Psychological constructs of AI speaking app adoption among Chinese tertiary learners: a mixed-methods investigation through the Technology Acceptance Model

2.1 Psychological factors in relation to AI acceptance

Several psychological factors have been well established as key concepts in foreign and second language learning, among which the constructs that are most studied and relevant to speaking exercises are perhaps WTC and SA. WTC, defined by MacIntyre et al. (1998), refers to learners’ readiness to engage in conversations using a foreign language, and its importance in EFL is due to the significance of interaction in language development. MacIntyre et al. (1998) proposed a pyramid model, in which WTC is conceptualized as learners’ behavioral intention that serves as a key antecedent of L2 communication behavior. Previous research on EFL learners at the tertiary level has indicated that individuals with higher WTC participate more actively in classroom speaking activities (Deng and Peng 2023; Ha 2011) and are more inclined to speak during tasks (Leeming et al. 2024). Additionally, it was found out by Huang and Zou (2024)’s study that WTC significantly influenced learners’ English speaking with AI, in dimensions of their perceived usefulness and satisfaction with AI-based speaking applications, which further affects their intentions to continue using these tools.

SA refers to the fear or anxiety associated with real or anticipated communication with others (McCroskey 1977). Research has demonstrated that SA negatively impacts language performance, particularly in speaking tasks (Hewitt and Stephenson 2012; Peng and Wang 2024). SA is also found to be negatively correlated with other psychological factors, such as intrinsic motivation (Ahmetovic et al. 2020) and WTC (Zarrinabadi et al. 2024), and this leads to learners’ decreased chances to participate in foreign language interactions. Research by Lee and Hsieh (2019) and Pavelescu (2023) found that anxiety hindered WTC in both in-class and out-of-class contexts. However, Lee and Hsieh (2019) noted that in digital settings, anxiety was not significantly related to WTC, explaining that digital environments may provide a less anxiety-provoking atmosphere for EFL students. Likewise, in Bárkányi’s (2021) study, the author indicated that learners with higher SA were more inclined to use mobile learning due to the anonymity it offers.

Moreover, a relatively less explored construct in foreign language learning, SPCC, also carries many potential implications for EFL learning. According to McCroskey and McCroskey (1988), SPCC refers to an individual’s self-assessment of their ability to effectively engage in communication activities. This concept has been applied to evaluating communication competence across various contexts, including interactions with acquaintances, friends, and strangers, as well as within small groups, public speaking scenarios, and interpersonal communication (Croucher et al. 2024). Research has shown that SPCC is associated with both Communication Apprehension (CA) and WTC; learners who perceive themselves as competent are more likely to engage in communication and experience lower levels of anxiety (Teven et al. 2010; Yashima 2002).

In general, the constructs reviewed above – WTC, SA, and SPCC – play a significant role in influencing EFL learners’ behaviors in speaking practice. Previous studies have mostly examined their interrelations (e.g., Liu and Jackson 2008; Nadeem et al. 2023) in formal English learning settings, and their predictions on speaking productions (e.g., Leeming et al. 2024; Peng and Wang 2024). Against the backdrop of increasing use of AI and the availability of mobile language learning resources, there is a need for further investigation into how these psychological constructs that play a role in English speaking practice can be associated with learners’ acceptance and use of AI speaking tools.

2.2 Technology Acceptance Model

The Technology Acceptance Model (TAM), originally introduced by Davis et al. (1989), is widely recognized as one of the most influential frameworks for understanding user acceptance of new technologies. The model outlines the factors that influence an individual’s behavioral intention, specifically identifying three key determinants: Perceived Usefulness (PU), Perceived Ease of Use (PEU) and Attitude Toward Using the System (ATUS). PU is understood as whether an individual believes that using a particular system would benefit their performance, while PEU indicates whether one believes that using a particular system would be free from efforts (Davis 1989). ATUS is conditioned by PU and PEU, exerting a direct influence on behavioral intention (BI). The specific relationship of the three constructs can be seen in Figure 1.

Figure 1:

Technology Acceptance Model (Davis 1989).

In the context of mobile learning in EFL, previous studies have incorporated TAM into their investigations, integrating psychological factors, such as motivation, self-efficacy, and anxiety caused by novel technologies, into the original TAM to examine AI adoption in the EFL context. For example, Wang et al. (2021) found that Chinese teachers’ self-efficacy of using novel technologies positively affected their PEU and ATUS in higher education. Zou et al. (2023c) revealed a correlation between the intrinsic motivation of Chinese tertiary-level students and their BI and PEU of an AI-based speech evaluation program. There are relatively few studies explicitly concentrating on English-speaking-related psychological factors in association with technological acceptance.

2.3 Learners’ perceptions of AI adoption in the EFL context

Previous studies have explored learners’ perceptions regarding the adoption of AI-driven applications for developing spoken English skills, involving both negative and positive attitudes. A primary concern revolves around the accuracy and reliability of AI-generated feedback, which may stem from limitations in speech recognition technology (Kholis 2021; Zou et al. 2023a, 2023c) and the inherent challenges in evaluating open-ended speaking topics (Litman et al. 2018; Zou et al. 2023c). Additionally, learners have reported that occasional inappropriate or unnatural responses from AI technologies can undermine their trust in these systems and lead to frustration (Kim and Kim 2020; Zou et al. 2023c). Furthermore, the simplistic feedback provided by AI educational tools is criticized for failing to identify specific issues or offer constructive corrections, which may result in confusion (Kim and Kim 2020; Zou et al. 2023a).

On the other hand, several notable benefits are recognized from the adoption of AI applications in education. AI tools offer flexibility, allowing learners to practice at their own pace and time, which is particularly advantageous for adult learners with busy schedules (Zhang and Pérez-Paredes 2021). Moreover, AI speaking apps can create a low-pressure environment for learners to practice speaking without the fear of making mistakes in front of others (Kim and Kim 2020; Zou et al. 2023c). This aspect can be especially beneficial for learners who experience anxiety or lack confidence when speaking English. Previous research has also corroborated the positive effects of AI technology on learners’ psychological factors, such as confidence, WTC, anxiety, and enjoyment (e.g., Fathi et al. 2024; Kim and Su 2024; Shafiee Rad 2024; Zhang et al. 2024), indicating that learners who utilize AI tools report increased willingness to communicate and confidence, as well as reduced anxiety in EFL speaking, attributable to the personalized and supportive nature of these tools.

Taken together, scant attention has been directed toward whether psychological factors predicted learners’ acceptance of AI, and few studies extended the TAM by adding psychological constructs (Hsu and Lin 2022). Focusing on AI speaking apps as an emergent technology with great potential to enhance learners’ English-speaking practice, we investigated two research questions (RQs) in this study:

1. How are the WTC, SA, and SPCC of China’s tertiary EFL learners associated with their acceptance of mobile AI speaking apps, indicated by BI, PU and PEU?

2. How do EFL students perceive the use of mobile AI speaking apps in practicing their spoken English?

3.1 Participants

In this study, we focused on Chinese tertiary EFL learners, a demographic increasingly engaging in mobile AI English speaking apps in China. An online questionnaire was administrated to students via Wenjuanxing (https://www.wjx.cn), an online survey distribution platform in China, generating participation of 337 students and ultimately yielding 280 valid responses. Informed consent was obtained with clarifications of participants’ rights. 98.9 % of the participants aged between 18 and 34, with only 1.1 % aged above. All participants were Chinese students who learned English as a second language. Their demographic data are summarized in Table 1. In terms of English language proficiency, all participants had at least passed the CET-4 exam, indicating that their vocabulary was at least 4,000 words and their English proficiency was at least B1-B2 level according to Common European Framework of Reference for Languages (CEFR). Additionally, 13.2 % of the participants had passed the TEM-8 exam, with a vocabulary base between 10,000 and 12,000 words, corresponding to the CEFR C1-C2 level.

3.2 Measures

3.3 Analytic approach

Prior to data analysis, 57 questionnaires with response time less than 2 mins or with highly repetitive values were excluded due to the concern that the respondents had not answered the questionnaire seriously. The data screening process resulted in 280 questionnaires that could be used for further analysis. SPSS was employed for data analysis. Exploratory factor analysis (EFA) and reliability test were conducted to examine the construct validity and reliability of the scales.

Then, we used logistic regression models to assess the respective relationship of the psychological constructs – WTC, SA and SPCC – and participants’ acceptance of mobile speaking apps indicated by BI, PU and PEU. In the logistic models, the log odds of the binary outcomes (BI, PU and PEU) were modeled as a linear combination of the predictor variables (WTC, SA and SPCC) and control variables of gender, academic year and English proficiency. We checked for multicollinearity using Variance Inflation Factor (VIF) to ensure the validity of the analysis. The Hosmer-Lemeshow test was also conducted to confirm the goodness of fit.

Finally, the interview data were audio-recorded, transcribed to computer by the first author and reviewed by the second author to ensure accuracy. Thematic analysis was used to extract themes emerging from the interview data.

4.1 Associations between students’ WTC, SA, and SPCC and their acceptance of mobile AI speaking apps

The descriptive results of the questionnaire response show that participants held a positive attitude towards AI speaking apps. Specifically, 80.4 % of participants expressed a behavioral intention to use mobile AI speaking apps in the future. Furthermore, 96.07 % perceived that using the apps was useful for improving their speaking skills, and 82.5 % found the apps easy to use.

To examine the predictive effects of psychological constructs on technological acceptance, we conducted binary logistic regression analyses with three separate models. Each model tested whether WTC, SA, and SPCC predicted one of the following dependent variables: behavioral intention to use, perceived usefulness, and perceived ease of use, while controlling for gender, academic year, and English proficiency (see Table 3).

The results revealed that WTC was significantly and positively related to behavioral intention (β = 1.22, p < 0.001), illustrating that for each 1 % increases in learner’s WTC, the log odds of BI increases by 1.22. Examining the odds ratio of this parameter estimate shows that for each 1 % increase in WTC, the odds of BI increases by a factor of 3.37. The Hosmer and Lemeshow’s test suggest that the model provides a good fit to the data.

In addition, Table 3 indicates that WTC is also significantly and positively related to PU (β = 1.34, p < 0.05), and the odds ratio of it shows that it increases the odds of perceiving the app to be useful by a factor of 3.81. Another significant predictor of PU is SPCC, which has a negative association (β = −1.27, p < 0.05) and increasing the odds of PU by a factor of 0.28. SA was not found significantly correlated with any construct of technological acceptance. No psychological construct that we concern in this study show an association with PEU.

4.2 Students’ perceptions of mobile AI speaking apps

To further explore participants’ perceptions of mobile AI speaking apps, semi-structured interviews were conducted with 8 participants, extracting two main themes revolving around the benefits and limitations of AI speaking apps.

5.1 WTC, SPCC, SA and technology acceptance

In the current study, the positive association between WTC and both BI and PU is illuminated by the qualitative account of AI speaking apps as “psychologically safe rehearsal spaces”. High-WTC learners in interviews emphasized how apps allowed them to bypass social judgment (e.g., S1’s comment on “nobody judges you”), which aligns with their heightened PU and BI. The results resonate with existing studies emphasizing the role of proactive communication tendencies in positive technology perception (e.g., Huang and Zou 2024) and that learners with higher WTC may view AI speaking apps as platforms to practice without social judgment, thereby amplifying their perceived utility – a pattern observed in gamified language apps (Lee and Drajati 2020). Our study extends the literature by revealing that WTC’s influence transcends PU to directly drive BI, a linkage less emphasized in previous work. This suggests that for AI tools requiring active interaction, WTC may serve as a gateway to both cognitive evaluation (PU) and decisional commitment (BI).

Contrary to previous studies (He and Li 2023; Yang and Lou 2024), which suggested that individuals with higher competence exhibit greater continuance intention or higher PU when utilizing mobile learning for English acquisition, we established that SPCC negatively predicted PU. This inverse relationship challenges the conventional “more competence, higher acceptance” assumption in TAM literature. A plausible explanation for this discrepancy lies in that previous studies examined competence in mobile language learning, whereas the present study focused more on the oral, communication-focused dimension. Moreover, learners with high SPCC, who already feel confident in face-to-face communication, may perceive AI apps as redundant or inferior to human interaction, thereby devaluing their usefulness. A “threshold effect” is thus observed here: for communication-focused AI speaking apps, learners with exceeding real-world competence (high SPCC) may perceive them as inadequate substitutes for human interaction, thereby suppressing PU. This aligns with “AI aversion” theories (e.g., Jussupow et al. 2022) but extends them to pedagogical contexts.

In addition, the lack of association between SA and technological acceptance found in the present study diverges from studies linking anxiety to technology avoidance (e.g., Barrot et al. 2021). This discrepancy may stem from the nature of interactions in the mobile learning setting. Unlike human-centered, face-to-face communication settings where anxiety directly impairs performance (Gregersen and MacIntyre 2014), AI apps may mitigate anxiety by offering non-judgmental, safe “rehearsal space” to learners – a feature highlighted by participants from our interview data. Alternatively, SA’s non-significance might also be due to measurement limitations; our binary acceptance measure (vs. graded scales in prior studies) could have obscured subtle anxiety effects on perceptions of usefulness or ease of use.

5.2 Toward human-AI synergy: pedagogical and technological implications

Our interview data revealed that mobile AI speaking apps could increase speaking practice frequency and reduce anxiety, aligning with previous research highlighting their motivational features, such as daily check-ins, and the advantages of anonymity (e.g., Zhang and Pérez-Paredes 2021). These tools help overcome time and space limitations and provide a non-judgmental space for practice, which has been shown to alleviate anxiety (Bárkányi 2021). However, despite these benefits, some significant challenges were revealed. First, ASR-based apps often deliver “vague feedback”, failing to identify specific errors like pronunciation subtleties, which limits their effectiveness for more advanced learners – a concern raised by Li and Bonk (2023). Second, the scripted nature of chatbot interactions detracts from their conversational authenticity, diminishing their perceived value, despite their interactive features (Zhang and Pérez-Paredes 2021).

Taken together, actionable insights are drawn from the study. EFL educators could consider integrating AI speaking apps into classroom practice, such as using apps for preparatory rehearsal (aligning with “low-stakes practice” comment), then transferring skills for human interactions. In addition, strategically positioning AI tools as complementary to human interaction may reconcile high-SPCC learners’ skepticism. For AI speaking apps’ developers, enhancing conversational authenticity and tiered feedback systems could address diverse learners’ needs, including creating real-world scenarios for interlocutors and giving error diagnosis with higher accuracy.