K-12 Teachers’ Adoption of Generative AI for Teaching: An Extended TAM Perspective

Abstract

This study investigates the factors influencing Chinese K-12 teachers’ adoption of generative artificial intelligence (GenAI) for instructional purposes by extending the Technology Acceptance Model (TAM) with pedagogical beliefs, perceived intelligence, perceived ethical risks, GenAI anxiety, and demographic moderators. Drawing on a theory-driven framework, survey data were collected from 218 in-service teachers across K-12 schools in China. The respondents were predominantly from urban schools and most had prior GenAI use experience. Eight latent constructs and fourteen hypotheses were tested using structural equation modeling and multi-group analysis. Results show that perceived usefulness and perceived ease of use are the strongest predictors of teachers’ intention to adopt GenAI. Constructivist pedagogical beliefs positively predict both perceived usefulness and intention, whereas transmissive beliefs negatively predict intention. Perceived intelligence exerts strong positive effects on perceived usefulness and ease of use but has no direct effect on intention. Perceived ethical risks significantly heighten GenAI anxiety, yet neither directly reduces adoption intention. Gender, teaching stage, and educational background further moderate key relationships, revealing heterogeneous adoption mechanisms across teacher subgroups. The study extends TAM for the GenAI era and highlights the need for professional development and policy initiatives that simultaneously strengthen perceived usefulness and ease of use, engage with pedagogical beliefs, and address ethical and emotional concerns in context-sensitive ways.

1. Introduction

The rapid advancement of artificial intelligence (AI) technology has brought profound transformations across multiple sectors of society (Luan et al., 2020). Among these developments, generative artificial intelligence (GenAI) has emerged as particularly important for educational innovation. By automating content creation, providing interactive feedback, and enabling new forms of human–machine collaboration, GenAI is reshaping the digital transformation of education (Mello et al., 2023) and altering the landscape of classroom instruction.

A growing array of GenAI tools has been developed to address diverse educational needs. In China, for example, tools such as DeepSeek, Doubao, and Kimi are increasingly adopted by educators for lesson planning, instructional delivery, and student assessment (Liao et al., 2025). Integrated across pre-class preparation, in-class support, and post-class assignment design, these tools are becoming embedded in teachers’ daily practice and are regarded as promising means of enabling more personalized, efficient, and data-informed learning experiences (Haroud & Saqri, 2025; Nasr et al., 2025; Beimel et al., 2025).

An emerging body of research has begun to document the educational potential of GenAI. Studies have highlighted its capacity to foster student creativity (R. Liu et al., 2025), enhance classroom engagement (Pahi et al., 2024), support the development of computational thinking (Liao et al., 2024), and streamline assessment processes (Jukiewicz, 2024). However, the effective integration of GenAI in schools depends not only on the technological affordances of these tools, but also on sound instructional design and, crucially, on teachers’ adoption and use decisions.

Although many studies have examined teachers’ attitudes toward using GenAI tools in education, most have relied on established frameworks such as the Technology Acceptance Model (TAM) and its extensions, or the Technological Pedagogical Content Knowledge (TPACK) framework. While these frameworks provide valuable insights on technology adoption, they were not originally designed with GenAI in mind, and therefore, tend to under-emphasize its distinctive attributes, such as its generative capabilities and conversational interfaces. Building on this line of work, Y. Wang et al. (2025) applied the Unified Theory of Acceptance and Use of Technology (UTAUT) to investigate the roles of social influence and facilitating conditions in pre-service teachers’ adoption of ChatGPT (https://chatgpt.com/, accessed on 12 January 2026) in China. Similarly, Yue et al. (2024) explored how demographic factors (e.g., gender, subject taught, grade level, teaching experience, and prior AI experience) shape teachers’ preparedness for and attitudes toward AI in education.

At the same time, existing literature has mainly focused on pre-service teachers (Gao et al., 2025; Yang & Appleget, 2024) and higher education instructors (Feng et al., 2025), with relatively limited attention paid to K-12 teachers. For instance, Y. Wang et al. (2021) investigated how anxiety, self-efficacy, and attitudes toward AI predict higher education instructors’ intentions to adopt AI. While such studies offer important insights, they emphasize individual-level perceptions such as attitudes, perceived ease of use, and perceived usefulness. These factors do not fully capture the complex, context-specific challenges that K-12 teachers encounter when integrating GenAI into everyday classroom practice, particularly within the Chinese education system.

Given these gaps in the literature, the present study seeks to develop a more nuanced understanding of the factors that influence Chinese K-12 teachers’ adoption of GenAI tools. Employing a quantitative survey research, the study examines how teachers’ intentions to adopt GenAI are shaped by constructs drawn from TAM, GenAI-specific attributes, and teaching-related factors. In doing so, this study contributes empirical evidence and theoretical refinement to the literature on GenAI in education and offers practical implications for supporting K-12 teachers’ meaningful and responsible use of GenAI in their instructional practices.

2. Research Hypotheses

TAM and UTAUT have been extensively applied in educational research to explain teachers’ technology adoption behaviors (Granić, 2022). The validity and reliability of these models have been supported across a range of educational contexts (Sun et al., 2025; Strzelecki & ElArabawy, 2024) and numerous adaptations of TAM have been employed to examine teachers’ and students’ acceptance of diverse educational tools (Koutromanos et al., 2015; Mac Callum & Jeffrey, 2014). In general, these models highlight three core determinants of users’ behavioral intention: perceived ease of use, perceived usefulness, and attitude toward use (Marangunić & Granić, 2015).

The present study primarily adopts TAM to investigate the factors shaping Chinese K-12 teachers’ intention to adopt generative artificial intelligence (GenAI) for teaching. In addition, existing research indicates that in educational settings, teachers’ decisions regarding the adoption of new technologies are not solely based on an evaluation of the tools’ functional capabilities, but are also significantly influenced by their pedagogical beliefs and the specific teaching context (Luik et al., 2024; Choi et al., 2023). Moreover, as an emerging technology characterized by a high degree of autonomy and content generation capabilities, GenAI’ s technical features may also play a significant role in shaping teachers’ attitudes towards its use and their intentions for employing it (Zhang et al., 2024). Therefore, the original TAM is extended with additional constructs that capture GenAI-specific perceptions and teachers’ pedagogical and contextual characteristics. This broadened perspective is expected to provide a more nuanced understanding of the mechanisms underpinning the integration of GenAI into classroom instruction.

2.1. Technology Acceptance Model

The Technology Acceptance Model (TAM) posits that perceived usefulness (PU) and perceived ease of use (PEU) are two key beliefs that form a causal chain with behavioral intention (BI) and external variables (Davis et al., 1989). In this study, behavioral intention is conceptualized as teachers’ intention to use or continue using GenAI for teaching in the near future. PU is defined as the extent to which users believe that using a given technology will enhance their work or learning efficiency, and it is conceptually similar to performance expectancy (Cox, 2012). When teachers perceive that a technology improves their performance, they are more inclined to adopt it (Anthony et al., 2023). Accordingly, novel educational technologies have often been enthusiastically embraced by teachers and students because they are perceived to increase efficiency (Zheng et al., 2024). In the context of GenAI, multiple studies have shown that PU exerts a significant positive effect on users’ behavioral intentions (Al-Emran et al., 2023; Sing et al., 2022; Zhao et al., 2025). PEU refers to the degree to which an individual believes that using a technology will be free of effort, and it is closely related to perceived simplicity. Technologies that are easy to operate and require a short learning curve are more likely to be accepted by pre-service teachers (Wijaya et al., 2022). At the early stage of technology adoption, PEU strongly influences behavioral intention, making users more willing to engage with user-friendly emerging technologies (Venkatesh & Davis, 2000).

In line with TAM and prior findings, this study proposes the following hypotheses:

H1. 

Perceived usefulness has a significant positive impact on teachers’ intention to adopt GenAI in teaching.

H2. 

Perceived ease of use has a significant positive impact on teachers’ intention to adopt GenAI in teaching.

2.2. Pedagogical Beliefs

Pedagogical beliefs (PB) refer to teachers’ fundamental conceptions of teaching and learning (Ertmer, 2005). These beliefs are commonly distinguished as transmissive pedagogical beliefs (TPB) and constructivist pedagogical beliefs (CPB). TPB emphasizes the transmission and supervision of students’ knowledge acquisition and is typically aligned with behaviorist approaches in classroom practice. In contrast, CPB stress students’ individuality, needs, and active participation in knowledge construction, and are usually associated with constructivist or social constructivist classroom practices (Deng et al., 2014). Transmissive and constructivist orientations are sometimes presented as theoretically opposed. Yet research on Chinese and other Confucian-heritage contexts suggests they can function as partially independent dimensions and may co-occur as teachers balance reform-oriented student-centered practices with examination- and curriculum-driven instructional demands (Chan & Elliott, 2004; M. Liu et al., 2022; Y. Wang et al., 2022; J. Wang & Kim, 2023).

Teachers’ pedagogical beliefs are related to how they select, integrate, and adapt technologies in teaching. Cheng et al. (2022) found that teachers’ competence beliefs and pedagogical beliefs jointly influence their intentions to integrate technology. Kim et al. (2013) further showed that teachers with CPB tend to adopt student-centered approaches and actively leverage digital tools to support collaborative and inquiry-based learning, whereas teachers with TPB typically use digital tools to reinforce teacher-centered, content-delivery activities. Recent research has begun to incorporate teachers’ pedagogical beliefs into TAM. Choi et al. (2023) reported that teachers holding stronger CPB were more likely to use educational AI tools than those with predominantly TPB. Moreover, CPB was found to exert a significant direct effect on both PU and PEU. These findings suggest that teachers’ underlying views about teaching and learning may shape not only their attitudes toward educational AI, but also their perceptions of its value and usability.

Drawing on this body of work, the present study proposes the following hypotheses:

H3. 

Constructivist pedagogical beliefs have a significant positive impact on teachers’ intention to adopt GenAI in teaching.

H4. 

Transmissive pedagogical beliefs have a significant negative impact on teachers’ intention to adopt GenAI in teaching.

H5. 

Constructivist pedagogical beliefs have a significant positive impact on teachers’ perceived usefulness of GenAI.

H6. 

Constructivist pedagogical beliefs have a significant positive impact on teachers’ perceived ease of use of GenAI.

H7. 

Transmissive pedagogical beliefs have a significant negative impact on teachers’ perceived usefulness of GenAI.

H8. 

Transmissive pedagogical beliefs have a significant negative impact on teachers’ perceived ease of use of GenAI.

2.3. Perceived Intelligence

Perceived intelligence (PI) refers to users’ overall perception of a GenAI system can understand input, act autonomously, provide effective responses, and engage in natural language interactions (Bawack, 2021). Systems that deliver rapid, contextually appropriate, and coherent feedback tend to be evaluated more positively and are seen as more competent. For example, GenAI tools such as ChatGPT can enhance users’ experience by adapting to their conversational style and preferences, thereby fostering more natural and engaging interactions. Conversely, when such systems generate inaccurate, irrelevant, or nonsensical responses, users’ perceptions of their usefulness and reliability may be undermined (Bernabei et al., 2023).

Within the classic TAM, system attributes often function as external variables whose influence on intention is transmitted primarily through core cognitive beliefs, particularly PU and PEU (Abdelkader, 2023; Moussawi et al., 2021; Sahari et al., 2023). From this perspective, PI can be conceptualized as a distal capability cue. When teachers perceive GenAI as more intelligent, they are more likely to judge it as beneficial for teaching (higher PU) and easier to work with in practice (higher PEU) (Zhong & Tang, 2025). At the same time, once PU and PEU are accounted for, PI may have a weak or non-significant direct relationship with behavioral intention because its primary explanatory role is to shape these proximal TAM beliefs rather than to independently motivate adoption.

Accordingly, this study examines both the direct and belief-mediated roles of PI in teachers’ adoption intention and proposes:

H9. 

Perceived intelligence has a significant positive impact on teachers’ intention to adopt GenAI in teaching.

H10. 

Perceived intelligence has a significant positive impact on teachers’ perceived usefulness of GenAI.

H11. 

Perceived intelligence has a significant positive impact on teachers’ perceived ease of use of GenAI.

2.4. Perceived Ethical Risks

Perceived ethical risks (PER) reflect teachers’ concerns and uncertainty about potential negative moral consequences of using GenAI in teaching, such as bias, misuse of student data, lack of transparency, academic integrity violations, and unclear responsibility for AI-generated outputs (Tan, 2002; Zhu et al., 2025). In TAM-related research, PER often functions as an inhibiting belief, hindering users’ willingness to adopt and continuously use technology (Chatterjee, 2020; Yim & Wegerif, 2024). Prior studies have shown that PER associated with AI products can significantly shape individuals’ technology-related behaviors, including their cautious use, avoidance, or selective adoption of AI systems (Wu et al., 2022; Gerlich, 2023). Accordingly, this study proposes the following hypotheses:

H12. 

Perceived ethical risks have a significant negative impact on teachers’ intention to adopt GenAI in teaching.

H13. 

Perceived ethical risks have a significant positive impact on teachers’ GenAI anxiety.

2.5. GenAI Anxiety

Anxiety is commonly defined as an emotional state characterized by fear, discomfort, frustration, rumination, and worry, which can shape individuals’ judgments and decision-making processes (Nayak, 2014; Thanigan et al., 2021). Building on this notion, GenAI anxiety (GAIA) refers to teachers’ uneasy or worried emotional responses when considering or using GenAI for teaching. This anxiety may involve concerns about losing instructional control, reduced professional autonomy, reputational risks from AI errors, uncertainty about acceptable practice, and broader worries about the long-term implications of AI for teachers’ roles and development (J. Li & Huang, 2020).

Such affective responses can influence teachers’ decisions and behaviors regarding the use of GenAI in teaching, leading to hesitation or resistance in adopting these tools (Jatileni et al., 2024). Furthermore, a recent meta-analysis by Xu et al. (2025) highlighted that the drawbacks of GenAI constitute important barriers to its adoption. These findings suggest that ethical risks associated with GenAI may heighten teachers’ perceived anxiety, which in turn dampen their intention to integrate GenAI into their teaching practice. Accordingly, the following hypotheses are proposed:

H14. 

GenAI anxiety has a significant negative impact on teachers’ intention to adopt GenAI in their teaching practices.

2.6. Gender, Educational Background, and Teaching Grade

This study examines heterogeneity in GenAI acceptance by testing whether key structural relationships differ across teacher subgroups. Moderation is theoretically meaningful in GenAI adoption because teachers’ evaluations of usefulness, usability, and pedagogical fit are shaped by their individual professional experiences and the instructional context in which GenAI would be implemented (Scherer & Teo, 2019).

Specifically, gender may moderate TAM pathways because prior TAM research suggests subgroup differences in the salience of usability burdens, confidence with new tools, and the extent to which PEU translates into adoption intention (Al-kfairy, 2024; Bazelais et al., 2024; Collie & Martin, 2024).

Educational background is conceptualized as a proxy for professional capital that can shape teachers’ capacity to translate pedagogical aspirations into workable GenAI-supported practice. Higher qualifications may be associated with stronger pedagogical knowledge, greater access to professional learning opportunities, and stronger confidence in evaluating AI output, which can amplify the extent to which constructivist beliefs translate into PU.

The teaching stage captures contextual differences in curriculum structure, exam pressure, classroom autonomy, and typical instructional routines. These differences can condition whether constructivist pedagogical beliefs align with feasible GenAI uses in daily teaching and thus can moderate the belief-to-usefulness pathway.

Accordingly, we treat gender, educational background, and teaching stage as moderators to identify which mechanisms are stable across teachers and which are context- and subgroup-sensitive, thereby improving the educational interpretability of the model.

The overall theoretical model is presented in Figure 1.

3. Methodology

3.1. Research Design

This study adopted a theory-driven quantitative design to examine the factors influencing Chinese K-12 teachers’ intention to adopt GenAI in teaching. The research process comprised two main components: (a) construction of a conceptual framework and (b) a questionnaire survey. First, a literature review of studies published between 2018 and 2025 was conducted in the Web of Science and ERIC databases using combined keywords (e.g., “generative AI + K-12 + teacher adoption,” “technology acceptance model + education”). Drawing on TAM and related technology adoption theories, seven influencing factors were identified. Together with teachers’ GenAI adoption intention, these formed an initial framework comprising eight latent constructs and fourteen hypotheses, grouped into three categories: (a) core TAM variables, (b) factors related to the distinctive characteristics of GenAI, and (c) teachers’ contextual factors.

Based on this framework, a structured questionnaire was developed to operationalize the constructs and quantitatively test the hypothesized relationships. Structural Equation Modeling (SEM) was employed to examine the direct and indirect effects among the influencing factors and to assess the overall fit of the proposed model.

3.2. Participants and Procedure

Questionnaires were distributed to K-12 teachers nationwide. A total of 283 responses were collected, of which 218 were retained as valid after data screening (e.g., removal of incomplete or patterned responses).

These 218 teachers constituted the sample for the quantitative analysis, which includes 62 primary school teachers (28.4%) and 156 middle school teachers (71.6%). Among them, 61 (28.0%) were male and 157 (72.0%) were female. Regarding educational background, 81.7% of the teachers held a bachelor’s degree or below, and 18.3% held a master’s degree or above. In terms of school location, 167 teachers (76.6%) worked in urban schools, 51 (23.4%) in county/town/rural schools. With respect to GenAI-related experience, 182 teachers (83.5%) had used GenAI, and 162 (74.3%) had already applied GenAI in their teaching.

The distribution of sample characteristics is presented in

Table 1. Distribution of sample characteristics.

Item	Option	Frequency	Percentage (%)
Gender	Male	61	28.0
	Female	157	72.0
Teaching Stage	Primary School	62	28.4
	Middle School	156	71.6
Educational background	Bachelor’s Degree or Below	178	81.7
	Master’s Degree or Above	40	18.3
School location	Urban	167	76.6
	County/Town/Rural	51	23.4
Prior GenAI experience	Yes	182	83.5
	No	36	16.5
GenAI teaching application	Yes	162	74.3
	No	56	25.7
Total		218	100.0

.

Because the survey relied on voluntary participation, the sample reflects self-selection that likely overrepresented teachers with stronger digital access and interest in GenAI. In particular, teachers from urban schools and teachers with prior GenAI experience were disproportionately represented. Therefore, the findings should be interpreted as explaining intention mechanisms among teachers with relatively high exposure to GenAI, rather than as nationally representative estimates or a comprehensive account of teachers with limited access or no prior experience. The results should be interpreted primarily as explaining teachers’ intention to sustain or expand GenAI use, rather than first-time adoption only.

3.3. Data Collection and Analysis

The research instruments consisted of a questionnaire on teachers’ GenAI adoption in teaching. The questionnaire comprised three parts: (a) demographic information (e.g., gender, educational background, teaching stage, school location), (b) teachers’ actual GenAI adoption behaviors in teaching, and (c) eight latent constructs capturing the factors influencing GenAI adoption.

All latent constructs were measured using a five-point Likert scale ranging from 1 (“strongly disagree”) to 5 (“strongly agree”). Behavioral intention to adopt GenAI for teaching, perceived usefulness, and perceived ease of use were adapted from TAM–related scales. Constructivist and transmissive pedagogical beliefs were adapted from Choi et al. (2023). Perceived intelligence was measured with items from Maheshwari (2024), perceived ethical risks from Zhu et al. (2025), and GenAI anxiety from Jatileni et al. (2024). The final instrument demonstrated satisfactory internal consistency (Cronbach’s α = 0.917) and sampling adequacy (KMO = 0.905).

Because all variables were collected via self-report in a single survey at one time point, common method variance (CMV) is a potential concern and may inflate observed associations. To reduce evaluation apprehension, the survey was administered anonymously, and several items were reverse-scored. We also conducted Harman’s single-factor test as a basic diagnostic. The first unrotated factor accounted for 26.03% of the variance, indicating that no single factor dominated the covariance structure (Cohen & Ehrlich, 2019; Fuller et al., 2016). Note that Harman’s test is a coarse and relatively insensitive procedure and cannot rule out more subtle CMV effects. Accordingly, the structural relationships should be interpreted with appropriate caution.

Structural equation modeling (SEM) was employed to test the proposed relationships among the eight constructs and the fourteen hypotheses. SEM was conducted in AMOS 24 using maximum likelihood estimation, which allows for the simultaneous estimation of the measurement and structural models, and is appropriate for examining complex causal relationships among latent variables.

The analytical procedure followed a two-step approach. First, the measurement model was evaluated in terms of reliability, convergent validity, discriminant validity, and overall model fit. Items with standardized factor loadings below 0.50 were designated for removal; all retained items met this criterion. Second, the structural model was estimated to test the hypothesized paths and assess the proportion of variance explained in each endogenous construct.

In addition to estimating the direct structural paths, we formally tested indirect effects using a bias-corrected bootstrap procedure in AMOS (5000 resamples). Bootstrapped 95% confidence intervals (CIs) were used to evaluate whether indirect effects differed from zero, providing an empirical test of mediation beyond inspection of direct-path significance.

To examine the moderating roles of gender, educational background, teaching stage, and school location, multi-group SEM analyses were carried out. For each moderator, models were estimated separately for the relevant subgroups. An unconstrained model with all parameters freely estimated across groups was first specified, followed by a constrained model in which structural paths were set to be equal between groups. Chi-square difference tests were used to examine whether constraining the paths significantly worsened model fit. For paths with significant between-group differences, subgroup-specific regression coefficients were compared to identify how the relations between predictor variables and behavioral intention varied across subgroups.

4. Results

4.1. Descriptive Results

Descriptive statistics for all key variables, including means, standard deviations, skewness, and kurtosis, are presented in

Table 2. Descriptive statistics.

Variable	Mean	SD	Skewness	Kurtosis
BI	3.72	0.80	−0.995	1.756
PU	3.96	0.71	−0.768	0.610
PEU	3.79	0.76	−0.694	0.473
CPB	4.15	0.52	−0.206	−0.060
TPB	3.74	0.69	−0.509	0.278
PI	3.72	0.59	−0.292	−0.372
PER	3.26	0.89	−0.302	−0.561
GAIA	2.46	1.01	0.560	−0.550

. The absolute values of skewness were all below 1 (ranging from −0.995 to 0.560), and the absolute values of kurtosis were below 2 (ranging from −0.561 to 1.756), indicating that the observed distributions approximated univariate normality and were suitable for subsequent structural equation modeling.

Consistent with the TAM-based framework, teachers reported relatively high levels of behavioral intention to adopt GenAI for teaching (BI; M = 3.72) as well as generally positive evaluations of GenAI’s usefulness (PU; M = 3.96) and ease of use (PEU; M = 3.79). Given that all constructs were measured on five-point Likert scales, these means are above the theoretical midpoints, suggesting that the sampled Chinese K-12 teachers tended to hold favorable views of GenAI and expressed a relatively strong willingness to integrate it into their instructional practice.

For the GenAI-specific variables, perceived intelligence (PI; M = 3.72) and perceived ethical risks (PER; M = 3.26) were both above the midpoint, indicating that teachers generally viewed GenAI tools as highly intelligent while also recognizing non-trivial ethical concerns. In contrast, the mean score for GenAI anxiety (GAIA; M = 2.46) was the lowest among all constructs and slightly below its theoretical midpoint. Combined with its relatively large standard deviation (SD = 1.01), this suggests that, on average, teachers did not experience high anxiety about GenAI, but there was considerable variation across individuals in their anxiety levels.

With respect to pedagogical beliefs, the mean scores for CPB (M = 4.15) and TPB (M = 3.74) were both clearly above their scale midpoints. This pattern indicates that teachers simultaneously endorsed both student-centered and teacher-centered beliefs. Rather than being mutually exclusive, their pedagogical belief systems appear to be multifaceted and context-sensitive. Importantly, the two belief dimensions are empirically distinguishable in the measurement model: the CPB–TPB correlation is moderate and is lower than the square roots of their AVEs, satisfying the Fornell–Larcker criterion and supporting discriminant validity.

4.2. Measurement Model Analysis

Before testing the structural relationships, the measurement model was evaluated for reliability and validity. Confirmatory factor analysis (CFA) was conducted. Items with standardized factor loadings below 0.50 were considered for removal; all retained items exceeded this threshold, and the final measurement model is reported in

Table 3. Test of convergent validity and composite reliability.

Latent Construct	Item	Estimate	AVE	CR	Cronbach’s α
BI	BI1	0.733	0.579	0.804	0.807
	BI2	0.826
	BI3	0.720
CPB	CPB1	0.744	0.561	0.793	0.723
	CPB2	0.740
	CPB3	0.763
TPB	TPB1	0.614	0.592	0.810	0.807
	TPB2	0.814
	TPB3	0.859
PU	PU1	0.865	0.715	0.881	0.870
	PU2	0.714
	PU3	0.943
PEU	PEU1	0.860	0.712	0.880	0.880
	PEU2	0.729
	PEU3	0.930
PI	PI1	0.820	0.582	0.805	0.788
	PI2	0.670
	PI3	0.791
GAIA	GAIA1	0.835	0.760	0.904	0.904
	GAIA2	0.873
	GAIA3	0.906
PER	PER1	0.692	0.595	0.814	0.811
	PER2	0.841
	PER3	0.775

.

Table 3 shows the standardized factor loadings, average variance extracted (AVE), composite reliability (CR), and Cronbach’s α for each construct. Standardized loadings ranged from 0.614 to 0.943 and were all statistically significant at p < 0.001. CR values ranged from 0.793 to 0.904, AVE values from 0.561 to 0.760, and Cronbach’s α coefficients from 0.723 to 0.904. These indices exceed commonly accepted benchmarks (CR ≥ 0.70; AVE ≥ 0.50), indicating satisfactory convergent validity and internal consistency for all constructs.

Discriminant validity was assessed using the Fornell–Larcker criterion. As shown in

Table 4. Discriminant validity of the model.

Variable	BI	CPB	TPB	PU	PEU	PI	GAIA	PER
BI	0.761
CPB	0.244	0.749
TPB	0.237	0.454	0.769
PU	0.708	0.723	0.443	0.846
PEU	0.686	0.269	0.494	0.776	0.843
PI	0.58	0.222	0.716	0.752	0.773	0.763
GAIA	−0.248	−0.011	0.04	−0.307	−0.265	−0.163	0.871
PER	0.017	−0.005	0.121	0.005	−0.058	0.102	0.496	0.771

Note: Diagonal elements are bolded to highlight the square roots of the AVEs. Off-diagonal elements are latent variable correlations. Discriminant validity is established when the diagonal value is higher than the values in its corresponding row and column.

, the square root of the AVE for each construct was greater than its correlations with all other constructs, indicating adequate discriminant validity and suggesting that the eight latent variables capture empirically distinct dimensions of teachers’ GenAI-related perceptions.

The overall fit of the measurement model was also examined. The CFA results indicated adequate, though not optimal, fit (CFI = 0.904, TLI = 0.881, IFI = 0.905, RMSEA = 0.077, and CMIN/df = 2.270; see

Table 5. Model fit test.

Indicators	Results	Reference Standards
CMIN/df	2.270	1–3 (excellent), 3–5 (good)
RMSEA	0.077	<0.05 (excellent), <0.08 (good)
IFI	0.905	>0.9 (excellent), >0.8 (good)
TLI	0.881	>0.9 (excellent), >0.8 (good)
CFI	0.904	>0.9 (excellent), >0.8 (good)

). RMSEA and the incremental indices CFI/IFI fall within commonly used “acceptable” ranges, whereas TLI is below the frequently cited 0.90 guideline, suggesting some residual misfit. We therefore interpret the measurement model as serviceable but improvable. We also note that relative fit indices such as TLI can be sensitive to sample size and model complexity, and should not be treated as a single decisive criterion (Hu & Bentler, 1999; Ximénez et al., 2022). Accordingly, we rely on a convergent evaluation across multiple indices. Taken together, other indices (CFI/IFI exceeding 0.90, RMSEA below 0.08, and CMIN/df below 3) support proceeding with structural path testing, while acknowledging that alternative theoretically grounded specifications may yield improved fit and should be examined in future work.

4.3. Structural Model Analysis

After confirming the adequacy of the measurement model, the structural model was estimated to test the fourteen hypothesized paths among the eight constructs. The final model, including standardized path coefficients and significance levels, is presented in Figure 2, and the detailed estimates are reported in

Table 6. Test of model path relationships.

Path			Estimate (β)	S.E.	C.R. (t)	p
BI	<---	PU	0.357	0.101	3.551	<0.001
BI	<---	PEU	0.357	0.105	3.397	<0.001
BI	<---	CPB	0.256	0.125	2.059	0.039
BI	<---	TPB	−0.139	0.05	−2.76	0.006
BI	<---	PER	0.069	0.067	1.038	0.299
BI	<---	GAIA	−0.048	0.056	−0.854	0.393
BI	<---	PI	0.035	0.116	0.304	0.761
PU	<---	CPB	0.275	0.135	2.043	0.041
PU	<---	TPB	0.05	0.054	0.926	0.354
PU	<---	PI	0.793	0.092	8.588	<0.001
PEU	<---	CPB	0.112	0.131	0.852	0.394
PEU	<---	TPB	0.087	0.055	1.566	0.117
PEU	<---	PI	0.875	0.097	8.985	<0.001
GAIA	<---	PER	0.486	0.094	5.169	<0.001

. The model explained a substantial proportion of variance in the endogenous constructs, as reflected in the R² values reported for behavioral intention, perceived usefulness, perceived ease of use, and GenAI anxiety in Figure 2.

4.3.1. Predictors of Teachers’ Intention to Adopt GenAI

Consistent with TAM, both perceived usefulness and perceived ease of use exerted significant positive effects on teachers’ behavioral intention to adopt GenAI in teaching (PU--->BI: β = 0.357, p < 0.001; PEU--->BI: β = 0.357, p < 0.001), supporting H1 and H2. These results suggest that teachers who believe that GenAI can effectively enhance teaching performance and who perceive it as easy to operate are more inclined to integrate GenAI into their instructional practice.

Pedagogical beliefs also played a significant role in shaping adoption intentions. Constructivist pedagogical beliefs had a significant positive effect on adoption intention (CPB--->BI: β = 0.256, p = 0.039), whereas transmissive pedagogical beliefs had a significant negative effect (TPB--->BI: β = −0.139, p = 0.006), thereby supporting H3 and H4. In other words, teachers who more strongly endorse student-centered, inquiry-oriented beliefs are more likely to adopt GenAI, whereas those who favor teacher-centered, content-transmission beliefs tend to be less willing to integrate GenAI into their teaching. This pattern aligns with the theoretical argument that pedagogical beliefs not only influence how teachers use technology but also whether they choose to adopt it.

By contrast, the direct effects of perceived ethical risks, GenAI anxiety, and perceived intelligence on BI were not statistically significant (ER--->BI: β = 0.069, p = 0.299; GAIA--->BI: β = −0.048, p = 0.393; PI--->BI: β = 0.035, p = 0.761). Accordingly, H9, H12, and H14 were not supported. These non-significant paths suggest that, in the current sample of Chinese K-12 teachers, adoption decisions are more strongly driven by evaluations of usefulness and ease of use and by pedagogical orientations than by direct effects of perceived intelligence, ethical risks, or anxiety. The implications of these findings are further discussed in Section 5.

4.3.2. Predictors of Perceived Usefulness and Ease of Use

With respect to perceived usefulness, both constructivist pedagogical beliefs and perceived intelligence showed significant positive effects (CPB--->PU: β = 0.275, p = 0.041; PI--->PU: β = 0.793, p < 0.001), providing support for H5 and H10. These findings indicate that teachers who hold constructivist beliefs and who perceive GenAI as highly intelligent are more likely to view GenAI as useful for enhancing teaching effectiveness. Conversely, the effect of transmissive pedagogical beliefs on perceived usefulness was not significant (TPB--->PU: β = 0.050, p = 0.354), and thus H7 was not supported.

For perceived ease of use, perceived intelligence again emerged as a key predictor, exerting a strong positive effect (PI--->PEU: β = 0.875, p < 0.001) and confirming H11. The paths from constructivist beliefs and transmissive beliefs to perceived ease of use were not significant (CPB--->PEU: β = 0.112, p = 0.394; TPB---> PEU: β = 0.087, p = 0.117), leading to the rejection of H6 and H8.

These results suggest that teachers’ judgments about how intelligent GenAI is are central to both their assessments of its usefulness and their perceptions of its ease of use, whereas pedagogical beliefs influence perceived usefulness but not perceived ease of use.

4.3.3. Relationship Between Perceived Ethical Risks and GenAI Anxiety

Perceived ethical risks had a significant positive effect on GenAI anxiety (PER--->GAIA: β = 0.486, p < 0.001), supporting H13. Teachers who perceived higher ethical risks, such as concerns regarding bias, misuse of data, or threats to academic integrity, also reported higher levels of anxiety about GenAI. However, as noted above, this heightened anxiety did not translate into a direct negative effect on behavioral intention in the structural model, so H14 was not supported.

4.4. Mediation Effects

Because PI significantly predicted PU and PEU, and PU/PEU were modeled as predictors of BI, we further tested whether PI influences BI indirectly via PU and PEU using a bias-corrected bootstrap procedure (5000 resamples).

Table 7. Bootstrapped Mediation Effects of PI on BI.

Effect Type	Path	Estimate	95%CI (Lower, Upper)	p	Interpretation
Indirect effect 1	PI --->PU---> BI	0.289	[0.000, 0.770]	0.050	Not supported
Indirect effect 2	PI ---> PEU ---> BI	0.229	[−0.173, 0.699]	0.205	Not supported
Direct effect	PI ---> BI	0.058	[−0.626, 0.800]	0.899	Not supported
Total effect	PI ---> BI	0.576	[0.392, 0.829]	0.001	Supported

reports the bootstrapped indirect, direct, and total effects with 95% CIs.

The results show that neither indirect effect is supported because the CIs include zero (PI--->PU--->BI: estimate = 0.289, 95% CI [0.000, 0.770], p = 0.050; PI--->PEU--->BI: estimate = 0.229, 95% CI [−0.173, 0.699], p = 0.205). The direct effect of PI on BI is also not supported (estimate = 0.058, 95% CI [−0.626, 0.800], p = 0.899). However, the total effect of PI on BI is significant (estimate = 0.576, 95% CI [0.392, 0.829], p = 0.001), indicating an overall relationship between PI and BI even though the hypothesized specific mediation pathways through PU and PEU are not empirically demonstrated in the bootstrap test.

4.5. Moderating Effects

To explore whether the structural relationships differ across teacher subgroups, multi-group SEM analyses were conducted for gender, teaching stage, educational background, and school location. Note that the subgroup of teachers with postgraduate qualifications was comparatively small (n = 40). Therefore, moderation findings involving educational background are interpreted as indicative patterns that warrant replication in larger, more balanced samples rather than as definitive population differences.

In the multi-group SEM, we applied a strict measurement invariance specification by constraining factor loadings (metric invariance), indicator intercepts (scalar invariance), and measurement residuals to be equal across groups. This specification therefore meets and exceeds the requirements of scalar invariance, which is the minimum condition for meaningful latent mean comparisons, and strengthens the comparability of constructs across groups when interpreting group differences in the subsequent multi-group analyses.

Table 8. Measurement invariance testing across groups (strict invariance: equal loadings, intercepts, and residuals).

Variable	Model Comparison	ΔX2	Δdf	p	Invariant
Gender	Measurement model VS. Unconstrained model	25.409	16	0.063	No
	Structural model VS. Unconstrained model	27.752	11	0.004	Yes
Teaching stage	Measurement model VS. Unconstrained model	24.554	16	0.078	No
	Structural model VS. Unconstrained model	22.958	11	0.018	Yes
Educational background	Measurement model VS. Unconstrained model	16.821	16	0.397	No
	Structural model VS. Unconstrained model	20.651	11	0.037	Yes

Note. Invariance testing followed a strict invariance procedure where factor loadings, indicator intercepts, and residuals were constrained to be equal across groups. The models compared include measurement invariance (configural + metric + scalar + residuals). Strict invariance ensures that group comparisons are meaningful for structural paths, but it is not required for comparing means (which was not the focus of this study).

presents the results of the cross-group invariance test. The measurement invariance tests for the three groups (gender, teaching grade, and educational background) did not reach a significant level (Δχ² = 25.41, 24.55, 16.82,, respectively; all p > 0.05), indicating that the factor loadings are invariant across different groups. However, the structural invariance tests for the three groups were all significant (Δχ² = 27.75, 22.96, 20.65, respectively; all p < 0.05), suggesting that there are significant inter-group differences in the path coefficients between latent variables. This finding supports the moderating roles of gender, teaching grade, and educational background in the core relationships of the model.

For each moderator, unconstrained and fully constrained models were compared using chi-square difference tests. The results indicated that three moderators exerted significant moderating effects on at least one hypothesized path, and the key effects are summarized in

Table 9. Moderating effect results.

Moderating Variables	Moderating Paths	Group 1	Group 2	t	p
Gender	PEU--->BI	Male (β = 0.017)	Female (β = 0.422)	−3.051	0.003
Teaching Level	CPB--->PU	Primary School (β = 0.413)	Middle School (β = −0.035)	2.753	0.006
Educational Background	CPB--->PU	Bachelor or Blow (β = 0.018)	Postgraduate or Above (β = 0.320)	−4.195	0.000

.

Gender significantly moderated the impact of perceived ease of use on behavioral intention (PEU--->BI; t = −3.051, p = 0.003). For female teachers, perceived ease of use had a substantial positive effect on adoption intention (β = 0.422), whereas for male teachers the effect was weak and non-significant (β = 0.017). This suggests that improvements in the usability and user-friendliness of GenAI tools may be particularly effective in fostering adoption among female teachers, while male teachers’ adoption decisions appear less sensitive to ease-of-use considerations.

The teaching stage moderated the relationship between constructivist pedagogical beliefs and perceived usefulness (CPB--->PU; t = 2.753, p = 0.006). Among primary school teachers, stronger constructivist beliefs were associated with higher perceived usefulness of GenAI (β = 0.413), whereas the effect was almost null among middle school teachers (β = −0.035). This pattern suggests that constructivist orientations are more readily translated into favorable evaluations of GenAI’s usefulness in primary school contexts, where student-centered, exploratory learning activities are more prevalent and may align more naturally with GenAI-supported instructional designs.

Educational background also moderated the CPB--->PU path (t = −4.195, p = 0.000). For teachers with a master’s degree or above, constructivist beliefs exert a significant positive impact on perceived usefulness (β = 0.320), whereas for teachers with a bachelor’s degree or below, the impact is weak and statistically insignificant (β = 0.018). These indicate that higher levels of academic qualification may enable teachers to more effectively leverage constructivist beliefs when evaluating the pedagogical value of GenAI.

5. Discussion

5.1. Summary of Key Findings

Four sets of findings stand out.

First, consistent with TAM, perceived usefulness (PU) and perceived ease of use (PEU) both exerted significant positive effects on behavioral intention (BI) to adopt GenAI. This confirms that, even in the context of advanced, conversational AI tools, classic TAM mechanisms remain central. Teachers are more inclined to integrate GenAI when they believe it will improve teaching performance and when they feel it is easy to use.

Second, teachers’ pedagogical beliefs emerged as powerful determinants of GenAI adoption. Constructivist pedagogical beliefs (CPB) positively predicted BI and PU, while transmissive pedagogical beliefs (TPB) negatively predicted BI but did not significantly affect PU or PEU. These findings suggest that teachers’ underlying conceptions of teaching and learning shape whether they view GenAI as pedagogically meaningful and whether they intend to use it.

Third, perceived intelligence (PI) of GenAI strongly predicted both PU and PEU, but the bootstrapped mediation analysis did not provide statistical support for the specific indirect pathways from PI to BI through PU or through PEU, and the direct effect of PI on BI was also non-significant. However, the total effect of PI on BI was significant. These results suggest that teachers who regard GenAI as more intelligent are more likely to perceive it as useful and easy to use, yet the hypothesized TAM-based mediation mechanism is not empirically confirmed in our data. Instead, PI appears to relate to adoption intention at the total-effect level, potentially through additional or more complex pathways beyond PU and PEU.

Fourth, perceived ethical risks (PER) significantly increased GenAI anxiety (GAIA), but neither PER nor GAIA had a significant direct effect on BI. Teachers who perceived higher ethical risks felt more anxious about GenAI, yet these concerns did not translate into lower adoption intentions in the structural model.

Finally, multi-group analyses showed that gender, teaching stage, and educational background moderated several key relationships: PEU was more strongly related to BI for female teachers, and the effect of constructivist beliefs on PU depended on both teaching stage and educational background.

5.2. Theoretical Implications

5.2.1. Extending TAM to GenAI in K-12 Settings

The strong, positive effects of PU and PEU on BI corroborate the robustness of TAM in the GenAI context and in K-12 settings. Prior studies have emphasized the centrality of PU and PEU for teachers’ adoption of various educational technologies (Granić & Marangunić, 2019; Scherer & Teo, 2019). The present findings extend this evidence base by demonstrating that, for highly complex and relatively novel technologies such as GenAI, teachers’ decisions still hinge on whether tools demonstrably support instructional goals and can be used without excessive effort (Cambra-Fierro et al., 2025; García de Blanes Sebastián et al., 2025).

However, it is important to interpret the PI mechanism cautiously. Although PI showed a moderate bivariate association with BI (Table 4), its direct effect on BI was not significant in the structural model, and the bootstrapped mediation analysis did not support the two specific indirect pathways via PU or via PEU (Table 7). In contrast, the total effect of PI on BI was significant. This pattern suggests that PI is not simply an “indirect-only” antecedent operating through classical TAM beliefs in a straightforward way. Rather, PI appears to be related to adoption intention at an overall level (e.g., Abdelkader, 2023; Moussawi et al., 2021), but the specific TAM channels tested here (PU and PEU separately) do not fully explain how that relationship is transmitted. One possibility is that PI influences BI through additional mediators not included in the current model (e.g., trust in GenAI outputs, perceived output quality, perceived compatibility with teaching routines, or normative expectations). Another possibility is that PI’s influence is distributed across multiple pathways simultaneously or operates conditionally, which may not be captured by testing the two specific indirect effects in isolation.

5.2.2. The Central Role of Pedagogical Beliefs

The strong positive effect of constructivist pedagogical beliefs on BI and PU, combined with the negative effect of transmissive beliefs on BI, highlights the importance of teachers’ pedagogical belief in the GenAI era. This is consistent with prior work indicating that teachers with constructivist orientations are more likely to adopt student-centered technologies and AI tools (Choi et al., 2023; Kim et al., 2013). The present study extends this line of research in several ways.

First, the descriptive results show that teachers simultaneously endorse relatively high levels of both CPB and TPB, suggesting that pedagogical beliefs are not dichotomous but layered and context-dependent. Chinese K-12 teachers may feel compelled to adopt transmissive practices in examination-oriented environments, even while endorsing constructivist ideals (S. H. Liu, 2011). Against this background, the positive association between CPB and both PU and BI suggests that constructivist orientations serve as a critical “bridge” that connects GenAI’ s capabilities to teachers’ visions of meaningful pedagogy. When teachers see learning as inquiry-driven and student-centered, they are better positioned to appreciate GenAI’ s potential to support exploration, differentiation, and co-creation of content (Cabero-Almenara et al., 2024; C. Cui & Li, 2024).

Second, the absence of significant paths from CPB and TPB to PEU indicates that pedagogical beliefs shape how teachers evaluate the value of GenAI (its usefulness), but not how easy they perceive it to be to use. Perceptions of ease of use appear to be driven more by technical features (captured by PI) than by pedagogical orientations (Can & Nguyen, 2025; Lewis et al., 2015). This differentiation clarifies the specific role of beliefs: they inform value judgements more than usability judgements.

Third, the negative effect of TPB on BI underscores that strong commitment to teacher-centered, content-delivery approaches may impede GenAI adoption, even when teachers do not necessarily perceive GenAI as difficult to use (Shankar et al., 2024). For teachers whose identity is tied to being the primary source of knowledge, GenAI’ s capacity to generate content and interact directly with students may be perceived as a threat to their instructional role (Reiss, 2021).

5.2.3. Ethical Risks, Anxiety, and the Complexity of Affective Responses

A particularly intriguing set of findings concerns the relationships among ethical risks, anxiety, and intention. Consistent with the Issue-Risk-Judgment (IRJ) framework (Tan, 2002) and recent AI research, perceived ethical risks significantly increased GenAI anxiety. Teachers who were more concerned about bias, data misuse, or academic integrity issues also reported higher anxiety levels, echoing earlier work on AI-related job insecurity and affective responses (J. Li & Huang, 2020; Xu et al., 2025).

However, neither PER nor GAIA exerted a significant direct negative effect on BI. This contrasts with some studies that have found risk perceptions and anxiety to be direct inhibitors of technology adoption (Chatterjee, 2020; Jatileni et al., 2024). Although this outcome did not statistically support the original hypothesis, these non-significant results should not be interpreted as evidence that ethical risks and anxiety are irrelevant. Instead, these null paths may indicate the following possibilities: (1) the effects operate through alternative mechanisms, such as indirect-only pathways; (2) the contextual forces offset or override risk-related deterrents; (3) the adoption dynamics differ by stage of diffusion; and (4) measurement or model specification issues attenuate the observed direct effects.

First of all, one plausible explanation is that PI, PER, and GAIA influence intention primarily indirectly through more proximal cognitive beliefs in technology acceptance, i.e., PU and PEU, rather than as independent direct inhibitors. This interpretation is consistent with the observation that GAIA is negatively associated with BI at the bivariate level but becomes non-significant once PU, PEU, and pedagogical beliefs enter the structural model. This may suggest that the variance attributable to affective deterrents may be absorbed by or transmitted through these proximal predictors. From this perspective, the null direct effects do not contradict prior literature so much as indicate that, in this sample and model configuration, anxiety and risk perceptions may shape intention via teachers’ judgments about utility and effort, rather than directly.

A second explanation concerns the institutional environment in which GenAI is promoted. In settings where GenAI is actively encouraged through national or local educational agendas, teachers may perceive that policy expectations and group norms outweigh personal reservations, leading to high stated intention even when anxiety remains (Farrukh Shahzad et al., 2025). In other words, teachers may express willingness to adapt to align with institutional priorities or professional norms while still holding ethical concerns internally (Bayaga, 2024). Importantly, because policy pressure and normative climate were not directly operationalized in this study, we present this as a testable hypothesis rather than a confirmed explanation.

Relatedly, teachers may prioritize perceived efficiency and instructional productivity gains, such as generating teaching materials, supporting lesson preparation, or assisting feedback (Barbieri & Nguyen, 2025). These perceived benefits could encourage teachers to proceed despite ethical reservations, thereby weakening the observable direct deterrent effect of risk and anxiety on intention in the structural model. Again, we emphasize that this is a plausible mechanism that should be directly tested in future work.

A third explanation is that our findings may reflect that, at an early diffusion stage, curiosity and perceived innovation benefits can dominate risk-related deterrents, making intention less sensitive to anxiety even when anxiety is present (Ivanov et al., 2024). Over time, as teachers gain more experience in use, ethical concerns may exert a stronger influence on sustained actual use.

Finally, the null paths also warrant consideration of methodological explanations. Restricted variance in anxiety or intention (e.g., relatively high intention among experienced users) can attenuate effects, and risk or anxiety may exhibit non-linear patterns that are not captured by linear direct paths. In addition, model misspecification is possible: anxiety and ethical risk may exert conditional effects (e.g., stronger among less experienced teachers), operate through omitted constructs such as trust, governance clarity, or perceived behavioral control, or appear primarily as indirect effects. Future studies could compare alternative theoretically grounded specifications, such as moderation by experience or inclusion of policy constructs, test non-linearities, and employ longitudinal designs to evaluate whether ethical concerns increasingly constrain continued use over time.

Overall, these findings suggest that ethical risk and anxiety should not be treated as “non-issues” simply because their direct paths to intention are non-significant. Rather, they appear to function as meaningful background conditions that shape teachers’ emotional readiness and highlight the need for concrete governance, training, and school-level norms to support responsible GenAI adoption.

5.2.4. Heterogeneity Across Teacher Subgroups

The moderation results indicate that GenAI acceptance mechanisms are not uniform. The same intervention may have different leverage depending on teacher backgrounds and instructional contexts. Note that given subgroup sample-size constraints, these differences are best viewed as hypothesis-generating patterns to inform targeted professional development and guide replication, rather than as definitive evidence of population-level moderation.

First, the stronger effect of PEU on BI among female teachers suggests that ease-of-use considerations are particularly salient for women, consistent with earlier findings in technology acceptance research (Scherer & Teo, 2019; Teo et al., 2015; Zhang et al., 2023). For female teachers, BI appears to be more sensitive to usability, as ease of use may function as a risk-management and workload-reduction heuristic, helping to minimize cognitive effort and instructional disruption in already demanding teaching environments (Teo et al., 2015; Qazi et al., 2022). For male teachers, BI may depend more on other considerations, such as perceived performance gains, institutional expectations, or professional identity-may play a more dominant role.

Second, educational background moderated the pathway from constructivist-belief to PU, suggesting that professional capital conditions whether constructivist aspirations can be translated into a concrete sense of GenAI’s instructional value. Teachers with higher qualifications may be better positioned to evaluate output quality, design prompts, and integrate GenAI into pedagogically coherent activities, whereas teachers with lower qualifications may be more likely to experience misalignment between constructivist ideals and the practical skills needed to realize those ideals through GenAI.

Third, the teaching stage moderated the pathway from constructivist-belief to PU. This finding is educationally interpretable because primary school contexts often provide greater flexibility for exploratory and student-centered learning activities, making it easier for teachers with constructivist orientations to perceive GenAI as instructionally useful. In contrast, in more exam-pressured settings, which are often concentrated at lower-secondary levels, teachers may experience tighter curricular pacing and assessment demands, limiting the perceived feasibility of constructivist, GenAI-supported activities even when such beliefs are endorsed (Chand, 2025; S. Li et al., 2022; Lin & Tan, 2025).

These findings suggest that one-size-fits-all models or intervention strategies may overlook the distinct underlying mechanisms at play among different groups of teachers. To foster the long-term integration of GenAI across diverse teaching populations, differentiated support strategies must be implemented to specifically address usability challenges, pedagogical adaptation issues, and structural constraints.

5.3. Practical Implications

The findings carry several practical implications for researchers, policymakers, school leaders, and GenAI developers.

First, since PU and PEU are central drivers of BI, professional development initiatives should move beyond generic introductions to GenAI and focus on concrete, subject- and grade-specific instructionally concrete cases, in order to clearly show how GenAI can improve lesson preparation, differentiation, formative assessment, and feedback (Blundell et al., 2025; Y. Cui & Li, 2025). At the same time, training should include hands-on practice that reduces perceived complexity and supports teachers in integrating GenAI into existing workflows (Al-Abdullatif, 2024).

Second, given the central role of CPB and TPB, professional learning should not treat GenAI as a purely technical innovation. Instead, it should create opportunities for teachers to critically reflect on how GenAI aligns with or challenges their beliefs about teaching and learning. For teachers with stronger transmissive orientations, facilitators might model how GenAI can support, rather than undermine, teacher authority and accountability (Zhou et al., 2025). For example, by using GenAI to generate alternative explanations, practice items, or visualizations that the teacher curates and contextualizes.

Third, developers should prioritize functionalities that make GenAI appear responsive, context-aware, and pedagogically aligned (Şimşek et al., 2025). However, this should be coupled with transparency about limitations, typical failure modes (e.g., AI hallucinations), and appropriate verification strategies so that heightened perceptions of intelligence do not lead to uncritical reliance (Samala et al., 2024).

Fourth, even though ER and GAIA did not directly reduce BI, the strong PER--->GAIA path indicates emotional and ethical concerns. School systems and policymakers should develop clear guidelines on responsible GenAI use in teaching, including data protection, plagiarism, academic integrity, and teacher-student roles (Hazzan-Bishara et al., 2025; Jayasinghe et al., 2025). Professional development can incorporate scenario-based discussions that help teachers rehearse how to handle ethical dilemmas, which may mitigate anxiety and foster more confident, reflective use.

Finally, moderation findings imply that supports should be differentiated. Usability-focused onboarding may be especially impactful for teachers for whom ease-of-use is most predictive of intention, while teachers in exam-pressured stages may need examples that fit constrained pacing and assessment demands, and teachers with lower academic qualifications may benefit from foundational support in prompt literacy and pedagogical integration strategies.

5.4. Limitations and Future Research Directions

Several limitations should be acknowledged when interpreting these findings.

First, the study employed a cross-sectional design and relied on self-reported data of behavioral intention rather than actual usage data. For instance, although CPB and TPB demonstrate acceptable reliability and discriminant validity, it remains possible that response styles (e.g., acquiescence) or socially desirable responding may inflate mean levels. In addition, the intention-behavior literature shows that intentions typically account for only a modest share of variance in actual behavior (Armitage & Conner, 2001), highlighting the need for caution when translating the results into claims about realized classroom practice. Future research should therefore adopt longitudinal and observational designs to examine how intentions translate into sustained GenAI use and how teachers’ perceptions and belief systems evolve as GenAI tools become more embedded in school systems. At the same time, researchers should triangulate teacher beliefs using balanced-keyed scales, social-desirability or acquiescence controls, and complementary qualitative or classroom-observation evidence, and should test measurement invariance across regions and experience groups.

In addition, despite anonymity and reverse-scored items, the single-source, cross-sectional self-report design may still introduce CMV and inflate observed relationships. Harman’s single-factor test suggests that CMV is unlikely to be dominated by a single general factor, but this test alone cannot exclude more subtle method effects. Future studies should incorporate stronger procedural remedies (e.g., temporal separation, collecting predictors and outcomes from different sources, or combining surveys with objective indicators) and consider established statistical controls such as marker-variable techniques or latent method-factor approaches.

Second, the sample consisted of Chinese K-12 teachers, most of whom were from urban schools and had prior GenAI experience. While this reflects a population at the forefront of GenAI adoption in education, the findings may not generalize to rural regions, other national contexts, or teachers with limited exposure to GenAI. Future research should use stratified or quota-based sampling to ensure adequate representation across urban–rural contexts and should deliberately recruit non-users and low-exposure teachers to model initial adoption barriers and the transition from non-use to first-time classroom integration.

An additional limitation concerns subgroup analyses. Although multi-group SEM provides useful insight into heterogeneity, moderation tests can be underpowered and parameter estimates may be unstable when subgroup sizes are small. Therefore, the detected subgroup differences should be interpreted as suggestive rather than conclusive. Future research should recruit larger and more balanced subgroup samples, replicate the moderation findings across independent datasets, and consider complementary approaches (e.g., interaction models or cross-validation) to test the robustness of heterogeneous effects.

Third, the non-significant direct effects of PER and GAIA on BI may reflect range restriction or omitted mechanisms. Although PER and GAIA demonstrated satisfactory reliability and validity in the measurement model and GAIA was negatively correlated with BI at the bivariate level, the multivariate model suggests that their unique effects may be indirect, threshold-based, or contingent on context and experience. Future research should test alternative specifications that allow PER and GAIA to influence PU/PEU and other proximal beliefs, examine potential non-linear effects, and evaluate measurement equivalence across user-experience groups and institutional contexts.

Relatedly, although PER significantly increased GAIA, the model explains a relatively modest portion of variance in GenAI anxiety (R² = 0.175). This is not unexpected given that anxiety is typically shaped by multiple individual, contextual, and experiential factors beyond perceived ethical risks. Accordingly, PER should be interpreted as a meaningful but partial driver of GAIA in our model. Future research should extend the anxiety formation model by incorporating additional predictors such as individual differences, technological self-efficacy, or situational support. Testing these factors may substantially improve explanatory power for GAIA and clarify how risk perceptions translate into affective readiness over time.

Finally, the present study focused on direct and moderating effects. Future work could test more complex models, including more mediation pathways and interactions between ethical concerns and pedagogical beliefs. Such models would further clarify the mechanisms through which GenAI’s technical characteristics, pedagogical orientations, and affective responses jointly shape adoption. A qualitative follow-up can help explain why certain relationships are absent or subgroup-specific.

Moreover, although perceived ethical risk significantly predicted GenAI anxiety, its explanatory power was limited, suggesting that future research should incorporate a wider range of predictors to more fully capture the formation and evolution of teachers’ GenAI-related anxiety.

6. Conclusions

This study examined the factors shaping Chinese K-12 teachers’ intention to adopt GenAI in teaching by extending the Technology Acceptance Model with pedagogical beliefs, perceived intelligence, perceived ethical risks, GenAI anxiety, and key demographic moderators. Using survey data from 218 in-service teachers, the study offers an integrated account of how teachers evaluate GenAI’s role in classroom practice.

The findings confirm that perceived usefulness and perceived ease of use remain the primary drivers of adoption intention, even in the context of advanced GenAI tools. At the same time, adoption is strongly shaped by teachers’ pedagogical belief systems. Perceived intelligence functions as a GenAI-specific antecedent that strengthens usefulness and ease-of-use judgments rather than directly influencing intention. Ethical risks significantly heighten GenAI anxiety, yet neither exerts a direct negative effect on intention, suggesting that teachers are cautiously willing to experiment despite underlying concerns. The study contributes to GenAI-in-education research by showing that technology-related beliefs, pedagogical orientations, GenAI-specific perceptions, and contextual characteristics must be considered together when supporting K-12 teachers’ meaningful and responsible use of GenAI.