Exploring the effects of AI literacy in teacher learning: an empirical study

The study addresses the gap that many K-12 teachers lack understanding of AI and feel unprepared to use or teach AI, which hampers effective AI integration in education. Drawing on the Theory of Reasoned Action (TRA), Theory of Planned Behavior (TPB), and the Technology Acceptance Model (TAM), the authors note that these models have rarely been applied to AI education and have neglected constructs such as literacy and ethical awareness. Teachers in China have reported anxiety about AI’s complexity and reluctance to learn AI, underscoring the need to examine determinants of teachers’ behavioral intentions to learn AI. The purpose is to propose and validate a model integrating AI literacy and awareness of AI ethics with TRA/TPB to explain K-12 teachers’ intentions to learn AI, thereby informing future research and professional development design.

The literature review integrates TRA and TPB with AI-specific constructs and develops hypotheses. Behavioral intentions to learn AI are conceptualized following TRA as teachers’ willingness to learn what constitutes AI and its application in teaching. Perceptions of the use of AI for social good (PAIS) are positioned as a key attitudinal facet; if teachers perceive AI benefits to society, they should be more motivated to learn AI (H1: PAIS → BI). Self-efficacy in learning AI reflects perceived behavioral control per TPB and is hypothesized to predict intentions (H2: SE → BI) and to influence attitudes (H3: SE → PAIS). AI literacy is treated as an epistemic factor influencing attitudinal and control beliefs (H4: AIL → PAIS; H5: AIL → SE) and indirectly intentions via PAIS (H6) and SE (H7). Awareness of AI ethics is argued to be positively related to AI literacy (H8: AIL → AAIE) and to affect attitudes (H9: AAIE → PAIS). Additional mediated effects are proposed: AIL → PAIS via AAIE (H10) and AAIE → BI via PAIS (H11). A conceptual model summarizes these relationships.

Design: Cross-sectional survey with two-step structural equation modeling (SEM). Participants: 318 K-12 teachers from 16 provinces/municipalities in China (final valid N=318 of 339 responses). Demographics included 38.68% male, 61.32% female; stages across primary, junior high, senior high; ages 18–60; 67.92% undergraduate degrees; 57.55% STEM majors; 66.98% urban. All participants used an AI-based product (Zhixue by iFLYTEK) in teaching. Recruitment: Random selection of ~50 partner schools and ~8 teachers per school; email invitations with online questionnaire link; ethics approval obtained. Instruments: Scales adapted for teachers—Awareness of AI ethics (four sub-dimensions: transparency, responsibility, justice, sustainability; 3 items each) from Lin et al. (2021) and Shih et al. (2021); AI literacy (4 items), self-efficacy in learning AI (4 items), PAIS (5 items), behavioral intention to learn AI (4 items) from Chai et al. (2021). Translation via forward/back-translation; expert consultation; pilot test and EFA-based revisions. Seven-point Likert (1=strongly disagree, 7=strongly agree). Data analysis: Two-step SEM using AMOS 21. First, confirmatory factor analysis (CFA) to validate measurement models, including first- and second-order modeling of AI ethics awareness. Three problematic items (SE2, SG5, BI1) were removed to improve fit. Then, structural model estimation tested direct and indirect effects. Model fit criteria followed Byrne (2010), Hu & Bentler (1999). Control variables tested: gender, school stage, age, school district, education background, major.

Measurement model: First-order CFA for AI ethics sub-dimensions (12 indicators) showed acceptable fit: χ²=157.223, df=48, RMSEA=0.085, CFI=0.965, TLI=0.952. Second-order factor (AAIE) fit: χ²=176.537, df=50, RMSEA=0.089, CFI=0.960, TLI=0.947. Revised overall measurement model after removing SE2, SG5, BI1: χ²=856.382, df=285, RMSEA=0.080, CFI=0.927, TLI=0.917. Convergent/discriminant validity supported: loadings 0.73–0.95; CR>0.70; AVE>0.50; Fornell–Larcker criteria met. Structural model fit (Fig. 3): χ²=958.646, df=366, RMSEA=0.071, CFI=0.925, TLI=0.917. Hypotheses: All 11 hypotheses supported. Direct effects: H1 PAIS → BI β=0.62, CR=9.037, p<0.001; H2 SE → BI β=0.29, CR=4.557, p<0.001; H3 SE → PAIS β=0.55, CR=7.625, p<0.001; H4 AIL → PAIS β=0.18, CR=2.406, p=0.016; H5 AIL → SE β=0.77, CR=12.238, p<0.001; H8 AIL → AAIE β=0.62, CR=10.063, p<0.001; H9 AAIE → PAIS β=0.22, CR=4.249, p<0.001. Indirect effects (Sobel tests): H6 AIL → BI via PAIS χ²=2.779, p=0.002; H7 AIL → BI via SE χ²=4.066, p<0.001; H10 AIL → PAIS via AAIE χ²=4.210, p<0.001; H11 AAIE → BI via PAIS χ²=3.835, p<0.001. Variance explained: SE 59%; AAIE 46%; PAIS 70%; BI 75%. Controls: Age positively impacted AAIE (β=0.24, p<0.001, CR=5.102); Major negatively impacted AAIE (β=−0.13, p=0.006, CR=−2.733); School district positively impacted PAIS (β=0.08, p=0.029, CR=2.177). Other controls were not significant. Overall, PAIS and SE are immediate predictors of intentions to learn AI; AIL and AAIE influence intentions indirectly; AIL is a key exogenous driver affecting SE, PAIS, and AAIE.

Findings confirm, in the context of K-12 teachers’ AI learning, the TRA/TPB proposition that attitudes (PAIS) and perceived behavioral control (self-efficacy in learning AI) directly shape behavioral intentions to learn AI. Importantly, the study integrates AI literacy and awareness of AI ethics into TRA/TPB, showing these as indirect antecedents of intentions: AI literacy strongly enhances self-efficacy and PAIS and raises awareness of AI ethics; awareness of AI ethics, in turn, positively influences attitudes (PAIS). The results clarify debates about the role of awareness in attitude formation by demonstrating a significant path from ethics awareness to attitudes toward AI for social good. Practically, enhancing teachers’ AI literacy emerges as central to boosting readiness to learn AI, as it cascades through self-efficacy, ethics awareness, and pro-social attitudes. Demographic/contextual influences (age, major, school district) further suggest tailoring professional development (e.g., addressing discipline differences, and rural–urban exposure) to strengthen PAIS and ethics awareness.

This study provides early empirical evidence on the mechanisms through which AI literacy shapes K-12 teachers’ behavioral intentions to learn AI. PAIS and self-efficacy in learning AI directly predict intentions, while AI literacy and awareness of AI ethics indirectly affect intentions. AI literacy directly increases self-efficacy, PAIS, and awareness of AI ethics and indirectly increases intentions through these mediators. The model explains 75% of the variance in intentions, underscoring its explanatory power. The work extends TRA/TPB by incorporating literacy and ethical awareness and highlights the importance of professional development that prioritizes AI literacy, addresses AI ethics, and fosters perceptions of AI’s social good. Future research could broaden the model (e.g., subjective norms), explore regional differences, and triangulate data sources.

Four limitations are acknowledged: (1) Limited regional coverage within China may affect generalizability; regional disparities in AI resources and opportunities suggest the need for regional comparisons. (2) Reliance on self-reported, subjective data may introduce bias; future studies should use multiple data sources. (3) Although explanatory power is strong, the model can be expanded (e.g., include subjective norms) to improve explanation. (4) Three instrument items (SE2, SG5, BI1) were not validated in this context, possibly due to cultural interpretation differences; further investigation is needed.