This study employs the Technology Acceptance Model 2 (TAM2) (Venkatesh and Davis, 2000) (Figure 1) as its primary theoretical foundation, and it focuses specifically on cognitive instrumental processes to explain English teachers’ BI to adopt AI for LP. TAM2 extends the original TAM by incorporating social influence and cognitive instrumental processes, retaining PU and perceived ease of use (PEU) as core constructs while removing attitude as a mediating variable. Within this TAM 2 framework, OQ, JR, and RD are conceptualized as cognitive instrumental antecedents of PU. Empirical evidence confirms the model’s broad validation across contexts including e-commerce (Paramaeswari and Sarno, 2020), mobile applications (Khoa et al., 2021), and educational technology adoption (Altawalbeh, 2023; Virdi and Mer, 2023), which supports its applicability to this study.

To further contextualize teachers’ intention to use AI for lesson planning, this study also draws on the Decomposed Technology Acceptance Model (DTAM) (Roca et al., 2006) and Self-Determination Theory (SDT) (Ryan and Deci, 2000). DTAM integrates elements of Theory of Planned Behavior (TPB) (Ajzen, 1991), Technology Acceptance Model (TAM) (Davis, 1989), and Expectancy Disconfirmation Theory (EDT) (Oliver, 1980), and positions user satisfaction as a central predictor of continuance intention. According to DATM, sustained usage depends not only on PU but also on user satisfaction derived from factors including information quality, service quality, and cognitive absorption (Roca et al., 2006). SDT further enriches this theoretical perspective by emphasizing that satisfaction of psychological needs—autonomy, competence, and relatedness—fosters intrinsic motivation and sustained behavioral engagement with a technology (Ryan and Deci, 2000).

Building on these theoretical insights, this study introduces NS as a mediating variable linking cognitive evaluations (e.g., OQ, JR, RD), PU, and BI. It posits that English teachers’ adoption of AI-driven lesson planning is motivated by both instrumental utility and its ability to fulfill teachers’ deeper psychological and professional needs. Therefore, OQ, JR, and RD act not only as cognitive antecedents of PU within TAM2 framework but also as predictors to NS, thereby fostering sustained AI adoption intention through both cognitive and affective pathways (Hadji and Degoulet, 2016; Liu and Sutunyarak, 2024; Luo, 2024).

In the Technology Acceptance Model 2 (TAM2), BI is positively influenced by two key factors: PU and PEU. PU refers to the degree to which an individual believes that using a particular technology or system would enhance their job performance. PU exerts a direct positive influence on BI, meaning that if users perceive a technology as useful, they are more likely to form an intention to use it (Venkatesh et al., 2003). Research has revealed that PU has a statistically significant direct effect on intention to use Web 2.0 technologies (Teo et al., 2019).

Moreover, the relationship between PU and user satisfaction has also received substantial empirical support. Recent studies demonstrate that PU significantly enhances user satisfaction with emerging technologies, particularly in the context of generative AI applications in educational settings. Zheng et al. (2024) found that PU positively influences pre-service teachers’ satisfaction with generative AI tools, while Han and Sa (2022) established that PU of online courses significantly contributes to educational satisfaction. In their study of ChatGPT usage in higher education, Yu et al. (2024) found that PEU and PU directly affect user satisfaction, which in turn positively influences continuance use intention.

Output quality, defined by Venkatesh and Davis (2000) as the perceived value and readiness of a system’s outputs for one’s professional use, has been widely recognized as a critical determinant of user satisfaction across various technology adoption contexts. Almufarreh (2024) identified content quality as a pivotal predictor of user satisfaction with generative AI tools in higher education, while Duong et al. (2024) found that both information and service quality significantly affect students’ satisfaction with ChatGPT for learning purposes. Ashfaq et al. (2020) further reinforced this relationship through their analysis of chatbot users.

Mateo et al. (2025) found that PEU, subjective norm, JR, OQ, and RD enhance the PU of the applications, which in turn influences users’ BI to adopt them. Additionally, factors such as subjective norms, professional reputation, JR, and OQ exert an indirect influence on intention, with PU serving as a key mediating variable (Garcia, 2024). Yazdanpanahi et al. (2024) further confirmed that subjective norm, JR, OQ, and RD have a significant positive influence PU. Furthermore, in their examination of EFL learners’ acceptance of ChatGPT, Hwang et al. (2025) found that RD has a significant positive predictor of PU.

Job relevance, defined as the perceived fit between AI tools and core instructional tasks (Venkatesh and Davis, 2000), directly fosters both competence and autonomy needs satisfaction. When AI applications are well aligned with teaching activities (e.g., lesson planning, resource customization), educators gain efficacy in achieving pedagogical goals and experience a greater sense of agency in integrating technology free from external pressure, thereby enhancing overall NS (Ryan and Deci, 2000). This finding is corroborated by Roca et al. (2006), who found that relevance exerts a positive influence on user satisfaction with e-learning systems.

Result demonstrability, defined as the clarity and communicability of the immediate value a system delivers (Venkatesh and Davis, 2000), also contributes to NS by reinforcing competence and autonomy. The positive effect of RD user satisfaction is not limited to education sector; similar outcomes have been observed in other fields, such as architecture (Algassim et al., 2025).

From the perspective of Self-Determination Theory (SDT), NS is underpinned by the fulfillment of three basic psychological needs: autonomy, competence, and relatedness (Deci and Ryan, 1985). BI represents a user’s willingness and behavioral tendency to adopt a certain technology in future practice (Ajzen, 1991; Davis, 1989). Prior research has yielded mixed yet insightful findings regarding the relationship between user satisfaction and BI in technology-mediated learning contexts. Liaw (2008) found that perceived user satisfaction exerts a significant positive effect on learners’ BI to adopt e-learning systems. Consistent positive effects of user satisfaction on BI have also been demonstrated across diverse contexts, including online professional learning communities (Jin et al., 2024), Massive Open Online Courses (MOOCs) (Pozón-López et al., 2021), electronic banking services (Khatoon et al., 2020), and service industries (Tuncer et al., 2021). In the specific context of AI tools, Yu et al. (2024) founded that user satisfaction significantly predicts continuance use intention for ChatGPT in higher education. Rekha et al. (2023) further confirmed that user satisfaction, alongside PU and computer self-efficacy, directly influences learners’ continuance intention in MOOC contexts. Wang et al. (2024) additionally demonstrated that both user satisfaction and task-technology fit both exert significant positive influences on continuance intentions toward educational video platforms.

This relationship, however, is more complex in the educator context. Within the teaching profession, NS derived from instructional and professional practice plays a crucial role in AI technology adoption. When teachers’ basic psychological needs are fulfilled, they are more inclined to demonstrate openness to sustained use of AI technologies. Despite the centrality of this construct, few empirical studies have examined the role of teachers’ NS in their acceptance and adoption of AI tools within educational contexts.

Based on prior theories and literature above, the following hypotheses are proposed as followed:

Based on the theoretical foundations and research hypotheses developed above, we develop an integrated conceptual framework (Figure 2). This model proposes that three external variables, namely OQ, JR, and RD, serve as antecedents influencing both PU and NS. Additionally, PU influences NS. In turn, both PU and NS jointly explain English teachers’ BI toward AI-based lesson planning. The proposed model offers a comprehensive structure for testing the hypothesized relationships and highlights the roles of both technological perceptions and psychological fulfillment in the context of educational AI integration.

Table 1 presents the demographic characteristics of the participating teachers. The total sample comprised 485 participants, including 306 secondary school English teachers and 179 higher education English teachers. For secondary school teachers, the cohort was predominantly female (86.3%). In terms of educational attainment, 84.3% held a bachelor’s degree and 12.4% possessed a postgraduate qualification. Regarding teaching experience, those with 6–15 years (30.1%) and 16–25 years (34.0%) of experience together made up 64.1% of this subgroup—over half the total. Additionally, 55.2% held a Level 1 professional title. These teachers were employed at regular middle schools (184, 60.1%) and key middle schools (122, 39.9%), and taught at the junior high (185, 60.5%) and senior high (121, 39.5%) levels.

For higher education English instructors, females constituted 74.3% of the subgroup. In terms of academic qualifications, 63.7% held a master’s degree and 20.1% a doctoral degree. Nearly half (49.7%) reported 16–25 years of teaching experience, and faculty ranks included Lecturers (49.2%), Associate Professors (36.9%), and other positions (13.9%). In terms of teaching courses, the number of general English teachers (41.9%) and English major teachers (44.1%) was nearly equivalent, with teachers of other English disciplines accounting for the remaining 14.0%.

The questionnaire comprised two sections. The first section collected demographic information, including gender, educational attainment, teaching experience, and professional title. Additional context-specific items were included for university and secondary school teachers respectively: university teachers provided information on the courses they taught, while secondary school teachers reported their school type and school level. The second section assessed teachers’ perceptions of OQ, JR, RD, PU, NS, and BI. The survey adopted a 5-point Likert scale, ranging from 1 (strongly disagree) to 5 (strongly agree), and all six constructs on the scale were adapted from prior empirical research. Table 2 presents the sample items for each construct and their respective original sources.

Data were collected through an online survey completed voluntarily by middle school and university English teachers in China. The online questionnaire link was distributed to participants through snowball sampling via WeChat. The study was conducted in accordance with the ethical guidelines approved by the Institutional Research Ethics Committee (Approval No. LWRW2025003). Informed consent was obtained online from all participants before data collection. At the outset of the questionnaire, all respondents were informed of the study’s aim, and assured of their anonymity. A total of 522 responses were initially collected with 37 excluded from the subsequent analyses due to excessively short completion time or identical responses across all items, indicating potential careless responding. Following data cleaning, the final analytical sample comprised of 485 teachers, including 179 university teachers and 306 secondary school teachers.

In this study, data analysis was performed using partial least squares structural equation modeling (PLS-SEM), a robust approach suited to exploratory and predictive research focused on testing complex relationships among the study’s core variables: OQ, JR, RD, PU, NS, and BI (Hair Jr et al., 2017). PLS-SEM was selected for three key reasons: its flexibility in accommodating both reflective and formative constructs, its suitability for model development in applied research contexts, and its capacity to analyze data with smaller sample sizes compared to covariance-based SEM (Henseler et al., 2015). Analyses followed a two-step procedure: assessment of the measurement model (including validity, reliability and model fit) and evaluation of the structural model (focused on hypothesis testing).

Table 3 presents the descriptive statistics for the constructs measured in the study. The mean scores of all constructs exceeded the theoretical midpoint of 3, suggesting generally positive perceptions of the focal variables among participating teachers. BI yielded the highest mean score (M = 4.16, SD = 0.72), reflecting a strong willingness to adopt AI for lesson planning. This was followed closely by JR (M = 3.90, SD = 0.87) and PU (M = 3.89, SD = 0.72). OQ (M = 3.84, SD = 0.75) and NS (M = 3.76, SD = 0.73) also received favorable ratings, while RD had a comparatively lower, yet still positive, mean score (M = 3.65, SD = 0.79). Standard deviations ranged from 0.72 to 0.87, suggesting responses were reasonably concentrated across the constructs.

We conducted a comprehensive reliability and validity evaluation of the measurement model in accordance with established methodological guidelines (Hair et al., 2019). For reliability assessment, composite reliability (CR) and Cronbach’s alpha were calculated, with all constructs meeting the recommended thresholds: CR values ranged from 0.948 to 0.969 (exceeding the 0.7 benchmark), and Cronbach’s alpha values fell between 0.931 and 0.953 (also surpassing 0.7), indicating strong internal consistency (Chin, 1998; Fornell and Larcker, 1981). Convergent validity was confirmed using two key indicators: (1) all item outer loadings, which ranged from 0.736 to 0.969 and exceeded the recommended threshold of 0.7 (Hair Jr et al., 2017), and (2) average variance extracted (AVE) values for each construct, which ranged from 0.787 to 0.911 and all surpassed the 0.5 cutoff (Fornell and Larcker, 1981) (Table 4). Discriminant validity was assessed via the Fornell-Larcker criterion, which stipulates that the square root of a construct’s AVE must exceed its correlation coefficients with all other constructs (Henseler et al., 2015). For example, the square root of its AVE for JR is 0.955, and its correlation coefficients with other constructs (e.g., 0.750 with BI and 0.641 with NS) were all lower than this value (Table 5). Thus, the measurement model satisfied the Fornell-Larcker criterion for discriminant validity. Collectively, these results confirms that the measurement model exhibits satisfactory reliability and validity, justifying its use in subsequent structural model analyses. Additionally, collinearity among latent variables in the structural model was assessed using variance inflation factors (VIF). VIF values ranged from 2.441 to 3.445, all below the conservative threshold of 5.0, indicating no severe multicollinearity issues that could bias path coefficient estimates (Hair Jr et al., 2017) (Table 4). Finally, prior to structural model analysis, the overall model fit was evaluated using two key indices: the Standardized Root Mean Square Residual (SRMR) and the Normed Fit Index (NFI). The estimated model yielded an SRMR value of 0.088. According to the conventional cutoff criteria proposed by Hu and Bentler (1999), an SRMR value close to 0.08 is considered indicative of a relatively good fit between the model and the observed data. The current SRMR of 0.088 falls just slightly above this threshold, suggesting an acceptable level of fit in terms of residual discrepancy. The model yielded an NFI of 0.897 (Figure 3). While a conventional cutoff of 0.90 is typically recommended for acceptable fit (Bentler and Bonett, 1980), social science and educational technology research has noted that NFI values between 0.85 and 0.90 are deemed reasonable for exploratory studies, particularly those with a theoretical foundation that test novel variable relationships (Henseler et al., 2016).

Having established the reliability and validity of the measurement model, we employed PLS-SEM to assess the structural model and examine the hypothesized relationships (Sarstedt et al., 2022). This analytical method is especially well-suited for complex models featuring latent variables, given its ability to handle non-normal data and optimize the explained variance of endogenous constructs (Hair and Alamer, 2022). The structural model evaluation focused on path coefficients, their significance levels evaluated through 5,000-bootstrap samples, and the coefficient of determination (R²) for endogenous variables, which indicates the proportion of variance explained by the model (Henseler et al., 2015). Additionally, we assessed the effect size (f²) to quantify the practical significance of exogenous variables on endogenous constructs and predictive relevance (Q²) using blindfolding to validate the model’s predictive capability (Hair et al., 2019). The results are below.

Table 6 presents the outcomes for the hypothesis test and Figure 3 shows path coefficients of the research model. The PLS-SEM analysis with 5,000 bootstrap samples confirmed seven of nine hypothesized paths were statistically significant and supported (p < 0.01). The results show that OQ, JR, and RD are all significant antecedents of PU, with RD exhibiting the strongest effect (β = 0.435, p < 0.001). Furthermore, PU significantly influences both BI (β = 0.428, p < 0.001) and NS (β = 0.529, p < 0.001), confirming its central role in the model. OQ also shows a significant positive effect on NS (β = 0.205, p < 0.001). However, the effects of JR and RD on NS were not statistically significant (β = 0.063, p = 0.162; β = 0.097, p = 0.062, respectively), leading to the rejection of H6 and H7. Finally, NS demonstrates a significant, though relatively modest, positive effect on BI (β = 0.180, p < 0.05).

The model demonstrated strong predictive power, with Q² values of 0.573 for PU, 0.594 for NS, and 0.264 for BI, all exceeding the 0.02 threshold for predictive relevance (Hair et al., 2019).

The mediation analysis reveals significant indirect pathways through which the exogenous variables influence NS and BI. OQ, RD, and JR all exert significant positive indirect effects on NS, exclusively mediated by PU (β = 0.142, p < 0.001; β = 0.230, p < 0.001; β = 0.098, p < 0.001, respectively) (Table 7).

All three external variables demonstrate (OQ, JR, RD) significant total indirect effects. OQ influences BI through three pathways: a strong direct mediation via PU (β = 0.115, p < 0.001), a weaker but significant chain mediation through PU → NS (β = 0.026, p < 0.05), and a non-significant path through NS alone, resulting in a substantial total indirect effect (β = 0.178, p < 0.001). Similarly, RD affects BI primarily through PU (β = 0.186, p < 0.001) with additional minor contributions from other pathways, yielding the strongest total indirect effect among all predictors (β = 0.245, p < 0.001). JR shows a more modest but still significant total indirect effect on BI (β = 0.108, p < 0.001), mainly driven by the PU pathway (β = 0.079, p < 0.01).

Notably, PU itself demonstrates a significant indirect effect on BI through NS (β = 0.095, p < 0.05), confirming the partial mediation role of needs satisfaction in the technology acceptance process.

The findings confirm that OQ significantly enhances both PU and NS, supporting H1 and H5, respectively. This demonstrates that high-quality outputs from AI tools, such as accurate teaching materials and pedagogically appropriate content, directly enhance teachers’ perceptions of the technology’s practical utility while simultaneously fulfilling their core psychological needs. The significant positive effect of OQ on PU aligns with prior findings from prior studies, where OQ consistently emerges as a crucial antecedent of PU across various technological contexts (Garcia, 2024; Jin et al., 2025b; Yazdanpanahi et al., 2024). Meanwhile, the direct relationship between OQ and NS corroborates research emphasizing quality dimensions as a fundamental driver of user needs satisfaction (Al-Fraihat et al., 2020; Almufarreh, 2024).

From a theoretical perspective, these dual effects can be understood through an integrated lens of technology acceptance and psychological need fulfillment. High-quality AI outputs directly boost teachers’ sense of professional competence by providing reliable, professionally aligned materials that reduce cognitive load in LP, thus addressing both the instrumental and psychological dimensions of educational technology adoption. When AI tools generate high-quality, instructionally relevant outputs, teachers are more likely to perceive these tools as genuinely useful for enhancing their instructional effectiveness. This heightened perception of usefulness subsequently fulfills their psychological needs for competence and autonomy by enabling more efficient and effective lesson planning practices (Roca et al., 2006; Ryan and Deci, 2000).

The results indicate that JR has a significant positive effect on PU, supporting H2. This finding is consistent with prior technology acceptance literature (Jin et al., 2025a), which identifies JR as a critical cognitive instrumental factor shaping users’ assessments of a technology’s utility. When teachers perceive AI tools as closely aligned with their core tasks, such as curriculum design and teaching material preparation, they are more likely to recognize the practical value of these tools, thereby elevating their perceptions of usefulness.

In contrast to the proposed hypothesis, JR did not exert a significant direct effect on NS, leading to the rejection of H6. This finding contrasts with certain prior studies in technology acceptance literature such as the work by Roca et al. (2006), which identified JR as a positive predictor of user satisfaction in e-learning systems. This discrepancy can be explained through a cognitive-affective mediating pathway. While AI-generated content (e.g., lesson plans, activity designs) may be relevant to core lesson-planning tasks, two key practical issues often emerge. First, such content often fails to meet teachers’ personalized needs (e.g., failing to align with the learning characteristics of specific grade levels or teachers’ habitual teaching styles) (Walter, 2024). Second, redundant information in some AI-generated content increases teachers’ cognitive load during the material screening and revision materials, thereby reducing overall lesson-planning efficiency (Mehta et al., 2025). In this context, teachers do not experience NS merely because AI-generated content is relevant to lesson planning. Instead, they first need to verify the practical utility of such content, for example, whether it reduces the workload of manual lesson planning or improves in-class effectiveness by optimizing activity designs (Ke and Gong, 2025). This process clearly demonstrates that the JR of AI-generated content to lesson-planning tasks can only influence teachers’ NS indirectly, through the mediating role of PU.

Our findings reveal that RD has a strong, significant positive effect on PU, supporting H3. This robust relationship indicates that when AI tools deliver tangible, observable benefits, such as clear efficiency gains or visible improvements to teaching materials, teachers are far more likely to perceive these tools as genuinely useful for their instructional practice.

This result aligns closely with prior technology acceptance research. The significant positive effect of RD on PU corroborates findings across multiple educational technology contexts, where the tangibility of outcomes consistently enhances perceived utility (Mateo et al., 2025; Yazdanpanahi et al., 2024). The substantial effect size (β = 0.435) further underscores RD’s crucial role as a cognitive instrumental factor in shaping teachers’ assessments of the practical value of AI tools. However, RD showed no significant direct effect on NS, leading to the rejection of H7. This finding contrasts with the recent work by Duong et al. (2024), which highlighted RD’s significant role in enhancing perceived value. This divergence can be explained by the inherent conceptual differences between these two constructs. RD refers to the tangibility and observability of an AI tool’s output and its effectiveness in communication (Moore and Benbasat, 1991; Venkatesh and Bala, 2008), a construct often associated with immediate, surface-level performance gains. In contrast, NS refers to the degree to which using a technology fulfills an individual’s innate psychological needs for autonomy, competence, and relatedness (Ryan and Deci, 2000). While AI-assisted lesson planning tools can demonstrably save time and propose enhanced instructional strategies, thus increasing perceived usefulness, their ultimate pedagogical efficacy is often delayed and must be evaluated post-hoc through student feedback, assessment results, and observations of classroom engagement (Zheng et al., 2024). This temporal disconnect creates a disjuncture whereby the immediate, demonstrable outcomes of the tool (high RD) fail to provide teachers with instant, intuitive feedback on their sense of professional competence or their pedagogical autonomy in the teaching process. Consequently, the direct pathway from RD to NS is disrupted, as the tangible results generated during the lesson planning phase do not immediately translate into intrinsic psychological need fulfillment in actual instructional use.

This intriguing divergence suggests that while demonstrable outcomes strongly shape cognitive evaluations of usefulness, they do not directly translate to psychological need fulfillment. The core distinction lies in their conceptual domains: RD addresses external, observable performance benefits, whereas NS encompasses internal psychological needs for autonomy, competence, and relatedness (Ryan and Deci, 2000). The immediate practical benefits captured by RD may therefore influence NS primarily through their enhancement of PU, rather than via direct psychological pathways.

The findings demonstrate that PU has a significant positive influence both NS and BI, providing strong support for H4 and H8. These results align with the core propositions of TAM2 (Venkatesh and Davis, 2000) and extend their applicability to the context of AI-assisted lesson planning for English teachers.

The substantial positive effect of PU on NS (β = 0.529) reinforces and extends previous research into the psychological mechanisms underlying educational technology adoption. This finding corroborates Zheng et al.’s (2024) observation that PU enhances pre-service teachers’ satisfaction with generative AI tools, and also supports Al-Fraihat et al.’s (2020) identification of PU as a significant determinant of user satisfaction in e-learning systems. The strong positive relationship indicates that when teachers recognize the practical value of AI tools in improving lesson planning efficiency and effectiveness, this cognitive appraisal directly contributes to the fulfillment of their psychological needs, particularly those for competence and autonomy.

Furthermore, the notable positive effect of PU on BI (β = 0.428) substantially reinforces the foundational TAM2 proposition that perceptions of usefulness directly drive technology usage intentions (Venkatesh et al., 2003). This relationship exhibits remarkable consistency across technological contexts, echoing Teo et al.’s (2019) findings on Web 2.0 technologies and confirming the robustness of advanced AI applications in educational settings. The exceptionally strong path coefficient underscores the paramount importance of PU in determining English teachers’ willingness to adopt AI tools for lesson planning.

The results confirm that NS exerts a significant positive direct effect on BI, thereby supporting H9. This aligns with a core proposition of SDT (Ryan and Deci, 2000), which posits that fulfilling individuals’ basic psychological needs, autonomy, competence, and relatedness enhances their intrinsic and extrinsic motivation to engage in a given a behavior. This finding is consistent with prior research on educational technology adoption (Ashfaq et al., 2020), confirming that when teachers’ psychological needs are fulfilled through AI tools, their intention to adopt such technologies is significantly strengthened.

Mediation analysis further reveals a more nuanced pattern of effects. While NS was influenced by several antecedents, OQ, RD, and JR through PU indirectly through PU, its mediating role was limited. Crucially, NS only served as a significant mediator in the PU → BI relationship.

This specific mediating pathway is strongly supported by SDT (Deci et al., 2017) and recent research in educational technology (Wang et al., 2024). It indicates that when teachers perceive AI tools as useful for enhancing lesson quality, this perception satisfies their underlying psychological needs for competence (by mastering an effective instructional tool) and autonomy (by providing flexible approaches to task completion). This user satisfaction then acts a key psychological mechanism that translates the cognitive judgment of usefulness into a stronger BI to adopt the technology.

However, the direct mediating paths from OQ, RD, and JR to BI via NS alone (i.e., OQ → NS → BI; RD → NS → BI; JR → NS → BI) were not statistically significant. This unexpected pattern can be explained by a potential “quality-threshold effect” (Kelchtermans, 2009). While factors such as information quality may meet teachers’ basic requirements for effective lesson planning, this baseline level of satisfaction may be insufficient on its own to directly drive adoption intention. Teachers may require additional, more compelling incentives, such as the promise of long-term efficiency gains or significant workload reduction, to fully convert psychological need satisfaction into a decisive behavioral intention to act.