteristics of quasi-social interaction and non-evaluative support of generative artificial intelligence, which is specifically reflected in the tolerance of language errors and no negative evaluation. The research deliberately adopts anthropomorphic expressions, such as \"comforting me like a partner\". The scale measures the functional emotional results brought about by AI interaction, not learners' ontological beliefs about AI. The design of this study constitutes an activation mechanism based on functional anthropomorphism theory [\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e] [\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e]. Discriminant validity evidence (HTMT = .711 \u0026lt; .85) confirms that participants did not confuse AI partner support (AIPS) with perceived teacher support (PTS) at the measurement level. Limitations of the operational definition are discussed in Section \u003cspan refid=\"Sec34\" class=\"InternalRef\"\u003e4.3\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eA pre-experiment with 56 participants verified the clarity of the items. In the formal survey, exploratory factor analysis (EFA) revealed a KMO value of .898, and the Bartlett test of sphericity was significant (χ\u0026sup2; = 919.762, \u003cem\u003edf\u003c/em\u003e\u0026thinsp;=\u0026thinsp;21, \u003cem\u003ep\u003c/em\u003e \u0026lt; .001). All items loaded on a single common factor, confirming unidimensionality. The Cronbach's α coefficient of the final 7-item scale is .899.\u003c/p\u003e \u003cp\u003eTo test the cross-sample stability of the unidimensional AIPS structure, this study implements split-half validation. The total sample (\u003cem\u003eN\u003c/em\u003e\u0026thinsp;=\u0026thinsp;249) is randomly divided into two subsamples using a uniform random number generator. Sub-sample 1 (\u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;128) is used for exploratory factor analysis. A single factor is extracted, with eigenvalue 4.17, explaining 59.57% of total variance. All item loadings are between .732\u0026ndash;.803. Sub-sample 2 (\u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;121) is used for confirmatory factor analysis (CFA) of the unidimensional model. Model fit is acceptable: CFI = .954, TLI = .931, RMR = .039, SRMR = .044. RMSEA (.118) exceeds the conventional threshold, which is consistent with the known characteristic that this index is sensitive to small samples[\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e]. Overall, EFA and CFA convergent evidence from independent subsamples support the replicability of the unidimensional AIPS structure.\u003c/p\u003e \u003cp\u003eTo test potential multidimensional structure, this study tests the competing model \u0026mdash; a first-order three-factor CFA model. The model shows a Heywood case (inter-factor correlation \u003cem\u003er\u003c/em\u003e\u0026thinsp;=\u0026thinsp;1.075), indicating the model is empirically unidentified under current sample constraints. Heywood cases in CFA are usually caused by: small sample size relative to model complexity, the construct being unidimensional, or item redundancy[\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e][\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e]. Since the competing model only estimates inter-factor correlations across seven items in a sample of 249, empirical underidentification is the most parsimonious interpretation. The emergence of a Heywood case therefore provides indirect evidence consistent with unidimensionality rather than against it. The unidimensional structure was retained as the only identifiable solution, and the pragmatic nature of this decision is addressed in Section \u003cspan refid=\"Sec34\" class=\"InternalRef\"\u003e4.3\u003c/span\u003e.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Data Collection Procedure\u003c/h2\u003e \u003cp\u003eA pilot study with 56 participants was conducted on November 6, 2025 to test the reliability and validity of the scale. A formal online questionnaire survey was then administered.\u003c/p\u003e \u003cp\u003e All participants provided implicit informed consent before participation. To ensure ecological validity, the study only included learners who regularly used generative AI tools to practice English. These tools included text-generating platforms (e.g., ChatGPT, ERNIE Bot) and voice-interaction programs (e.g., Doubao, Duolingo). After data collection, responses were screened and excluded if participants reported no AI usage, submitted incomplete questionnaires, or completed the survey in an unreasonably short time. Finally, 249 valid cases were obtained.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Statistical Analysis\u003c/h2\u003e \u003cp\u003eAll analyses were conducted using IBM SPSS Statistics 27.0 and AMOS 29.0. The study first computed descriptive statistics, Pearson correlations, and Harman's single-factor test to evaluate common method bias.Then, we conducted CFA to test the validity and reliability of the measurement model.Hierarchical regression was employed to examine the study hypotheses. All variance inflation factor (VIF) values were below 2.0, confirming the independence assumption.We adopted bootstrapping with 5,000 resamples to obtain robust 95% bias-corrected confidence intervals. An exploratory moderation analysis was additionally conducted using PROCESS macro version 5.0 [\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e] with 5,000 bootstrap resamples to examine whether AIPS moderated the ELA\u0026rarr;L2 WTC relationship.\u003c/p\u003e \u003cp\u003eTo examine potential sample heterogeneity, this study performed multi-group analysis (MGA) in Mplus 8.11 using the Maximum Likelihood with Robust Standard Errors (MLR) estimator. Structural paths were freely estimated across English majors and STEM learners, and the Wald test of parameter constraints was used to exploratorily examine the hypothesized emotion-behavior decoupling in the STEM group.\u003c/p\u003e \u003c/div\u003e"},{"header":"3. Results","content":"\u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Data Suitability and Common Method Bias Test\u003c/h2\u003e \u003cp\u003eKMO values (.851\u0026ndash;.910) and Bartlett's test of sphericity (\u003cem\u003ep\u003c/em\u003e \u0026lt; .001) indicate that the data are suitable for factor analysis[\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e].We further checked skewness (\u0026le;\u0026thinsp;.83) and kurtosis (\u0026le;\u0026thinsp;.156) to verify normality, which provides justification for subsequent parametric tests [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eDuring the questionnaire design stage, we adopted several procedural controls to reduce common method bias (CMB).These steps included guaranteeing participant anonymity, counterbalancing item blocks, and spatially separating support items from outcome scales [\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e]. Harman's single-factor test shows that the first unrotated factor accounts for 36.81% of the total variance, below the 50% threshold. However, the test has well-documented limitations, so the result has limited diagnostic value. This study attempted to model an Unmeasured Latent Method Construct (ULMC) through confirmatory factor analysis (CFA) but encountered a Heywood case, making it impossible to statistically quantify method variance. CMB therefore remains an unresolved threat to internal validity.\u003c/p\u003e \u003cp\u003eOne structural feature of the data bears on this concern. The bivariate correlations of PTS and AIPS with ELA were similar in magnitude (PTS: \u003cem\u003er\u003c/em\u003e = .329; AIPS: \u003cem\u003er\u003c/em\u003e = .287). If CMB inflation were uniform across both pathways, the regression model would be expected to treat them comparably. It did not: PTS\u0026rarr;ELA reached significance (β\u0026thinsp;\u003cem\u003e=\u003c/em\u003e\u0026thinsp;.246, \u003cem\u003ep\u003c/em\u003e = .001) while AIPS\u0026rarr;ELA did not (β\u0026thinsp;\u003cem\u003e=\u003c/em\u003e\u0026thinsp;.141, \u003cem\u003ep\u003c/em\u003e = .064). A uniform inflation account has difficulty explaining this divergence, which lends indirect support to the substantive interpretation of the PTS\u0026rarr;ELA path. That said, without a valid ULMC estimate, residual method variance cannot be definitively quantified, and all path coefficients should be interpreted with this caveat.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Measurement Model Assessment\u003c/h2\u003e \u003cp\u003eFollowing the item purification procedures outlined in Section \u003cspan refid=\"Sec10\" class=\"InternalRef\"\u003e2.2\u003c/span\u003e (removing PTS D1\u0026ndash;D3, AIPS A1, and L2 WTC C1), the measurement model achieved acceptable fit (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). The GFI (.883) is considered adequate for complex models [\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e][\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e].\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e\u003cb\u003eConfirmatory Factor Analysis (CFA) Model Fit Indices\u003c/b\u003e\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFit Indices\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eStudy Result\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAcademic Threshold\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eInterpretation\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCMIN/DF\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2.057\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eGood\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRMSEA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e.065\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;.08\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAcceptable\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRMR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e.045\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;.08\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eGood\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGFI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e.883\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026gt;\u0026thinsp;.8 (Acceptable)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAcceptable\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAGFI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e.851\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026gt;\u0026thinsp;.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAcceptable adjusted fit\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCFI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e.929\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026gt;\u0026thinsp;.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eGood\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTLI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e.918\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026gt;\u0026thinsp;.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eGood\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePNFI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e.753\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026gt;\u0026thinsp;.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eGood\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePCFI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e.802\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026gt;\u0026thinsp;.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eGood\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"4\"\u003e\u003cem\u003eNote\u003c/em\u003e. N\u0026thinsp;=\u0026thinsp;249. Thresholds follow [\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e] .\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cdiv id=\"Sec20\" class=\"Section3\"\u003e \u003ch2\u003e3.2.1 Reliability and Convergent Validity\u003c/h2\u003e \u003cp\u003eInternal consistency and convergent validity were strong (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). All Cronbach's α and Composite Reliability (CR) values were above .80, while Average Variance Extracted (AVE) scores exceeded the .50 threshold. Additionally, all factor loadings were statistically significant (\u003cem\u003ep \u0026lt;\u003c/em\u003e .001).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e.\u003cb\u003eReliability and Convergent Validity Results (Cronbach's α\u0026thinsp;+\u0026thinsp;CR\u0026thinsp;+\u0026thinsp;AVE)\u003c/b\u003e\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eScale\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eItems\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCronbach's α\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCR\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAVE\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAIPS (A2-A8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e.899\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e.897\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.554\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eELA (B1-B6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e.864\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e.865\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.517\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eL2 WTC (C2-C5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e.826\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e.828\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.548\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePTS (D4-D6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e.820\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e.833\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.624\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"5\"\u003e\u003cem\u003eNote\u003c/em\u003e. \u003cem\u003eN\u003c/em\u003e\u0026thinsp;=\u0026thinsp;249. \u003cem\u003eCr\u0026thinsp;=\u003c/em\u003e\u0026thinsp;Composite Reliability; \u003cem\u003eAVE\u003c/em\u003e\u0026thinsp;=\u0026thinsp;Average Variance Extracted.\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec21\" class=\"Section3\"\u003e \u003ch2\u003e3.2.2 Discriminant Validity\u003c/h2\u003e \u003cp\u003eDiscriminant validity was assessed using the Fornell-Larcker criterion (Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). The square roots of the AVE values for each construct, shown on the diagonal, were greater than the correlations between the constructs, indicating that all constructs were statistically distinct. Crucially, the Fornell-Larcker criterion results provide robust statistical evidence against potential measurement confounding regarding the AIPS scale. Despite the functional anthropomorphism deliberately utilized in the AIPS items (e.g., 'comforts me'), the square roots of the AVEs for both PTS (.790) and AIPS (.744) strictly exceeded their inter-construct correlation (r = .615).\u003c/p\u003e \u003cp\u003eWe computed the Heterotrait-Monotrait Ratio (HTMT) to better establish discriminant validity and formally evaluate the presence of common method bias (CMB).Between PTS and AIPS, the largest HTMT value reached .711, well beneath the .85 cutoff widely adopted in structural equation modeling research [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e] [\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e].These results offer additional empirical support for the distinctiveness of all focal constructs at the measurement stage.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e\u003cb\u003eDiscriminant Validity Test (Square Root of AVE vs. Correlation Coefficients)\u003c/b\u003e\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVariables\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePTS\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eL2 WTC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eELA\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAI PS\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePTS\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003e(.790)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eL2 WTC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e.494**\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e(.740)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eELA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e.329**\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e.298**\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e(.719)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAIPS\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e.615**\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e.501**\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e.287**\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e(.744)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"5\"\u003e\u003cem\u003eNote\u003c/em\u003e. \u003cem\u003eN\u003c/em\u003e\u0026thinsp;=\u0026thinsp;249. Diagonal values in bold and parentheses are the square roots of the AVE; off-diagonal values are Pearson correlations. ** \u003cem\u003ep \u0026lt;\u003c/em\u003e .01.\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec22\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Descriptive Statistics and Correlation Analysis\u003c/h2\u003e \u003cp\u003eDescriptive statistics and Pearson correlations for all focal variables are displayed in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e.These correlational results offer initial empirical support for the proposed research model. Both PTS and AIPS showed significant positive associations with L2 WTC (\u003cem\u003ep\u003c/em\u003e \u0026lt; .01). Of particular interest, ELA also yielded a significant positive correlation with L2 WTC (\u003cem\u003er\u003c/em\u003e = .298, \u003cem\u003ep\u003c/em\u003e \u0026lt; .01), a pattern whose theoretical implications are elaborated in the note below.\u003c/p\u003e \u003cp\u003eTheoretical Anomaly Note. Classical SLA theories yield a clear directional expectation for the current sample: with a mean ELA score of 3.68 and widespread self-reported communication apprehension, ELA should relate negatively and significantly to L2 WTC[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e][\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. However, the observations deviate from this prediction at two distinct levels. At the bivariate level, there is a significant positive correlation between ELA and L2 WTC (\u003cem\u003er\u003c/em\u003e = .298, \u003cem\u003ep\u003c/em\u003e \u0026lt; .01), which is directly contrary to the prediction of classical theory. After controlling the double support source in the multivariate model, ELA does not have a significant inhibitory effect on L2 WTC (β\u0026thinsp;=\u0026thinsp;.105, \u003cem\u003ep\u003c/em\u003e = .063), and the coefficient is still positive.\u003c/p\u003e \u003cp\u003eOverall, these findings reflect a structural anomaly that cannot be fully explained by insufficient statistical testing: under the framework of traditional second language acquisition (SLA) theory, even if social support is controlled, samples with this average level of anxiety should still present a negative ELA\u0026rarr;WTC path. Therefore, the consistent positive correlation across the analytical level needs to go beyond the simple interpretation of \"no significant effect\" and carry out in-depth theoretical discussion, which also provides a basis for the analysis of Section \u003cspan refid=\"Sec30\" class=\"InternalRef\"\u003e4.1.2\u003c/span\u003e of this study.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e\u003cb\u003eCorrelation Matrix of Key Constructs\u003c/b\u003e\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVariable\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eM\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSD\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1. PTS\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e3.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e.77\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2. L2 WTC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e3.62\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e.84\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e.494**\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e3. ELA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e3.68\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e.81\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e.329**\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.298**\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e4. AIPS\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e3.89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e.75\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e.615**\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.501**\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e.287**\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"7\"\u003e\u003cem\u003eNote. N\u003c/em\u003e\u0026thinsp;=\u0026thinsp;249. ** \u003cem\u003ep\u003c/em\u003e \u0026lt; .01 (two-tailed).\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec23\" class=\"Section2\"\u003e \u003ch2\u003e3.4 Hypothesis Testing: Hierarchical Regression Results\u003c/h2\u003e \u003cp\u003eHierarchical regression analyses were conducted to sequentially test the proposed hypotheses. The results predicting L2 WTC are presented in Table\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e5\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eHierarchical Regression Analysis Results\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVariables\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eModel 1(Controls )\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eModel 2(Dual Support)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eModel 3(Anxiety Inclusion)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eModel 4 (Predicting ELA)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDependent Variable\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cem\u003eL2 WTC\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cem\u003eL2 WTC\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eL2 WTC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eELA\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGender\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eβ\u0026thinsp;\u003cem\u003e=\u0026thinsp;.032, p=.636\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eβ\u003cem\u003e= \u0026minus;\u0026thinsp;.011\u003c/em\u003e,\u003c/p\u003e \u003cp\u003e\u003cem\u003ep=.846\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eβ= \u0026minus;\u0026thinsp;.018,\u003c/p\u003e \u003cp\u003e\u003cem\u003ep\u003c/em\u003e=.752\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eβ\u0026thinsp;=\u0026thinsp;.064,\u003c/p\u003e \u003cp\u003e\u003cem\u003ep\u003c/em\u003e=.312\u003c/p\u003e\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMajor\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eβ\u003cem\u003e= \u0026minus;\u0026thinsp;.109, p=.105\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eβ= \u0026minus;\u0026thinsp;.164**,\u003c/p\u003e \u003cp\u003e\u003cem\u003ep\u003c/em\u003e=.003\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eβ= \u0026minus;\u0026thinsp;.154**,\u003c/p\u003e \u003cp\u003e\u003cem\u003ep\u003c/em\u003e=.006\u003c/p\u003e\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eβ= \u0026minus;\u0026thinsp;.098,\u003c/p\u003e \u003cp\u003e\u003cem\u003ep\u003c/em\u003e = .120\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePTS (Centered)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eβ\u0026thinsp;=\u0026thinsp;.311***,\u003c/p\u003e \u003cp\u003e\u003cem\u003ep\u003c/em\u003e\u0026lt;.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eβ\u0026thinsp;=\u0026thinsp;.285***,\u003c/p\u003e \u003cp\u003e\u003cem\u003ep\u003c/em\u003e\u0026lt;.001\u003c/p\u003e\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eβ\u0026thinsp;=\u0026thinsp;.246***,\u003c/p\u003e \u003cp\u003e\u003cem\u003ep\u003c/em\u003e = .001\u003c/p\u003e\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAIPS(Centered)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eβ\u0026thinsp;=\u0026thinsp;.319***,\u003c/p\u003e \u003cp\u003e\u003cem\u003ep\u003c/em\u003e\u0026lt;.001\u003c/p\u003e\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eβ\u0026thinsp;=\u0026thinsp;.304***,\u003c/p\u003e \u003cp\u003e\u003cem\u003ep\u003c/em\u003e\u0026lt;.001\u003c/p\u003e\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eβ\u0026thinsp;=\u0026thinsp;.141,\u003c/p\u003e \u003cp\u003e\u003cem\u003ep\u003c/em\u003e= .064\u003c/p\u003e\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eELA (Anxiety)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eβ\u0026thinsp;=\u0026thinsp;.105,\u003c/p\u003e \u003cp\u003ep=.063\u003c/p\u003e\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eR\u0026sup2;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e.015\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e.332\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e.342\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.137\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eΔR\u0026sup2;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e.015\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e.317\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e.009\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.137\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eF\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.882\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e57.923***\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3.497\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e9.712***\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"5\"\u003e\u003cem\u003eNote. N\u003c/em\u003e\u0026thinsp;=\u0026thinsp;249. Standardized coefficients (\u003cem\u003eβ\u003c/em\u003e) are reported. * \u003cem\u003ep\u003c/em\u003e \u0026lt; .05, ** \u003cem\u003ep\u003c/em\u003e \u0026lt; .01, *** \u003cem\u003ep\u003c/em\u003e \u0026lt; .001.\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cdiv id=\"Sec24\" class=\"Section3\"\u003e \u003ch2\u003e3.4.1 Regression Model Interpretations\u003c/h2\u003e \u003cp\u003eStep 1: Robustness Check with Control Variables\u003c/p\u003e \u003cp\u003eThis study incorporates demographic variables (gender and major) to test whether there is a systematic deviation in a male-dominated STEM sample. The explanatory power of the benchmark model is extremely low (\u003cem\u003eR\u0026sup2;\u003c/em\u003e = .015), and gender is not a significant predictor.\u003c/p\u003e \u003cp\u003eStep 2: Dual-Path Support Validation\u003c/p\u003e \u003cp\u003eAfter adding PTS and AIPS, the model\u0026rsquo;s explained variance increased significantly. Both predictors reached significance (β\u0026thinsp;\u003cem\u003e=\u003c/em\u003e\u0026thinsp;.311, \u003cem\u003ep\u003c/em\u003e \u0026lt; .001 for PTS; β\u0026thinsp;=\u0026thinsp;.319, \u003cem\u003ep\u003c/em\u003e \u0026lt; .001 for AIPS), supporting H1. Learners maintained communicative intent through both human and AI support.\u003c/p\u003e \u003cp\u003eStep 3: Differentiated effect on anxiety (test H2)\u003c/p\u003e \u003cp\u003eModel 4 is tested with ELA as the dependent variable.PTS has a significant positive correlation with anxiety (β\u0026thinsp;=\u0026thinsp;.246, \u003cem\u003ep\u003c/em\u003e = .001), which is consistent with the inferential operation definition described in Section \u003cspan refid=\"Sec11\" class=\"InternalRef\"\u003e2.2.1\u003c/span\u003e. This direction is in line with theoretical expectations, but in view of the cross-sectional design, the reverse interpretation is still valid: anxious learners may actively seek more teacher support as a compensatory response. In contrast, AIPS has no significant effect on anxiety (β\u0026thinsp;=\u0026thinsp;.141, \u003cem\u003ep\u003c/em\u003e = .064). This model supports hypothesis H2 in direction: Authority-based support will aggravate evaluation anxiety, while AI support will not.\u003c/p\u003e \u003cp\u003eIn order to formally test whether there is a significant difference between the two coefficients, this study uses the non-standardized regression coefficient to conduct the Paternoster z test[\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e]. The test result is not significant (\u003cem\u003ez\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.943, \u003cem\u003ep\u003c/em\u003e = .346), indicating that under the current sample size, the PTS\u0026rarr;ELA path (\u003cem\u003eB\u003c/em\u003e = .260, \u003cem\u003eSE\u003c/em\u003e = .080) and the AIPS\u0026rarr;ELA path (\u003cem\u003eB\u003c/em\u003e = .152, \u003cem\u003eSE\u003c/em\u003e = .082) is within the sampling error range. The result recalibrates the interpretation of H2: this study data does not support the formal distinction between the two effects, but supports the directional model consistent with H2. The positive correlation strength of PTS and ELA is higher than that of AIPS, but to achieve the distinction at the coefficient level, repeated verification with a larger sample is needed (recommended sample size \u003cem\u003eN\u003c/em\u003e\u0026thinsp;\u0026gt;\u0026thinsp;500). Nevertheless, this directional difference is consistent with the theoretical distinction between evaluative support and non-evaluative support, and the structural argument of the common method deviation presented in Section \u003cspan refid=\"Sec18\" class=\"InternalRef\"\u003e3.1\u003c/span\u003e further supports its substantive interpretation.\u003c/p\u003e \u003cp\u003eStep 4: The Non-Inhibitory Pattern of ELA (Model 3)\u003c/p\u003e \u003cp\u003eModel 3 incorporates ELA with the dual support variables to predict L2 WTC. ELA did not emerge as a significant negative predictor (β\u0026thinsp;=\u0026thinsp;.105, \u003cem\u003ep\u003c/em\u003e = .063). The positive direction of the coefficient warrants an alternative interpretation: the positive bivariate correlation between ELA and WTC (\u003cem\u003er\u003c/em\u003e = .298) may reflect shared variance from an unmeasured third variable (most plausibly academic survival motivation). After controlling for PTS and AIPS, the path became non-significantly positive, which to some extent reflects a suppression-like pattern in the dual-support ecology.\u003c/p\u003e \u003cp\u003eExisting cross-sectional data cannot rule out this explanation. To empirically arbitrate among these competing accounts, an exploratory moderation analysis was conducted and is reported in Section \u003cspan refid=\"Sec25\" class=\"InternalRef\"\u003e3.5\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eTo examine whether the non-significant result reflects a genuine absence of inhibitory effects rather than insufficient statistical power, this study adopted the Two One-Sided Tests (TOST) equivalence procedure [\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e]. The equivalence bound was set at Δβ = \u0026plusmn;0.20, based on the smallest effect size of practical significance for instructional design in applied linguistics [\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e]. Effects below this threshold are practically meaningless for teaching design.The TOST result was significant at this bound (\u003cem\u003ep\u003c/em\u003e = .026). A stricter bound of \u0026plusmn;\u0026thinsp;0.10 was also tested; equivalence was not established (\u003cem\u003ep\u003c/em\u003e = .210), indicating that under a more conservative standard, the effect cannot be judged as negligible.Therefore, the main equivalence conclusion holds at \u0026plusmn;\u0026thinsp;0.20 and is explicitly boundary-dependent.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec25\" class=\"Section2\"\u003e \u003ch2\u003e3.5 Exploratory Moderation Analysis\u003c/h2\u003e \u003cp\u003eIn order to provide an empirical basis for the preliminary distinction of the three competitive interpretations proposed in Section \u003cspan refid=\"Sec30\" class=\"InternalRef\"\u003e4.1.2\u003c/span\u003e, this study uses PROCESS Model 1 [\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e] for exploratory moderation effect analysis and sets up 5,000 Bootstrap resampling. And centralize the processing of the predicted variables. Take English Language Anxiety (ELA) as the predictive variable, L2 Willingness to Communicate (L2 WTC) as the result variable, and AI Partner Support (AIPS) as the moderator. The interaction term ELA \u0026times; AIPS is not significant (\u003cem\u003eB\u003c/em\u003e\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;.007, \u003cem\u003eSE\u003c/em\u003e = .005, p = .171, 95% CI [\u0026minus;\u0026thinsp;.020, .007]), indicating that AIPS does not moderate the relationship between ELA and L2 WTC. The overall model accounted for 28.25% of variance in L2 WTC (R\u003cem\u003e\u0026sup2;\u003c/em\u003e = .283, \u003cem\u003eF\u003c/em\u003e(3, 245)\u0026thinsp;=\u0026thinsp;32.15, \u003cem\u003ep\u003c/em\u003e \u0026lt; .001). Theoretical interpretation of this result is provided in Section \u003cspan refid=\"Sec30\" class=\"InternalRef\"\u003e4.1.2\u003c/span\u003e.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec26\" class=\"Section2\"\u003e \u003ch2\u003e3.6 Multi-Group Analysis: An Exploratory Cross-Discipline Signal\u003c/h2\u003e \u003cp\u003eThis study uses the MLR estimation method to conduct an exploratory multi-group analysis in Mplus 8.11. Full metric invariance between groups is supported (ΔCFI = .002) [\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e], indicating that the measurement framework is functionally equivalent in two subsamples.\u003c/p\u003e \u003cp\u003eAlthough the sub-sample size is relatively small, the baseline configural model still shows an acceptable fit (CFI = .865, RMSEA = .070).The interdisciplinary structure path shows a clear direction pattern:For English majors, anxiety has a significant positive predictive effect on willingness to communicate (WTC) (β\u0026thinsp;=\u0026thinsp;.314, \u003cem\u003ep\u003c/em\u003e = .025); There is no significant predictive effect on STEM students (β\u0026thinsp;=\u0026thinsp;.029, \u003cem\u003ep\u003c/em\u003e = .742). The bootstrapped contrast did not reach the conventional significant level (Δβ\u0026thinsp;=\u0026thinsp;.285, \u003cem\u003ep\u003c/em\u003e = .088). The post-hoc power analysis showed that the statistical power was 62%, which was lower than the generally accepted 80% standard.\u003c/p\u003e \u003cp\u003eTherefore, the differences between the groups should be regarded as a suggestive trend rather than a solid empirical result that can be strongly inferred. Discipline-based comparative research usually requires 150\u0026ndash;200 subjects per group to obtain sufficient statistical power to make effective cross-group inference.\u003c/p\u003e \u003c/div\u003e"},{"header":"4. Discussion","content":"\u003cdiv id=\"Sec28\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Findings\u003c/h2\u003e \u003cp\u003eWe examined how PTS and AIPS jointly shape L2 WTC and ELA in Chinese STEM learners.Our findings offer robust support for the additive positive effect of the dual-support system (H1).\u003c/p\u003e \u003cp\u003eFor Hypothesis 2, the Paternoster z-test showed that the two coefficients were not statistically different under the current sample size (\u003cem\u003ez\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.943, \u003cem\u003ep\u003c/em\u003e = .346). Nevertheless, the observed directional pattern was consistent with H2: the positive link between PTS and ELA (β\u0026thinsp;=\u0026thinsp;.246, \u003cem\u003ep\u003c/em\u003e = .001) was stronger than the corresponding association for AIPS (β\u0026thinsp;=\u0026thinsp;.141, \u003cem\u003ep\u003c/em\u003e = .064).\u003c/p\u003e \u003cp\u003eThis pattern suggests that PTS and AIPS differ in strength rather than nature in their relations with ELA: both convey evaluative signals, yet PTS conveys a markedly stronger signal.The formal statistical separation of the two paths is regarded as a target for subsequent replication, rather than a confirmed research finding. A pattern consistent with H3 also emerged: ELA did not inhibit L2 WTC, and its positive coefficient across both bivariate and multivariate levels constitutes a dual-level anomaly elaborated in Section \u003cspan refid=\"Sec30\" class=\"InternalRef\"\u003e4.1.2\u003c/span\u003e. All paths involving PTS and ELA represent culturally inferred associations, bounded by the CHC heuristic framework established in Section \u003cspan refid=\"Sec11\" class=\"InternalRef\"\u003e2.2.1\u003c/span\u003e, and the cross-sectional design does not permit definitive causal inferences.\u003c/p\u003e \u003cdiv id=\"Sec29\" class=\"Section3\"\u003e \u003ch2\u003e4.1.1 The Synergistic Drivers: Dual-Path Support on WTC\u003c/h2\u003e \u003cp\u003eBoth PTS and AIPS have become powerful parallel predictors of L2 WTC, which is consistent with H1. These findings extend the classic social support theories [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e][\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e] to the AI-assisted learning context. The correlation coefficient between PTS and WTC (\u003cem\u003er\u003c/em\u003e = .494) shows that even if there is evaluative pressure, authority-based teacher support still retains its scaffolding function in Confucian culture. At the same time, AIPS demonstrates independent predictive power\u0026mdash;which suggests that generative AI may not only be used as a practice tool, but also as a structurally distinct support source that works in parallel with human scaffolding. The two forms of support appear to work in tandem rather than overlap: students receive instructional support from human instructors while drawing affective reassurance from AI.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec30\" class=\"Section3\"\u003e \u003ch2\u003e4.1.2 A Directional Anomaly in the Classic Affective Filter: Boundary Conditions and Theoretical Implications\u003c/h2\u003e \u003cp\u003eThe dual-level oddity in Section \u003cspan refid=\"Sec22\" class=\"InternalRef\"\u003e3.3\u003c/span\u003e requires comprehensive theoretical consideration.Traditional SLA frameworks [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e][\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e] indicate a significant negative ELA\u0026rarr;WTC route in our sample, with a mean ELA of 3.68.ELA positively correlates with WTC at the bivariate level (\u003cem\u003er\u003c/em\u003e =\u0026thinsp;.298, \u003cem\u003ep\u003c/em\u003e \u0026lt;\u0026thinsp;.01) but does not inhibit WTC at the multivariate level (β\u0026thinsp;=\u0026thinsp;.105, \u003cem\u003ep\u003c/em\u003e =\u0026thinsp;.063).Insufficient statistical power is hard to explain when both pieces of data point in the same unusual direction.Cross-sectional data cannot discriminate between three plausible competing theories.\u003c/p\u003e \u003cp\u003eThe first is a suppressor explanation. An unmeasured third variable, most likely academic survival motivation driven by the communicative imperative [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e], simultaneously increases both ELA and WTC. English is high-stakes and thus anxiety-inducing for ELA, and high-stakes and thus worth attempting for WTC. This pattern produces the observed positive bivariate correlation while suppressing the inhibitory path in regression. This account requires no new theoretical assumptions and only invokes an unmeasured variable.This is consistent with the communication necessity framework in Section \u003cspan refid=\"Sec2\" class=\"InternalRef\"\u003e1.1\u003c/span\u003e: Academic survival motivation drives both English Language Anxiety (ELA) and Willingness to Communicate (WTC).\u003c/p\u003e \u003cp\u003eThe second explanation is the facilitative anxiety perspective. Under task pressure, STEM learners' ELA may manifest as facilitative rather than debilitating anxiety [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e][\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e], that is, activating cognitive resources rather than triggering avoidance behavior. The FLCAS revised scale adopted in this study does not distinguish between facilitative debilitating anxiety subtypes, so this possibility cannot be ruled out.\u003c/p\u003e \u003cp\u003eThe third explanation is the functional substitution perspective. Learners can hand over the highly anxious trial and error process to non-evaluative AI, so as to structurally remove the connection between ELA and communication avoidance. The view is that with the improvement of AIPS, ELA's inhibition path to WTC will weaken, which is different from the first two explanations.\u003c/p\u003e \u003cp\u003eIn order to test whether functional substitution can explain the non-suppressive mode of ELA, this study uses PROCESS Model 1[\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e] for exploratory moderation analysis, set up 5,000 Bootstrap resampling and centralize the prediction variables. The interaction term ELA \u0026times; AIPS is not significant (\u003cem\u003eB\u003c/em\u003e\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;.007, \u003cem\u003eSE\u003c/em\u003e = .005, \u003cem\u003ep\u003c/em\u003e = .171, 95% CI [\u0026minus;\u0026thinsp;.020, .007]), indicating that AIPS does not moderate the relationship between ELA and L2 WTC. This null interaction result suggests that the non-inhibitory ELA pattern is unlikely to be driven by functional substitution alone, lending indirect support to the suppressor variable interpretation as the more parsimonious account under the present data.\u003c/p\u003e \u003cp\u003ePragmatic Resilience serves as a conceptual label to guide systematic investigation of this anomaly.The three mechanisms identified are not mutually exclusive: academic survival motivation may establish the baseline, facilitative anxiety may shape task engagement, and AI-enabled functional substitution may act as a secondary mechanism.Distinguishing among these accounts requires longitudinal designs, experimental manipulation of AI use, and direct measures of academic survival motivation.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec31\" class=\"Section3\"\u003e \u003ch2\u003e4.1.3 AI's Nonjudgmental Role: A Safety Valve for Learner Anxiety\u003c/h2\u003e \u003cp\u003ePTS aids communication, however statistics show its dual nature in CHC situations.Students often regard the intensive support of teachers as the embodiment of academic care and strong authority[\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. The mechanism of AIPS is completely different: non-evaluative gaze [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e] creates a unique emotional space that has a non-significant impact on anxiety (β\u0026thinsp;=\u0026thinsp;.141, \u003cem\u003ep\u003c/em\u003e = .064).\u003c/p\u003e \u003cp\u003eThis contradiction is not a simple opposition. This study calls it \"ontological dualism\".\u003c/p\u003e \u003cp\u003eLearners can get psychological comfort in the anthropomorphic language of AI without worrying about evaluation. Based on the recent expansion of media equation theory [\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e] and the mind perception framework [\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e], this study believes that users can anthropomorphize the intelligent body for social and emotional benefits, while maintaining non-human ontology awareness at the metacognitive level.\u003c/p\u003e \u003cp\u003eSTEM students may adopt the attitude of \"willing suspension of disbelief\" in this digital learning ecology: using AI's functional anthropomorphism to meet emotional needs, while avoiding the risk of \"face\" in reality with the help of its machine ontological attributes.\u003c/p\u003e \u003cp\u003eThis makes AIPS a unique psychological affordance, rather than a simple reproduction of human interaction.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec32\" class=\"Section3\"\u003e \u003ch2\u003e4.1.4 Alternative Explanations: The Bidirectionality of Support and Anxiety\u003c/h2\u003e \u003cp\u003eThe cross-sectional design limits the causal interpretation of the PTS\u0026ndash;ELA path.In the Chinese cultural context (CHC), evaluative support may exacerbate anxiety due to \u0026ldquo;face\u0026rdquo; concerns [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e];but highly anxious learners may also actively seek more teacher support as a compensatory response driven by communicative imperative.Longitudinal data are needed to clarify these causal directions.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec33\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Practical Implications\u003c/h2\u003e \u003cp\u003eThe structural model supports testable teaching heuristic strategies, such as the Strategic Sandwich model (Human-AI-Human). This paper proposes a design-based research agenda to carry out experimental verification instead of constructing a deterministic causal model (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eBased on control-value theory [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e], the process will hand over anxiety-inducing practice to AI, while retaining the motivational authority of teachers.\u003c/p\u003e \u003cp\u003eStage 1: Teachers lead input and motivation. Teachers set academic goals and emphasize the value of tasks; evaluative stakes and communicative imperative are established at this stage.\u003c/p\u003e \u003cp\u003eStage 2: AI-assisted buffer. Students use generative AI for low-risk private exercises. The non-evaluative environment promotes trial and error and no \"face\" cost, so that learners can gain a sense of control before high-stakes performance and improve linguistic accuracy.\u003c/p\u003e \u003cp\u003eStage 3: Output and Performance (Interpersonal Interaction). Classroom display and peer communication enable students to return to the interpersonal environment. Evaluative pressure still exists, but the early AI practice reduces the cognitive load and avoids anxiety-induced avoidance behavior. Anxiety is addressed rather than suppressed.\u003c/p\u003e \u003cp\u003eThe model does not advocate that AI replaces the affective needs of interpersonal interaction, but proposes that putting AI exercises before interpersonal performance can reduce the threshold of behavioral disruption caused by anxiety. This hypothesis can be tested by the experimental design of random assignment to sequencing conditions.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eNote\u003c/strong\u003e \u003cp\u003eThe \"Strategic Sandwich\" model integrates the dual-path ecology into a three-stage sequence: teacher-led goal setting (stage 1) to induce evaluation pressure, AI-assisted drill (stage 2) to provide a nonjudgmental buffer, and interpersonal interaction (stage 3) to implement the necessity of communication. AI acts as an emotional scaffold, not a substitute for real social interaction.\u003c/p\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec34\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Limitations and Future Research\u003c/h2\u003e \u003cp\u003eThere are significant interpretation limitations in this research, and the most fundamental constraint is cross-sectional design. There may be a bidirectional relationship between ELA and WTC: continuous positive AI interaction may reduce baseline anxiety, but existing research cannot verify this, and it is necessary to use longitudinal design to track such dynamic reciprocal relationships.\u003c/p\u003e \u003cp\u003eThere are two factors that affect the generalizability: English majors in a single institution may not represent all language majors in each university, and it is necessary to carry out multi-institutional replication to judge whether the discipline-specific effects observed in this study are universal; the sub-sample size of multi-group analysis (\u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;95; \u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;154) results in statistical power insufficient (62%), which is lower than the threshold of 80%. In future studies, each subgroup should ensure 150\u0026ndash;200 subjects before drawing discipline-specific conclusions.\u003c/p\u003e \u003cp\u003eSampling criteria bring inherent tension: From a methodological point of view, limiting AIPS evaluation to learners with experience in using AI will limit the generalizability to non-AI-user groups and may lead to a ceiling effect on the baseline AIPS score. Future research should include a control group without AI experience to establish the boundary conditions of the support effect.\u003c/p\u003e \u003cp\u003eFuture research must solve the two major measurement limitations of the AIPS scale: because its anthropomorphic language is not verified in combination with the respondents' explicit AI ontology awareness, some subjects may confuse AI support with social support; Heywood case appears in the competing CFA model, indicating the unidimensional structure is more practical than the theoretically grounded structure. The psychometric properties of the scale must be confirmed by cross-sample replication and MTMM-based convergent validity testing.\u003c/p\u003e \u003cp\u003eThe conceptual boundary of PTS is deliberate methodological choice, not omission. Items D4\u0026ndash;D6 measure teacher care and academic scaffolding; The evaluative pressure at the core of this study's PTS-ELA theory is inferred based on Chinese cultural logic (CHC) \u0026mdash; in this culture, authority-based support is intrinsically related to face expectations and social appraisal [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThis inferential gap means that the observed PTS-ELA significant correlation (β\u0026thinsp;=\u0026thinsp;.246, \u003cem\u003ep\u003c/em\u003e = .001) should be regarded as a statistically observed pattern, and the evaluative pressure mechanism behind it comes from cultural inference rather than direct verification. In order to bridge this conceptual gap, future research should develop and validate a dedicated \u0026ldquo;Perceived Academic Evaluative Threat\u0026rdquo; scale, which should be combined with PTS scaffolding items to directly empirically test the evaluative pressure mechanism proposed by this study.\u003c/p\u003e \u003cp\u003eThe data of this study only partially distinguish the three competing explanations for the ELA non-inhibitory pattern proposed in Section \u003cspan refid=\"Sec30\" class=\"InternalRef\"\u003e4.1.2\u003c/span\u003e. Exploratory moderation analysis provides indirect evidence against the \"functional substitution as main effect\", but limited by cross-sectional design, the suppressor variable account and the facilitative anxiety account cannot be empirically separated. Future research should directly measure academic survival motivation and distinguish these explanations through experimental manipulation of AI availability.\u003c/p\u003e \u003cp\u003eQuantitative designs can portray the structural pattern, but they cannot touch learners' lived experiences. This study constitutes the first stage of an explanatory sequential mixed-methods design; the subsequent qualitative stage will use in-depth interviews to analyze how STEM students cope with teacher evaluation pressure and why AI is constructed as an affectively safe space, thus enriching the boundary conditions of this study.\u003c/p\u003e \u003c/div\u003e"},{"header":"5. Conclusion","content":"\u003cp\u003eUnder the human-AI dual support context, this study examines how PTS and AIPS jointly affect L2 WTC and ELA among Chinese STEM learners.\u003c/p\u003e \u003cp\u003eAt the confirmatory analysis level, the two sources of support independently and additively predict L2 WTC (ΔR\u0026sup2; = .317). PTS provides academic scaffolding, while AIPS offers a non-judgmental space that does not amplify the inherent evaluative pressure of teacher feedback. This dual-pathway structure constitutes a replicable empirical pattern with direct implications for instructional design.\u003c/p\u003e \u003cp\u003eAt the exploratory level, ELA did not act as a significant inhibitor of L2 WTC. Compared with classical SLA predictions, this finding reveals a dual-level anomaly. Pragmatic Resilience serves as a conceptual framework to guide future research on this boundary condition, rather than functioning as a settled theoretical explanation.\u003c/p\u003e \u003cp\u003eInterpersonal support and AI support can coexist as functionally distinct resources: the former establishes academic standards, while the latter buffers affective costs. Whether such coexistence can structurally reshape the classic anxiety\u0026ndash;communication relationship remains an open empirical question.\u003c/p\u003e"},{"header":"Abbreviations","content":" \u003cdiv class=\"Description\"\u003e \u003cp\u003e \u003cb\u003eFull term\u003c/b\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eAIPS\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eAI Partner Support\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eCFA\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eConfirmatory Factor Analysis\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eCHC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eChinese Heritage Culture\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eEFA\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eExploratory Factor Analysis\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eEFL\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eEnglish as a Foreign Language\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eELA\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eEnglish Language Anxiety\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eL2 WTC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eSecond Language Willingness to Communicate\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eMGA\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eMulti-Group Analysis\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003ePTS\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003ePerceived Teacher Support\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eSLA\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eSecond Language Acquisition\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eSTEM\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eScience, Technology, Engineering, and Mathematics\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e \u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e \u003cp\u003e This study was approved by the Ethics Review Committee of the University International College (UIC), Macau University of Science and Technology (MUST) (Ref. No. : UIC/L/26/132; approved November 1, 2025), and was conducted in accordance with the Declaration of Helsinki. All participants received a comprehensive introduction prior to participation; completing and submitting the questionnaire was treated as implied informed consent. Participant anonymity and data confidentiality were strictly maintained throughout, and participants retained the right to withdraw at any time without consequence.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eConsent for publication\u003c/strong\u003e \u003cp\u003eNot applicable.\u003c/p\u003e \u003c/p\u003e\u003cp\u003e \u003ch2\u003eCompeting interests\u003c/h2\u003e \u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eFunding\u003c/h2\u003e \u003cp\u003eNo external funding was received for this study.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eCX conceived and designed the study, collected and analysed the data, and wrote the manuscript. WZ supervised the study and provided critical revisions. All authors read and approved the final manuscript.\u003c/p\u003e\u003ch2\u003eAcknowledgement\u003c/h2\u003e\u003cp\u003eThe author used ChatGPT (OpenAI) for grammar checking and language editing during manuscript preparation. All content was reviewed and revised by the author, who takes full responsibility for the final manuscript.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eMacIntyre PD, Cl\u0026eacute;ment R, D\u0026ouml;rnyei Z, Noels KA. Conceptualizing willingness to communicate in a L2: A situational model of L2 confidence and affiliation. Mod Lang J. 1998;82(4):545\u0026ndash;62. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1111/j.1540-4781.1998.tb05543.x\u003c/span\u003e\u003cspan address=\"10.1111/j.1540-4781.1998.tb05543.x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMystkowska-Wiertelak A, Pawlak M. Willingness to communicate in instructed second language acquisition: Combining a macro- and micro-perspective. Multilingual Matters; 2017.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYashima T, MacIntyre PD, Ikeda M. Situated willingness to communicate in an L2: Interplay of individual characteristics and context. Lang Teach Res. 2018;22(1):115\u0026ndash;37. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1177/1362168816657851\u003c/span\u003e\u003cspan address=\"10.1177/1362168816657851\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWei X, Xu Q. Predictors of willingness to communicate in a second language (L2 WTC): Toward an integrated L2 WTC model from the socio-psychological perspective. Foreign Lang Ann. 2022;55(1):258\u0026ndash;82. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1111/flan.12595\u003c/span\u003e\u003cspan address=\"10.1111/flan.12595\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYu Q, Xu D, Wan Q. Exploring the relationship between L2 motivational self system, academic emotions, and L2 willingness to communicate among university EFL learners in central and western China. Acta Psychol. 2025;261:105969. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.actpsy.2025.105969\u003c/span\u003e\u003cspan address=\"10.1016/j.actpsy.2025.105969\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKatsaris T. The Willingness to Communicate (WTC): Origins, significance, and propositions for the L2/FL classroom. \u003cem\u003eJ Appl Lang Linguist\u003c/em\u003e. 2019;3(2):31\u0026ndash;42. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://als-edu.wixsite.com/jall\u003c/span\u003e\u003cspan address=\"https://als-edu.wixsite.com/jall\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Accessed 24 Feb 2026.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHorwitz EK, Horwitz MB, Cope J. Foreign language classroom anxiety. Mod Lang J. 1986;70(2):125\u0026ndash;32. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1111/j.1540-4781.1986.tb05256.x\u003c/span\u003e\u003cspan address=\"10.1111/j.1540-4781.1986.tb05256.x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDornyei Z. The L2 motivational self system. In: Dornyei Z, Ushioda E, editors. Motivation, language identity and the L2 self. Multilingual Matters; 2009. pp. 9\u0026ndash;42.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDuff PA. Language socialization into academic discourse communities. Annu Rev Appl Linguist. 2010;30:169\u0026ndash;92. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1017/S0267190510000048\u003c/span\u003e\u003cspan address=\"10.1017/S0267190510000048\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePun J, Fu BX, Cheung KKC. Language challenges and coping strategies in English Medium Instruction (EMI) science classrooms: A critical review of literature. Stud Sci Educ. 2023;60(1):1\u0026ndash;32. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1080/03057267.2023.2188704\u003c/span\u003e\u003cspan address=\"10.1080/03057267.2023.2188704\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXie W, Curle S. Success in English Medium Instruction in China: Significant indicators and implications. Int J Biling Educ Biling. 2022;25(2):585\u0026ndash;97. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1080/13670050.2019.1703898\u003c/span\u003e\u003cspan address=\"10.1080/13670050.2019.1703898\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDeci EL, Ryan RM. The What and Why of goal pursuits: Human needs and the self-determination of behavior. Psychol Inq. 2000;11(4):227\u0026ndash;68. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1207/S15327965PLI1104_01\u003c/span\u003e\u003cspan address=\"10.1207/S15327965PLI1104_01\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKohnke L, Moorhouse BL, Zou D. ChatGPT for language teaching and learning. RELC J. 2023;54(2):537\u0026ndash;53. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1177/00336882231162868\u003c/span\u003e\u003cspan address=\"10.1177/00336882231162868\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang C, Meng Y, Ma X. Artificial intelligence in EFL speaking: Impact on enjoyment, anxiety, and willingness to communicate. System. 2024;121:103259. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.system.2024.103259\u003c/span\u003e\u003cspan address=\"10.1016/j.system.2024.103259\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKalsoom A, Soomro NH, Pathan ZH. How social support and foreign language anxiety impact willingness to communicate in English in an EFL classroom. Int J Engl Linguist. 2020;10(2):80\u0026ndash;91. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.5539/ijel.v10n2p80\u003c/span\u003e\u003cspan address=\"10.5539/ijel.v10n2p80\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMa M, Noordin N, Razali AB. Effects of an AI Chatbot Mobile Application on Foreign Language Anxiety among Chinese EFL Undergraduates. \u003cem\u003eInt J Acad Res Prog Educ Dev\u003c/em\u003e. 2024;13(4). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://ijarped.com/index.php/journal/article/view/3301\u003c/span\u003e\u003cspan address=\"https://ijarped.com/index.php/journal/article/view/3301\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Accessed 24 Feb 2026.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTeimouri Y, Goetze J, Plonsky L. Second language anxiety and achievement: A meta-analysis. Stud Second Lang Acquis. 2019;41(2):363\u0026ndash;87. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1017/S0272263118000311\u003c/span\u003e\u003cspan address=\"10.1017/S0272263118000311\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDewaele J-M. The effect of classroom emotions, attitudes toward English, and teacher behavior on willingness to communicate among English foreign language learners. J Lang Soc Psychol. 2019;38(4):523\u0026ndash;35. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1177/0261927X19864996\u003c/span\u003e\u003cspan address=\"10.1177/0261927X19864996\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAlpert R, Haber RN. Anxiety in academic achievement situations. J Abnorm Soc Psychol. 1960;61(2):207\u0026ndash;15. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1037/h0045464\u003c/span\u003e\u003cspan address=\"10.1037/h0045464\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePekrun R. The control-value theory of achievement emotions: Assumptions, corollaries, and implications for educational research and practice. Educ Psychol Rev. 2006;18(4):315\u0026ndash;41. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s10648-006-9029-9\u003c/span\u003e\u003cspan address=\"10.1007/s10648-006-9029-9\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWu C-P, Lin H-J. Anxiety about speaking a foreign language as a mediator of the relation between motivation and willingness to communicate. Percept Mot Skills. 2014;119(3):785\u0026ndash;98. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.2466/22.PMS.119c32z7\u003c/span\u003e\u003cspan address=\"10.2466/22.PMS.119c32z7\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAlmukhaild H, King J. Scaling the apex of engagement: Anxious language learners' positive flow experiences during speaking tasks. System. 2024;124:103373. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.system.2024.103373\u003c/span\u003e\u003cspan address=\"10.1016/j.system.2024.103373\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBonanno GA. Loss, trauma, and human resilience: Have we underestimated the human capacity to thrive after extremely aversive events? Am Psychol. 2004;59(1):20\u0026ndash;8. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1037/0003-066X.59.1.20\u003c/span\u003e\u003cspan address=\"10.1037/0003-066X.59.1.20\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLuthar SS, Cicchetti D, Becker B. The construct of resilience: A critical evaluation and guidelines for future work. Child Dev. 2000;71(3):543\u0026ndash;62. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1111/1467-8624.00164\u003c/span\u003e\u003cspan address=\"10.1111/1467-8624.00164\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZimet GD, Dahlem NW, Zimet SG, Farley GK. The multidimensional scale of perceived social support. J Pers Assess. 1988;52(1):30\u0026ndash;41. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1207/s15327752jpa5201_2\u003c/span\u003e\u003cspan address=\"10.1207/s15327752jpa5201_2\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHouse JS. Work stress and social support. Addison-Wesley; 1981.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHwang KK. Face and favor: The Chinese power game. Am J Sociol. 1987;92(4):944\u0026ndash;74. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1086/228588\u003c/span\u003e\u003cspan address=\"10.1086/228588\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYoung DJ. Creating a low-anxiety classroom environment: What does language anxiety research suggest? Mod Lang J. 1991;75(4):426\u0026ndash;39. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.2307/329492\u003c/span\u003e\u003cspan address=\"10.2307/329492\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNoels KA, Pelletier LG, Clement R, Vallerand RJ. Why are you learning a second language? Motivational orientations and self-determination theory. Lang Learn. 2000;50(1):57\u0026ndash;85. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1111/0023-8333.00111\u003c/span\u003e\u003cspan address=\"10.1111/0023-8333.00111\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShort J, Williams E, Christie B. The social psychology of telecommunications. Wiley; 1976.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAbdelsaheed BSM, Alsalamah LA. ChatGPT as a tool for reducing anxiety and enhancing presentation skills: Perspectives of Saudi EFL students. Int J Instr. 2026;19(1):719\u0026ndash;48. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.29333/iji.2026.19137a\u003c/span\u003e\u003cspan address=\"10.29333/iji.2026.19137a\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFathi J, Rahimi M, Derakhshan A, Improving. EFL learners' speaking skills and willingness to communicate via artificial intelligence-mediated interactions. System. 2024;121:103254. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.system.2024.103254\u003c/span\u003e\u003cspan address=\"10.1016/j.system.2024.103254\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCohen S, Wills TA. Stress, social support, and the buffering hypothesis. Psychol Bull. 1985;98(2):310\u0026ndash;57. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1037/0033-2909.98.2.310\u003c/span\u003e\u003cspan address=\"10.1037/0033-2909.98.2.310\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen A, Jia J, Li Y, Fu L. Investigating the effect of role-play activity with GenAI agent on EFL students' speaking performance. J Educ Comput Res. 2025;63(1):99\u0026ndash;125. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1177/07356331241299058\u003c/span\u003e\u003cspan address=\"10.1177/07356331241299058\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKline RB. Principles and practice of structural equation modeling. 5th ed. Guilford; 2023.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHair JF, Black WC, Babin BJ, Anderson RE. Multivariate data analysis. 7th ed. Pearson; 2010.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGambino A, Fox J, Ratan RA. Building a stronger CASA: Extending the computers are social actors paradigm. Hum Mach Commun. 2020;1:71\u0026ndash;85. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.30658/hmc.1.5\u003c/span\u003e\u003cspan address=\"10.30658/hmc.1.5\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWaytz A, Cacioppo J, Epley N. Who sees human? The stability and importance of individual differences in anthropomorphism. Perspect Psychol Sci. 2010;5(3):219\u0026ndash;32. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1177/1745691610369336\u003c/span\u003e\u003cspan address=\"10.1177/1745691610369336\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKenny DA, Kaniskan B, McCoach DB. The performance of RMSEA in models with small degrees of freedom. Sociol Methods Res. 2015;44(3):486\u0026ndash;507. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1177/0049124114543236\u003c/span\u003e\u003cspan address=\"10.1177/0049124114543236\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003evan Driel OP. On various causes of improper solutions in maximum likelihood factor analysis. Psychometrika. 1978;43(2):225\u0026ndash;43. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/BF02293865\u003c/span\u003e\u003cspan address=\"10.1007/BF02293865\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKolenikov S, Bollen KA. Testing negative error variances: Is a Heywood case a symptom of misspecification? Sociol Methods Res. 2012;41(1):124\u0026ndash;67. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1177/0049124112442138\u003c/span\u003e\u003cspan address=\"10.1177/0049124112442138\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHayes AF. Introduction to mediation, moderation, and conditional process analysis: A regression-based approach. 3rd ed. Guilford Press; 2022.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKaiser HF, Rice J, Little Jiffy MIV. Educ Psychol Meas. 1974;34(1):111\u0026ndash;7. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1177/001316447403400115\u003c/span\u003e\u003cspan address=\"10.1177/001316447403400115\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePodsakoff PM, MacKenzie SB, Lee JY, Podsakoff NP. Common method biases in behavioral research. J Appl Psychol. 2003;88(5):879\u0026ndash;903. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1037/0021-9010.88.5.879\u003c/span\u003e\u003cspan address=\"10.1037/0021-9010.88.5.879\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBaumgartner H, Homburg C. Applications of structural equation modeling in marketing and consumer research: A review. Int J Res Mark. 1996;13(2):139\u0026ndash;61. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/0167-8116(95)00038-0\u003c/span\u003e\u003cspan address=\"10.1016/0167-8116(95)00038-0\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDoll WJ, Xia W, Torkzadeh G. A confirmatory factor analysis of the end-user computing satisfaction instrument. MIS Q. 1994;18(4):453\u0026ndash;61. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.2307/249524\u003c/span\u003e\u003cspan address=\"10.2307/249524\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHenseler J, Ringle CM, Sarstedt M. A new criterion for assessing discriminant validity in variance-based structural equation modeling. J Acad Mark Sci. 2015;43(1):115\u0026ndash;35. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s11747-014-0403-8\u003c/span\u003e\u003cspan address=\"10.1007/s11747-014-0403-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePaternoster R, Brame R, Mazerolle P, Piquero A. Using the correct statistical test for the equality of regression coefficients. Criminology. 1998;36(4):859\u0026ndash;66. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1111/j.1745-9125.1998.tb01268.x\u003c/span\u003e\u003cspan address=\"10.1111/j.1745-9125.1998.tb01268.x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLakens D, Scheel AM, Isager PM. Equivalence testing for psychological research: A tutorial. Adv Methods Pract Psychol Sci. 2018;1(2):259\u0026ndash;69. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1177/2515245918770963\u003c/span\u003e\u003cspan address=\"10.1177/2515245918770963\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePlonsky L, Oswald FL. How big is big? Interpreting effect sizes in L2 research. Lang Learn. 2014;64(4):878\u0026ndash;912. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1111/lang.12079\u003c/span\u003e\u003cspan address=\"10.1111/lang.12079\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCheung GW, Rensvold RB. Evaluating goodness-of-fit indexes for testing measurement invariance. Struct Equ Model. 2002;9(2):233\u0026ndash;55. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1207/S15328007SEM0902_5\u003c/span\u003e\u003cspan address=\"10.1207/S15328007SEM0902_5\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-psychology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"psyo","sideBox":"Learn more about [BMC Psychology](http://bmcpsychology.biomedcentral.com/)","snPcode":"","submissionUrl":"","title":"BMC Psychology","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"AI partner support, Perceived Teacher Support, L2 willingness to communicate, English language anxiety, Emotion-behavior decoupling, STEM learners","lastPublishedDoi":"10.21203/rs.3.rs-9578013/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9578013/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eEven though generative artificial intelligence is transforming language teaching worldwide, little research has investigated how Perceived Teacher Support (PTS) and AI Partner Support (AIPS) jointly influence learners' Second Language Willingness to Communicate (L2 WTC) and affective states.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eA cross-sectional survey was administered to 249 EFL learners at a STEM-focused university in China. Hierarchical regression, confirmatory factor analysis, exploratory moderation analysis, and multi-group analysis were conducted to test the proposed hypotheses.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eFindings are reported at two levels. At the confirmatory level, the additive positive effect of PTS and AIPS on L2 WTC was consistently supported across all analytical stages (Δ\u003cem\u003eR\u0026sup2;\u003c/em\u003e = .317), establishing the dual-pathway human-AI support structure as a robust and replicable predictor of communicative intention within STEM learning ecologies. At the exploratory level, ELA did not emerge as a significant inhibitor of L2 WTC, yet yielded a positive bivariate correlation (\u003cem\u003er\u003c/em\u003e = .298). This dual-level anomaly departs from classical SLA predictions. An exploratory moderation analysis provided initial evidence against functional substitution as the primary driver, lending indirect support to a suppressor variable interpretation; longitudinal designs are needed to adjudicate among competing accounts. Exploratory multi-group analysis (MGA) detected a directional cross-disciplinary signal that did not reach conventional significance (bootstrap \u003cem\u003ep\u003c/em\u003e = .088; post-hoc power\u0026thinsp;=\u0026thinsp;62%), and is therefore treated as hypothesis-generating rather than confirmatory evidence of discipline-specific differences.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e \u003cp\u003eThese findings represent preliminary empirical evidence whose generalizability requires replication across more diverse institutional and demographic contexts.\u003c/p\u003e","manuscriptTitle":"Anxious but Active: The Additive Effects of Generative AI and Teacher Support on L2 Willingness to Communicate","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-05-15 19:03:35","doi":"10.21203/rs.3.rs-9578013/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"reviewersInvited","content":"","date":"2026-05-06T12:05:06+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2026-05-06T02:22:40+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-05-04T11:02:01+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-05-04T11:01:37+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Psychology","date":"2026-04-30T13:36:21+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-psychology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"psyo","sideBox":"Learn more about [BMC Psychology](http://bmcpsychology.biomedcentral.com/)","snPcode":"","submissionUrl":"","title":"BMC Psychology","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"a0f0159e-1de0-48db-8ddd-4a7b2d6c78dc","owner":[],"postedDate":"May 15th, 2026","published":true,"recentEditorialEvents":[{"type":"reviewersInvited","content":"6","date":"2026-05-06T12:05:06+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2026-05-06T02:22:40+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-05-04T11:02:01+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-05-04T11:01:37+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Psychology","date":"2026-04-30T13:36:21+00:00","index":"","fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-05-15T19:03:35+00:00","versionOfRecord":[],"versionCreatedAt":"2026-05-15 19:03:35","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9578013","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9578013","identity":"rs-9578013","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}