Exploring EFL learners’ behavioral intention toward AI-based conversational tools through the lens of UTAUT and Self-Determination Theory Authors: Li Pan, Qihuan Chen Published: 2026 ABSTRACT AbstractAI-based conversational tools have attracted increasing attention in EFL speaking practice because they offer real-time interaction and flexible opportunities for oral communication, yet learners’ intention to use these tools remains insufficiently understood. Drawing primarily on the Unified Theory of Acceptance and Use of Technology, with Self-Determination Theory as a complementary perspective, this study examines factors associated with EFL learners’ behavioral intention to use AI-based conversational tools. A cross-sectional survey was conducted among 568 Chinese undergraduate EFL majors, most of whom were early-stage, low-frequency users of AI-based conversational tools. Structural equation modeling showed that behavioral intention was mainly associated with performance expectancy, hedonic motivation, and effort expectancy, while social influence and facilitating conditions showed weaker associations. The analysis identified significant indirect associations involving autonomy or competence and intrinsic motivation, whereas the corresponding association involving relatedness was not significant. Taken together, these findings suggest that learners’ intention to use AI-based conversational tools is related both to how they evaluate the technology and to the motivational quality of the speaking practice it supports. In practice, teachers should organize purposeful speaking activities that offer meaningful learner choice and make progress visible, while developers should ensure that AI-based conversational tools are useful, enjoyable, and easy to use.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s40359-026-05059-3.Keywords: AI-based Conversational Tools, Behavioral Intention, EFL, SDT, UTAUT Introduction AI-based conversational tools have become increasingly prevalent in language learning, particularly for speaking practice in English as a Foreign Language (EFL) setting [71]. Such tools, including Call Annie, Duolingo’s AI Video Call, and ELSA Speak, can provide around-the-clock speaking partners, immediate feedback, and emotional support, creating low-stakes opportunities for language output and communicative practice [67]. Compared with general-purpose text-based AI chatbots such as ChatGPT, these tools are designed more specifically for language-learning conversations, offering real-time speech or video interaction, simulated dialogue scenarios, and speaking feedback [18, 67]. However, despite their growing use and unique features for language learning, relatively little research has examined EFL learners’ behavioral intention (BI) to use these AI-based conversational tools. The Unified Theory of Acceptance and Use of Technology (UTAUT) provides a widely applied framework for examining how performance expectancy (PE), effort expectancy (EE), hedonic motivation (HM), social influence (SI) and facilitating conditions (FC) relate to learners’ intention to use AI-based conversational tools [66]. In the context of AI-supported speaking practice, where conversational tools are used to pursue speaking-development goals, BI may also be associated with learners’ intrinsic motivation (IM) to engage in English-speaking practice. Given that speaking development requires active and sustained oral engagement, learners with stronger IM may be more inclined to seek additional practice opportunities, which may correspond with a stronger intention to use AI-based conversational tools [30, 37]. To explain the psychological basis of this learning-related IM, the present study draws on Self-Determination Theory (SDT). According to SDT, satisfaction of autonomy, competence, and relatedness during AI-supported speaking practice may be associated with learners experiencing the learning process as more self-directed, effective, and connected [16]. These learning-oriented experiences may correspond with stronger IM, which may also be associated with learners’ intention to use the tools [69]. At present, research considering UTAUT alongside SDT remains limited. In technology-supported language learning, Hsu [28] and Luo and Day [40] represent two notable examples, with SDT-related constructs mainly examined as direct predictors of BI or as moderators of the relationships between UTAUT constructs and BI. Given that SDT links basic psychological need satisfaction with learning-related IM, the present study further explores whether PE, EE, and HM are linked to these learning-oriented experiences during AI-supported speaking practice. Unlike SI and FC, which reflect external expectations and support [65], PE, EE, and HM capture learners’ evaluations of the tool itself during practice and may be associated with their sense of control, progress, and connection in English-speaking development [26, 75, 78]. The present study therefore examines whether PE, EE, and HM show statistical indirect associations with BI involving learning-related psychological needs and IM, thereby offering an exploratory account of how tool evaluations and learning-oriented experiences may be connected in relation to BI. Therefore, with UTAUT as the primary framework and SDT as a complementary perspective, the present study primarily examines factors associated with EFL learners’ intention to use AI-based conversational tools. It also examines indirect associations between learners’ technology-related evaluations and BI involving basic psychological need satisfaction and IM during AI-supported speaking practice. Literature review Unified theory of acceptance and use of technology The Unified Theory of Acceptance and Use of Technology (UTAUT), proposed by Venkatesh et al. [65], is a widely used framework for examining factors associated with BI and technology use. Drawing on UTAUT, this study examines five constructs. Performance expectancy (PE), effort expectancy (EE), and hedonic motivation (HM) represent learners’ evaluations of the technology, whereas social influence (SI) and facilitating conditions (FC) represent influences arising from the surrounding social and support environment. Performance expectancy (PE) PE refers to individuals’ belief that using a system will help them improve job performance [65]. PE is particularly relevant to AI-based conversational tools because their real-time speaking practice and immediate feedback may support learners’ communicative fluency development [21]. Prior research has consistently found that PE is positively associated with BI in AI-supported learning contexts [5, 38]. When learners believe that a technology can support their performance and learning goals, they may regard its use as purposeful rather than of uncertain value [11]. This perceived instrumental value may be associated with a stronger sense that the time and effort involved in using the technology are worthwhile, which may correspond with stronger BI [20]. This consideration may be especially salient in AI-supported speaking practice because repeated oral production requires sustained effort, making expected gains in fluency, accuracy, or communicative effectiveness particularly relevant to learners’ evaluations of the tool [18]. Therefore, this study proposes the following hypothesis. H1: PE is significantly associated with BI. H1: PE is significantly associated with BI. Effort expectancy (EE) EE refers to the degree of ease associated with the use of the system [65]. Easy-to-use tools may impose lower cognitive and operational demands, potentially enabling learners to devote greater attention to language expression during real-time speaking interactions [34]. Prior research has often shown that EE is positively associated with BI in AI-supported learning contexts [5, 38]. When learners expect that initiating and managing speaking interactions will require relatively little time, effort, and technical input, the anticipated burden of using the tools is reduced [73]. This more favorable evaluation of the effort required relative to the expected speaking benefits may make tool use appear more worthwhile, which is associated with stronger BI [36]. This consideration may be especially relevant for early-stage and low-frequency users, whose limited familiarity with the tools may make anticipated effort a more salient factor in their evaluations of whether to use them for speaking practice. Therefore, this study proposes the following hypothesis. H2: EE is significantly associated with BI. H2: EE is significantly associated with BI. Hedonic motivation (HM) HM refers to the fun or pleasure derived from using a technology [66]. AI-based conversational tools can create enjoyable use experiences through lively dialogue, playful interaction, and encouraging responses [67, 76]. Extensive research in AI-supported learning contexts has shown that HM is positively associated with BI [38, 52, 77]. These enjoyable experiences may foster a more favorable affective evaluation of using the tools, making speaking practice feel more appealing rather than merely functional [74]. Lively, playful, and encouraging exchanges may make repeated oral production feel less monotonous and more emotionally rewarding [67]. These experiences may strengthen learners’ favorable affective evaluations of AI-based conversational tools, making them more attractive for speaking practice and helping to explain their association with stronger BI [79]. Therefore, this study proposes the following hypothesis. H3: HM is significantly associated with BI. H3: HM is significantly associated with BI. Social influence (SI) SI refers to the degree to which individuals perceive that important others believe they should use a particular system or technology [65]. As AI-based conversational tools are still emerging in EFL speaking practice, early-stage learners may rely on teachers and peers for social cues when judging whether these tools are suitable for learning [5, 67]. Several empirical studies have shown that SI is positively associated with learners’ BI in AI-supported learning contexts [5, 52]. Positive endorsement from teachers and peers may signal that using AI-based learning tools is socially recognized and educationally legitimate [8]. By making an unfamiliar form of speaking practice appear more credible and appropriate, such validation may be associated with stronger BI [32]. Therefore, this study proposes the following hypothesis. H4: SI is significantly associated with BI. H4: SI is significantly associated with BI. Facilitating conditions (FC) FC refers to the degree to which an individual believes that an organizational and technical infrastructure exists to support the use of a system or technology [65]. In AI-supported EFL speaking practice, stable technical conditions and timely support may reduce practical barriers to using conversational tools [32]. In AI-supported learning contexts, some studies have found a significant association between FC and BI [38, 52], whereas others have reported a non-significant association [5, 77]. This inconsistency may reflect differences in support availability, as FC may be less relevant to BI when support is readily accessible and more salient when access or assistance is uncertain [61]. In addition, FC may also be particularly relevant to BI among early-stage and low-frequency users, who may depend more heavily on external support because of their limited experience [41]. Therefore, this study proposes the following hypothesis. H5: FC is significantly associated with BI. H5: FC is significantly associated with BI. Self-Determination Theory Self-Determination Theory (SDT), proposed by Deci and Ryan [15], emphasizes the quality of motivation, particularly intrinsic motivation (IM), rather than behavior driven mainly by external rewards or pressure. It posits that individuals have three basic psychological needs: autonomy, competence, and relatedness. When these needs are satisfied, learners are more likely to develop higher levels of IM and experience learning activities as more self-endorsed and personally meaningful [16]. Relatedness (RE) Relatedness refers to individuals’ need to feel connected to others, to care for others, and to be cared for by others [16, 54]. In the present study, relatedness refers to learners’ sense of connection, support, and belonging in language learning as experienced during AI-supported speaking practice. From an SDT perspective, relatedness may provide a secure and supportive relational context in which IM can flourish [54]. Prior studies in technology-supported EFL learning contexts have also suggested that relatedness can be positively associated with IM [3, 38]. However, these prior findings are based on human-supported learning contexts, whereas in AI-supported speaking practice, relatedness may arise primarily from AI-mediated interaction. Although AI-based conversational tools can simulate dialogue and provide responsive interaction, they may not offer the emotional reciprocity, continuity, and interpersonal depth usually found in human–human communication [58]. Accordingly, the association between relatedness and IM in AI-supported speaking practice is expected to be context-dependent and exploratory. Therefore, this study proposes the following hypothesis. H6: Relatedness is significantly associated with IM. H6: Relatedness is significantly associated with IM. Competence (CO) Competence refers to individuals’ need to feel effective in their actions, experience mastery, and develop their capacities [16, 54]. In the present study, competence refers to learners’ sense that they can learn English effectively, manage speaking tasks, and make progress during AI-supported speaking practice. From an SDT perspective, satisfaction of the need for competence supports IM by fostering feelings of effectiveness and mastery [54]. Prior studies in technology-supported EFL learning contexts have also suggested that competence is positively associated with IM [19, 57]. In AI-supported speaking practice, repeated interaction and immediate feedback may make the connection between learners’ effort, successful task completion, and speaking progress more visible [18]. When learners interpret these outcomes as evidence of their own capability, speaking practice may feel more rewarding and inherently satisfying rather than merely externally required, providing a basis for stronger learning-related IM [17, 46]. Therefore, this study proposes the following hypothesis: H7: Competence is significantly associated with IM. H7: Competence is significantly associated with IM. Autonomy (AU) Autonomy refers to individuals’ need to experience volition and self-endorsement in their actions [16, 54]. In the present study, autonomy refers to learners’ sense of choice, control, and self-direction in engaging in AI-supported speaking practice in line with their English-learning goals, needs, and preferences. From an SDT perspective, autonomy supports IM by making engagement feel self-determined and inherently satisfying [54]. Prior research in technology-supported EFL learning contexts has suggested that autonomy is positively associated with learners’ IM [19, 43]. AI-supported speaking practice allows learners to adjust the topics, pace, and format of practice according to their own goals and preferences [67]. This alignment may make the activity feel more consistent with what they personally value and want to achieve. Such self-congruent engagement may make speaking practice more personally meaningful and inherently satisfying, thereby supporting learners’ IM for English-speaking development [45]. Therefore, this study proposes the following hypothesis: H8: Autonomy is significantly associated with IM. H8: Autonomy is significantly associated with IM. Intrinsic motivation (IM) IM refers to engaging in an activity because the activity itself is interesting, enjoyable, or inherently satisfying, rather than because of external rewards or pressures [16, 54]. In the present study, IM is treated as a learning-related motivational state, referring to learners’ enjoyment, perceived meaningfulness, and satisfaction associated with developing their English-speaking ability through AI-supported speaking practice. Unlike HM, which captures the immediate pleasure derived from using the AI tool, IM is directed toward the speaking-learning activity itself. Previous studies in in technology-supported EFL learning contexts showing that IM is positively associated with BI [29, 59]. This association may reflect the role of IM in making English-speaking development personally meaningful and satisfying. When learners value the learning activity itself, AI-supported practice may be viewed not merely as tool use but as a self-endorsed way of developing their speaking ability [45]. Accordingly, learners’ intrinsic interest in speaking development may be associated with stronger BI when the tool is viewed as supporting an activity they genuinely value. Therefore, this study proposes the following hypothesis: H9: IM is significantly associated with BI. H9: IM is significantly associated with BI. Relationships between selected UTAUT constructs and basic psychological needs Performance expectancy and competence EFL learners who perceive AI-based conversational tools as useful for improving speaking fluency, accuracy, or communicative effectiveness may also feel more effective and capable during speaking practice, as successful task experiences supported by these tools may provide positive information about their own speaking capability [46]. This interpretation may be more likely when learners attribute at least part of their successful performance to their own effort or ability, viewing it as evidence that they can learn effectively, manage speaking tasks, and make progress during AI-supported speaking practice [27, 56]. Previous research has similarly linked perceived usefulness and successful technology-supported learning experiences with greater learner confidence, stronger self-efficacy, and more positive competence beliefs [4, 44, 70]. Therefore, this study proposes the following hypothesis. H1a: PE is significantly associated with competence. H1a: PE is significantly associated with competence. Effort expectancy and autonomy In technology-supported learning, lower cognitive and operational demands may reduce the extent to which technical procedures constrain how learners manage their learning [39]. When AI-based conversational tools also provide flexible practice options, ease of use may help learners exercise these options and organize speaking practice according to their preferred pace, time, and learning goals [6]. When the content, pace, and mode of practice are largely determined by learners and experienced as self-endorsed, they may develop a stronger sense of choice, control, and personal ownership during speaking practice, which is consistent with greater autonomy [55, 64]. Some empirical studies have suggested that EE may be associated with autonomy. Pan and He [48] found that lower cognitive and operational barriers were associated with stronger autonomy in flipped classroom teaching. Similarly, Pan [47] reported that Chinese college English learners who perceived technology as easy to use were more willing to regulate their own learning. Therefore, this study proposes the following hypothesis: H2a: EE is significantly associated with autonomy. H2a: EE is significantly associated with autonomy. Hedonic motivation and relatedness Lively dialogue, playful interaction, and encouraging responses during AI-supported speaking practice may be associated with learners’ hedonic experience and sense of social presence [26]. Such experiences may make speaking practice feel less isolated and more supportive, as learners may perceive the AI tool as a responsive practice partner and develop a limited sense of companionship or connection [62]. However, this AI-mediated connection may represent only a limited form of relatedness because it does not fully provide the mutual understanding, emotional closeness, and stable relational bonds found in genuine interpersonal relationships [10]. As Al-Hoorie et al. [2] cautioned, prolonged reliance on AI-based conversation, particularly when it replaces authentic interaction with teachers and peers, may be associated with fewer opportunities for interpersonal communication, greater “social thinning,” and weaker relatedness need satisfaction. Therefore, the relationship between HM and relatedness may be context-dependent, with a positive association more likely when AI-mediated interaction supplements rather than replaces meaningful human interaction. This study proposes the following hypothesis: H3a: HM is significantly associated with relatedness. H3a: HM is significantly associated with relatedness. Hedonic motivation and autonomy In AI-supported speaking practice, encouraging responses, personalized interaction, and low-pressure design may contribute to more enjoyable and less psychologically demanding tool use [71, 74]. When the tool also provides flexible practice options, such enjoyable experiences may be associated with learners’ greater willingness to organize speaking practice around their own goals, needs, and preferences, which may correspond with a stronger sense of choice, control, and self-direction [72]. This self-directed participation may be experienced as a self-endorsed learning choice rather than as engagement driven mainly by external pressure, which is consistent with autonomy satisfaction [53, 55]. Thus, HM may be more closely related to autonomy when it is accompanied by meaningful opportunities to organize speaking practice according to learners’ own goals and preferences. Therefore, this study proposes the following hypothesis. Figure 1 presents the theoretical framework of this study. Conceptual framework H3b: HM is significantly associated with autonomy . H3b: HM is significantly associated with autonomy . Methods Research design This study employed a cross-sectional quantitative research design to examine factors associated with EFL learners’ behavioral intention to use AI-based conversational tools, with UTAUT as the primary framework and SDT as a complementary perspective. Purposive sampling was employed to recruit undergraduate English majors with prior experience using AI-based conversational tools. To reduce the potential for common method bias, several procedural remedies were implemented during questionnaire administration. Participants first read an electronic informed consent form explaining the study purpose, the anonymity and confidentiality of their responses, the voluntary nature of participation, and their right to withdraw at any time without penalty. They were also informed that there were no right or wrong answers and were encouraged to respond based on their own experiences and perceptions. In addition, items were not presented in strict construct order, and each page contained fewer than 10 items to reduce hypothesis guessing and response fatigue. A total of 652 questionnaires were collected from English majors with prior experience using AI-based conversational tools. After excluding 52 incomplete responses and 32 responses with unusually short completion times, 568 valid questionnaires were retained for analysis. Regarding demographic characteristics, 296 participants were male (52.1%) and 272 were female (47.9%). Most participants had limited experience with AI-based conversational tools: 192 learners (33.8%) had used such tools for less than one month, 187 (32.9%) for 1–2 months, 159 (28.0%) for 2–3 months, 16 (2.8%) for 3–6 months, and 14 (2.5%) for more than six months. Weekly usage was also relatively low, with 300 participants (52.8%) using these tools once a week or less, 192 (33.8%) using them 2–3 times per week, 38 (6.7%) using them 3–5 times per week, and 38 (6.7%) using them more than five times per week. Overall, the sample mainly consisted of early-stage and low-frequency users of AI-based conversational tools. Research instruments Data were collected through an online questionnaire administered via the WJX platform. The questionnaire consisted of four sections. The first section included screening questions, and only Chinese English majors with prior experience using AI-based conversational tools were eligible to continue. The second section presented an electronic informed consent form, which explained the study purpose and assured participants of confidentiality and data security; participants were required to select “I consent” before proceeding. The third section collected demographic and usage information, including gender, length of experience with AI-based conversational tools, and weekly usage frequency. The final section included measurement scales adapted from established instruments and modified to fit the context of AI-based conversational tool use for EFL speaking. These scales included UTAUT-related constructs adapted from Tang et al. [60]: PE and BI were each measured with three items, while EE, SI, and FC were each measured with four items. HM was adapted from Venkatesh et al. [66] and included four items. The SDT-related constructs were adapted from Hsu et al. [29], with four items for AU, four for CO, three for RE, and three for IM. Because the original scale items were in English, they were translated into Chinese by a professional translation agency and independently back-translated into English by another agency. Three doctoral-level experts, two from English-related fields and one from educational technology, reviewed the original, translated, and back-translated versions and adjusted them to fit the Chinese EFL context for AI-based conversational tool use while preserving the original construct meanings. The final version was confirmed through expert consensus after online discussion. Prior to the main data collection, a pilot test was conducted with 50 participants to examine item clarity and preliminary reliability, with all Cronbach’s alpha coefficients exceeding 0.70 and indicating acceptable internal consistency. The adapted measurement scales were therefore retained for the formal survey, while their reliability and validity were further assessed in the main analysis. The full items for the measures are provided in the Supplementary Materials (Table S1). Data analysis This study employed Partial Least Squares Structural Equation Modeling (PLS-SEM) to assess construct reliability and validity, test the hypothesized associations, and estimate specific indirect associations. PLS-SEM was selected because the proposed model extends an established technology-acceptance framework with several learning-oriented motivational constructs and places emphasis on explaining variance and prediction. SmartPLS 4.1 was used to evaluate the measurement model and estimate the structural associations, with bootstrapping based on 5,000 resamples used to assess the significance of the direct and indirect estimates. Results Common method bias was assessed using Harman’s single-factor test and the full collinearity approach. Harman’s test showed that the first unrotated factor explained 24.842% of the total variance, below the commonly used threshold of 40% [49]. The full collinearity VIF values ranged from 1.662 to 2.454, also below the recommended threshold of 3.3 [35]. These results suggest that common method bias is unlikely to pose a serious threat to the findings, although it cannot be completely ruled out given the same-source, same-time, self-report nature of the data. Multivariate outliers were assessed using the Mahalanobis distance based on the latent variable scores. Fourteen cases (2.46%) exceeded the recommended chi-square threshold (p < 0.001). These cases were retained because they showed no signs of invalid responding and likely reflect genuine variability among learners’ responses rather than data errors. Consequently, all 568 observations were retained for subsequent analyses. As shown in Table 1, Cronbach’s alpha values for all constructs ranged from 0.765 to 0.909, exceeding the 0.70 threshold and indicating acceptable internal consistency. Convergent validity was evaluated using factor loadings, composite reliability (CR), and average variance extracted (AVE). All item loadings ranged from 0.768 to 0.931, exceeding the cutoff of 0.70. The CR values ranged from 0.864 to 0.938, above the 0.70 benchmark, while the AVE values ranged from 0.619 to 0.834, above the 0.50 criterion [23]. Taken together, these indices indicate satisfactory convergent validity. Table 1Convergent validityConstructsItemsFactor LoadingAlphaCRAVEPEPE10.8140.7650.8640.680PE20.845PE30.814EEEE10.8550.8940.9260.759EE20.875EE30.884EE40.871SISI10.8930.9020.9310.772SI20.871SI30.888SI40.864FCFC10.8810.9090.9360.785FC20.894FC30.896FC40.874HMHM10.8690.8290.8980.745HM20.840HM30.881RERE10.9100.9030.9380.834RE20.931RE30.898COCO10.7710.7950.8670.619CO20.810CO30.798CO40.768AUAU10.8260.8450.8960.682AU20.847AU30.811AU40.820IMIM10.8900.8630.9160.785IM20.879IM30.888BIBI10.8880.8600.9150.782BI20.882BI30.882 Convergent validity Discriminant validity was assessed using the Heterotrait-Monotrait Ratio of Correlations (HTMT). Table 2 reports the HTMT values, which ranged from 0.029 to 0.792 and fell below the commonly recommended threshold of 0.80 [25], thereby providing further evidence that the constructs are empirically distinct. Table 2HTMTPEEESIFCHMRECOAUIMBIPEEE0.265SI0.3470.147FC0.2240.0390.118HM0.4650.2160.2280.119RE0.1180.1130.1190.0290.212CO0.6150.1950.2420.1250.3070.116AU0.5020.2970.2720.0600.5790.2200.395IM0.4450.2270.2550.0660.3290.1390.4550.537BI0.7920.3770.3510.2380.5820.1980.5640.5650.536 HTMT Before assessing the structural relationships, collinearity among the predictor constructs was examined using VIF values. All inner VIF values ranged from 1.486 to 3.140 and were below the recommended threshold of 3.3 [23], indicating no multicollinearity issues. The structural relationships were examined using bootstrapping with 5,000 resamples. The results of the direct effects are presented in Table 3, while the indirect effects are summarized in Table 4. For clarity, the findings are reported in groups based on their theoretical roles. First, all UTAUT-related constructs were significantly associated with BI, with PE showing the strongest association (β = 0.418, p < 0.001), followed by HM (β = 0.232, p < 0.001) and EE (β = 0.146, p < 0.001), whereas FC (β = 0.085, p = 0.002) and SI (β = 0.072, p = 0.015) showed a comparatively weaker association. In terms of incremental explanatory contribution, PE had a medium effect on BI (f2 = 0.292), while HM (f2 = 0.101) and EE (f2 = 0.050) had small effects. FC (f2 = 0.016) and SI (f2 = 0.011) were statistically significant, but their effects on BI were negligible. Second, regarding the relationships between selected UTAUT constructs and basic psychological needs, PE was significantly associated with CO (β = 0.481, p < 0.001), and HM was significantly associated with RE (β = 0.192, p < 0.001). In addition, EE (β = 0.172, p < 0.001) and HM (β = 0.454, p < 0.001) were significantly associated with AU, with HM showing a particularly strong association. IM was significantly associated with BI (β = 0.196, p < 0.001), with a small incremental effect (f2 = 0.072). Third, within the SDT framework, CO (β = 0.254, p < 0.001) and AU (β = 0.371, p < 0.001) were significantly associated with IM, whereas RE (β = 0.028, p = 0.434) was not. Measurement invariance across gender was also assessed using the MICOM procedure, with the detailed results reported in Supplementary Table S2 and Table S3. Table 3Hypothesis testingHypothesisPathβSDtpConclusionH1PE → BI0.4180.02914.324< 0.001SupportedH1aPE → CO0.4810.03015.899< 0.001SupportedH2EE → BI0.1460.0294.979< 0.001SupportedH2aEE → AU0.1720.0364.789< 0.001SupportedH3HM → BI0.2320.0307.725< 0.001SupportedH3aHM → RE0.1920.0375.184< 0.001SupportedH3bHM → AU0.4540.03214.185< 0.001SupportedH4SI → BI0.0720.0302.4260.015SupportedH5FC → BI0.0850.0283.0370.002SupportedH6RE → IM0.0280.0350.7830.434Not SupportedH7CO → IM0.2540.0386.609< 0.001SupportedH8AU → IM0.3710.03610.305< 0.001SupportedH9IM → BI0.1960.0306.624< 0.001Supported Hypothesis testing Table 4Indirect effectsPathβSDtpPE → CO → IM → BI0.0240.0064.297< 0.001EE → AU → IM → BI0.0120.0043.4980.001HM → RE → IM → BI0.0010.0010.7440.457HM → AU → IM → BI0.0330.0065.244< 0.001 Indirect effects In the analysis of indirect associations, PE showed a significant indirect association with BI involving CO and IM (β = 0.024, p < 0.001). Similarly, EE (β = 0.012, p = 0.001) and HM (β = 0.033, p < 0.001) showed significant indirect associations with BI involving AU and IM. By contrast, the indirect association between HM and BI involving RE and IM was not significant (β = 0.001, p = 0.457). The R2 value for BI was 0.562, indicating that the model accounted for 56.2% of the variance in BI. The path results of PLS-SEM are shown in Fig. 2. PLS-SEM results PLSpredict analysis was conducted to assess the out-of-sample predictive power of the model [22]. As shown in Table 5, all Q2predict values were above zero, indicating predictive relevance. The PLS-SEM model produced lower RMSE and MAE values than the linear benchmark for most indicators, including all three BI indicators. Although AU1 and AU2 showed slightly higher prediction errors, the overall results support moderate predictive power for the model. Table 5PLSpredict results for out-of-sample predictive powerItemsQ2predictPLS-SEM_RMSEPLS-SEM_MAELM_RMSELM_MAEAU10.1940.6210.5120.6140.500AU20.2050.6030.4880.5960.483AU30.1790.6040.4810.6040.485AU40.1680.5840.4610.5850.465BI10.4100.5700.4740.5750.476BI20.3820.5900.4840.5990.493BI30.4150.6310.5000.6380.503CO10.1300.6150.4840.6200.484CO20.1510.6110.4950.6140.495CO30.1600.6050.5110.6100.511CO40.1210.6330.5150.6380.517IM10.1010.7390.6170.7440.626IM20.1280.7440.6190.7450.637IM30.1160.7530.6160.7590.625RE10.0130.9400.7790.9520.790RE20.0430.9040.7320.9090.744RE30.0180.9040.7310.9160.741 PLSpredict results for out-of-sample predictive power Discussion Discussion of UTAUT-related variables PE was positively associated with BI and showed the largest effect size among the examined predictors, consistent with previous studies [5, 38]. This finding suggests that learners’ intention to use AI-based conversational tools is closely aligned with the extent to which they perceive these tools as relevant and useful for their speaking goals [11, 21]. The strength of this association may indicate that learners’ BI is largely goal-directed, reflecting their expectation that using these tools is relevant to their speaking development. PE also showed a strong positive association with competence. This pattern may suggest that perceptions of usefulness and competence can share a basis in successful speaking experiences, as such experiences may inform both judgments of the tool’s value and learners’ perceptions of their own capability [17, 46]. Successful experiences may be particularly relevant to competence when learners interpret at least part of their performance as reflecting their own effort or ability rather than only the tool’s support. PE also showed a significant indirect association with BI involving competence and IM. This pattern places PE within a broader motivational configuration in which perceived usefulness, competence in speaking practice, intrinsic engagement, and BI are statistically interconnected. It suggests that the relevance of PE may extend beyond an instrumental evaluation of the tool and be embedded in learners’ broader experience of speaking development. Accordingly, clarifying the potential speaking benefits of different AI tool functions may be a useful approach. Developers could explain how specific practice modes and feedback are intended to support fluency, pronunciation, accuracy, or conversational interaction, and provide brief examples illustrating the learning purpose of each function. HM was positively associated with BI and showed the second-largest incremental explanatory contribution among the examined predictors, consistent with previous studies [52, 77]. The comparatively stronger association may indicate that, in optional and repetitive speaking practice, learners consider not only what the tool may achieve but also how the interaction feels [50]. Lively, encouraging, and low-pressure exchanges may be associated with repeated oral production being experienced as less monotonous and more emotionally rewarding, although such affective appeal should not be equated with actual speaking improvement [71]. HM was also positively associated with relatedness. Enjoyable and responsive AI interaction may be associated with speaking practice feeling less isolated and more supportive, as well as with learners’ sense of connection during language learning [62]. However, this AI-mediated relatedness may remain limited and context-dependent because it may not provide the same level of mutual understanding and belonging as interaction with teachers and peers. Although HM was positively associated with relatedness, the indirect association between HM and BI involving relatedness and IM was not significant. This pattern appears to reflect the non-significant association between relatedness and IM rather than the HM-relatedness link, possibly because the limited connection available in AI interaction was not closely aligned with learning-oriented intrinsic motivation. Furthermore, HM was strongly and positively associated with autonomy. This relationship may indicate that enjoyment and learning-related autonomy are closely intertwined in AI-supported speaking practice, as positive interaction may coincide with learners engaging more willingly and organizing practice around their own speaking goals and preferences. HM also showed a significant indirect association with BI involving autonomy and IM. This pattern places HM within a broader motivational configuration in which enjoyment, autonomy, intrinsic engagement, and BI are statistically interconnected. The relevance of HM may therefore extend beyond immediate affective appeal and be embedded in learners’ broader experience of self-directed speaking development. In practice, developers should make AI-supported speaking practice enjoyable without reducing it to entertainment. Varied scenarios, responsive dialogue, encouraging prompts, and low-pressure feedback may help sustain learners’ interest during repeated practice. These features should be integrated into purposeful speaking tasks that reflect learners’ goals and preferences, so that enjoyment remains connected to meaningful learning. EE was positively associated with BI but showed the smallest effect size among the technology-related evaluations, consistent with previous studies [5, 38]. The modest effect may suggest that ease of use was not a central basis for BI among these EFL learners, particularly as university students may already be familiar with emerging digital technologies [14, 42]. For early-stage users, smooth operation may still reduce uncertainty and make initial interactions more manageable; however, once basic usability is achieved, lower effort may do little to distinguish stronger from weaker BI [24]. Learners’ intention may therefore be more closely aligned with whether the tools provide clear and efficient support for speaking development than with ease of operation alone. EE was also positively associated with autonomy. This association may suggest that lower operational demands are linked to learners having greater practical control over how they organize speaking practice around their own goals and preferences. However, ease of use should not be equated with autonomy itself, as it may support autonomy more likely when learners are also given meaningful choices and can align tool use with personally endorsed speaking goals [33]. EE also showed a significant but small indirect association with BI involving autonomy and IM. Beyond its direct association with BI, EE was also embedded in a broader learning-oriented pattern involving autonomy and IM. This suggests that ease of use was relevant both as a technology evaluation and as part of learners’ wider experience of control and intrinsic engagement in speaking practice. In practice, the main priority for EE is to reduce friction in the core interaction process, particularly for first-time users. Developers should streamline onboarding, dialogue initiation, mode switching, turn-taking, and error recovery while avoiding unnecessary steps or feature overload. FC was positively associated with BI, but its effect size was negligible, consistent with previous studies [38, 52]. The significant association may partly reflect the present sample’s early-stage and low-frequency use, as unfamiliar functions or occasional problems may still make technical support relevant. However, the negligible effect may indicate that the tools were generally easy enough to use, leaving FC relevant mainly when learners encountered difficulties rather than as a central basis for BI [9]. As another contextual influence, SI was also positively associated with BI, and its effect size was similarly negligible, consistent with previous studies [5, 52]. The positive association may reflect that recommendations from teachers and peers signal that AI-based conversational tools are an acceptable form of speaking practice, particularly for learners with limited prior experience [18]. However, one possible explanation for the negligible effect is that this practice is often completed individually outside class and is not strongly embedded in formal or shared learning activities, limiting the relevance of social expectations to learners’ BI [68]. Taken together, contextual influences arising from learners’ social and support environments showed much weaker associations with BI than their evaluations of the technology itself. External support and social endorsement may therefore represent supplementary considerations for BI, while favorable evaluations of the tool and the learning experience appear to be more central [9, 63]. Discussion of SDT-related variables IM, as a key motivational outcome of basic psychological need satisfaction in SDT, was significantly associated with BI, consistent with previous studies [29, 59]. This finding suggests that learners’ BI may be linked not only to their evaluations of the technology but also to their intrinsic interest in speaking practice itself. From an SDT perspective, learners who experience speaking practice as enjoyable and satisfying may view AI-based conversational tools as opportunities to pursue an activity they already find intrinsically engaging [7]. In practice, greater attention should be paid to making the speaking activity itself interesting and satisfying rather than relying on the appeal of the technology. Developers could provide open-ended tasks involving the expression of personal viewpoints, communicative problem-solving, and story development, as these activities offer forms of speaking practice that may be experienced as rewarding in their own right. Among the three basic psychological needs, autonomy showed the strongest association with IM, consistent with previous studies [19, 43]. This pattern may reflect the self-directed nature of AI-supported speaking practice, which allows learners greater discretion over when, what, and how to practice in accordance with their own goals [64]. Such control may be associated with speaking practice being experienced as personally endorsed rather than externally imposed, which may help explain the stronger association between autonomy and IM. However, flexibility does not necessarily satisfy autonomy, because choices that are excessive or poorly aligned with learners’ goals may offer nominal control without genuine volition [33]. In practice, teachers and developers should prioritize meaningful rather than numerous choices by allowing learners to set personal goals and control key aspects of practice, such as topics, pace, feedback timing, and repetition. Rigid requirements, excessive monitoring, and superficial customization should be minimized because they may be inconsistent with learners’ sense of volition. Competence was positively associated with IM, consistent with previous studies [19, 57]. This finding supports the relevance of SDT to AI-supported EFL speaking practice, suggesting that learners’ feelings of effectiveness, mastery, and progress remain closely associated with intrinsically motivated learning in technology-mediated contexts [13]. Feeling capable of learning English effectively, managing speaking tasks, and making progress may be associated with experiencing AI-supported speaking practice as more rewarding and inherently satisfying [46]. This motivational experience may be particularly evident when learners interpret successful performance as evidence of their developing speaking ability. Accordingly, helping learners recognize their progress during AI-supported speaking practice may be a useful approach. Teachers could encourage learners to compare their initial and revised responses and reflect on how their effort and strategies relate to observed improvement, while developers could support side-by-side comparisons of recordings, transcripts, and feedback across repeated tasks. Relatedness was not significantly associated with IM, which differs from findings in some human–human language classroom settings [1, 12]. One possible explanation is that relatedness in AI-supported speaking practice may be experienced more in connection with AI-mediated interaction than with reciprocal relationships involving teachers or peers [13]. Although current AI-based conversational tools can provide immediate responses, encouragement, and a certain sense of social presence [31], human–machine interaction may still lack the reciprocal understanding, emotional closeness, and stable relational recognition that make relatedness motivationally powerful in human–human learning contexts [51]. Thus, learners may view AI-based conversational tools as responsive practice partners, but this limited sense of connection may not be sufficiently strong to show a significant association with learning-oriented IM. This distinction may also help explain why relatedness showed a different pattern from autonomy and competence. Autonomy and competence are more directly reflected in learners’ sense of ownership over their speaking practice and their perceived capability and progress, whereas relatedness depends more heavily on whether that practice is embedded in supportive social relationships with teachers and peers [13, 45]. Theoretical implications Several theoretical implications can be drawn from these findings. First, this study provides additional evidence on the application of UTAUT in AI-based conversational tool use by showing that its constructs are not equally associated with BI. Specifically, learners’ BI toward AI-based conversational tools is more strongly associated with technology-related evaluations, such as usefulness, ease of use, and enjoyment, than with external social or facilitating conditions. This finding suggests that, in early-stage use of AI-based conversational tools, UTAUT explanations of BI should pay particular attention to learners’ direct experience with the tool itself. Second, using SDT as a complementary perspective brings attention to learning-oriented psychological experiences in English-speaking practice, rather than only learners’ evaluations of the technology. The findings further suggest that selected technology-related evaluations (PE, EE, and HM) may show indirect associations with BI involving competence or autonomy and IM. Their relevance to BI may therefore extend beyond favorable judgments of the tool to whether its use is associated with learners feeling capable and self-directed in speaking development. This distinction helps separate technology acceptance from the motivational quality of the language-learning activity it supports. Third, although relatedness has long been considered an important factor in fostering IM, its non-significant role in this study does not imply that relatedness is unimportant in language learning. Rather, it suggests that relatedness in human–machine interaction should not be treated as equivalent to the social support and relational experience typically observed in human–human interaction. The nature of relatedness in AI-supported speaking practice, and the conditions under which it is associated with IM and BI, therefore warrant further theoretical and empirical investigation. Conclusion This study aimed to examine the factors associated with Chinese undergraduate EFL learners’ behavioral intention to use AI-based conversational tools, focusing particularly on early-stage, low-frequency users. The findings indicate that BI was more strongly associated with learners’ evaluations of the tool itself, particularly PE, followed by HM and EE, whereas SI and FC, which reflect social and support conditions surrounding tool use, showed comparatively weaker associations. In addition, learning-oriented IM was positively associated with BI, suggesting that learners’ intention to use AI-based conversational tools was also linked to their intrinsic interest and satisfaction in developing English-speaking ability. Selected technology-related evaluations (PE, EE, and HM) showed indirect associations with BI involving learning-oriented autonomy or competence and IM. This suggests that favorable evaluations of the tools may be linked to BI partly when their use is accompanied by stronger feelings of self-direction or capability in speaking development. By contrast, relatedness showed a limited role, suggesting that the relatively shallow connection provided by human–machine interaction may not yet become a strong source of learning-oriented motivation. In practice, teachers and developers should combine useful, enjoyable, and accessible tool design with speaking activities that support learners’ autonomy, competence, and intrinsic interest, rather than relying on favorable tool features or AI-mediated connection alone. Limitations Several limitations should be acknowledged when interpreting the findings of this study. First, although Harman’s single-factor test and the full collinearity assessment suggested no serious common method bias, this bias cannot be completely ruled out because the data were collected from the same respondents at a single time point using self-report questionnaires. The cross-sectional nature of the data also limits causal interpretations, as the direction and development of the relationships among UTAUT constructs, SDT needs, IM, and BI cannot be firmly established. Moreover, indirect effects estimated from cross-sectional data may overestimate or otherwise misrepresent mediation because temporal ordering, reverse associations, and omitted-variable explanations cannot be ruled out. Future research should adopt longitudinal or experimental designs and incorporate multi-source or behavioral log data to further validate the findings and examine causal mechanisms over time. Second, although the adapted scales showed acceptable reliability and validity, their psychometric structure may be partly context-bound because the items were modified for Chinese EFL learners’ use of AI-based conversational tools. Therefore, claims regarding content validity and generalizability should be interpreted cautiously, and future studies should further validate these measures across different learner groups, learning contexts, and AI-based conversational tools. Third, most participants in this study were Chinese English-major undergraduates with early-stage and low-frequency experience using AI-based conversational tools, so the findings may not fully apply to more experienced users, higher-frequency users, or learners from other contexts. Future studies should test the model with more diverse samples and usage stages. Supplementary Information Supplementary Material 1. Acknowledgements During the preparation of this work the author(s) used Grammarly and QuillBot to enhance the readability of the manuscript. After using this tool/service, the author(s) reviewed and edited the content as needed and take(s) full responsibility for the content of the published article. Authors’ contributions Li Pan: Conceptualization; Formal analysis; Methodology; Project administration; Software; Supervision; Visualization; Writing – original draft; Writing – review & editing Qihuan Chen: Data curation; Funding Acquisition; Investigation; Resources; Validation; Writing – review & editing. Funding This research received no external funding. Data availability In accordance with local regulations and the commitments that researchers have made to participants in the informed consent form, the researchers shall not disclose the dataset to the public. Nevertheless, the data can be made available from the corresponding author upon a reasonable request. Declarations Ethics and consent to participate The study was conducted in accordance with the Declaration of Helsinki and approved by Shandong Xiehe University Institutional Review Board (protocol code LLSC-KY03-2025027, October 21, 2025). Informed consent was obtained from all participants involved in the study. Consent for publication All participants provided consent for their anonymized data to be used in scientific publications. Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1. Data Availability Statement In accordance with local regulations and the commitments that researchers have made to participants in the informed consent form, the researchers shall not disclose the dataset to the public. Nevertheless, the data can be made available from the corresponding author upon a reasonable request.