Assessing AI-TPACK readiness in mathematics teacher education: The role of self-efficacy and teaching beliefs

Synthesis: Xie and Luo (2026) develop and validate a domain-specific AI-Technological Pedagogical Content Knowledge (TPACK) instrument tailored to the unique pedagogical and logical demands of Math Education, then survey 412 Chinese mathematics Professional Development students (MTES) across seven universities (289 female, 123 male; 262 senior undergraduates, 105 first-year and 45 second-year graduate students). Their AI-TPACK readiness is currently at a preliminary stage, with all six construct means falling in the medium range (3.76–5.33 on 7-point scales): teaching beliefs were highest (M = 5.24) and AI-TK lowest (M = 4.23). A structural equation model shows that Self-Efficacy is a significant positive predictor of AI-TK (β = 0.69), AI-TCK (β = 0.78), and AI-TPK (β = 0.73), whereas strong traditional teaching beliefs act as a weak cognitive barrier (negative paths to AI-TCK and AI-TPK), and AI-TK's influence on AI-TPACK is fully mediated through AI-TPK and AI-TCK. The findings provide empirical evidence for redesigning mathematics teacher training to address both technical proficiency and psychological readiness.

Key Findings

Domain-specific instrument. A 24-item AI-TPACK scale across six dimensions (AI-TK, AI-TCK, AI-TPK, AI-TPACK, self-efficacy, teaching beliefs) was developed following DeVellis & Thorpe's framework, starting from a 35-item pool rated on 7-point Likert scales and screened by a five-expert panel (three educational-technology professors, two mathematics-education specialists) that removed or rephrased five ambiguous items. EFA on a pilot sample of 128 MTES from one university (KMO = .930, Bartlett's test significant, p < .001; item-to-participant ratio 1:4.3) then CFA confirmed the structure, with final factor loadings of .485–.863, 84.62% total variance explained, and Cronbach's alphas of .883 (AI-TK), .917 (AI-TPK), .919 (AI-TCK), .930 (AI-TPACK), .909 (self-efficacy), and .894 (teaching beliefs).

Preliminary readiness. The final survey of 412 MTES (item-to-participant ratio ≈ 1:17, satisfying the 10–20-observations-per-parameter rule) revealed AI-TPACK readiness at a preliminary stage: teaching beliefs were highest (M = 5.24) and AI-TK lowest (M = 4.23), with AI-TCK (M = 4.53) the highest AI-TPACK component. AI use was largely "consumption-oriented" (searching answers, designing exam questions, drafting basic lesson plans) rather than advanced (e.g., VR for 3D geometry, intelligent student-observation systems), which the authors link to the tension between heuristic exploration and exam-oriented performance in Chinese mathematics education.

No grade-level differences in AI-TPACK. ANOVA found no significant differences across senior undergraduates and first- and second-year graduate students for AI-TK (F = 1.79, p = .170), AI-TCK (F = 2.14, p = .119), AI-TPK (F = 0.37, p = .708), or AI-TPACK (F = 0.83, p = .435), although self-efficacy (F = 4.68, p = .010) and teaching beliefs (F = 7.26, p < .001) did differ; teaching experience also did not significantly enhance AI-TPACK. The authors attribute the stagnation to fragmented, informal internet-based learning and limited AI exposure during internships.

Self-efficacy as predictor. In the final AI-TPACK-SEM (χ²/df = 3.35, RMSEA = 0.078, NFI = 0.915, CFI = 0.938, TLI = 0.927), self-efficacy strongly and positively predicted AI-TK (β = 0.69), AI-TCK (β = 0.78), and AI-TPK (β = 0.73), supporting all H1 paths and reflecting a psychological catalyst that reduces perceived complexity of AI tools.

Teaching beliefs as cognitive barrier. Strong traditional teaching beliefs showed weak negative associations with AI-TCK (β = −0.03) and AI-TPK (β = −0.15), contrary to the H2 hypotheses, interpreted via second-order barriers and the conflict between traditional mathematical rigor and AI's perceived unpredictability. AI-TK did not directly predict AI-TPACK but was mediated through AI-TPK (β = 0.30) and AI-TCK (β = 0.35), which in turn predicted overall AI-TPACK (β = 0.59 and 0.64) — consistent with Ouyang et al.'s "know-how/know-why/know-how-to-teach" synthesis.

Implication. Mathematics teacher training must be redesigned to address both technical proficiency and psychological readiness (self-efficacy and beliefs), embedding AI-pedagogical training continuously across the four-year curriculum rather than as a single elective, and linking to the knowledge base's Professional Development and Math Education concepts. Limitations include cross-sectional self-reported data with potential social-desirability bias, restriction to the Chinese mathematics-education context, and exploratory pruning of non-significant paths (H2b, H3c) pending larger-sample validation.

What this means for practice

• Faculty developers. Embed AI-pedagogical training continuously across the four-year curriculum rather than offering it as a single elective, because grade level and teaching experience did not raise AI-TPACK in this sample.
• Faculty developers. Target Self-Efficacy directly with scaffolded hands-on AI use, since it was the strongest positive predictor of AI-TK (β = 0.69), AI-TCK (β = 0.78), and AI-TPK (β = 0.73).
• Faculty developers. Move training past consumption-oriented uses such as searching for answers and drafting basic lesson plans toward advanced applications, because that is where the surveyed pre-service teachers were weakest (AI-TK M = 4.23, versus AI-TCK M = 4.53).
• Researchers. Treat traditional teaching beliefs as a barrier to name and work on, since strong belief in mathematical rigor showed weak negative paths to AI-TCK (β = −0.03) and AI-TPK (β = −0.15).
• Researchers. Use the 24-item domain-specific instrument to diagnose which of the six dimensions is weak before designing an intervention, and report the mediation structure — AI-TK reached AI-TPACK only through AI-TPK (β = 0.30) and AI-TCK (β = 0.35).

Limitations

• The 412 pre-service mathematics teachers were surveyed at a single time point, so the structural paths, including self-efficacy's large coefficients, are correlational and the authors call for longitudinal work.
• Every measure is self-report on 7-point Likert scales, which the authors identify as vulnerable to social-desirability bias in which participants overstate technical readiness.
• The instrument was developed and tested only in the Chinese mathematics-education context, and the authors state that the high-stakes, rigor-oriented curriculum there limits direct application to other disciplines or cultures.
• Two hypothesized paths (H2b and H3c) were non-significant and pruned from the final model, so the account of teaching beliefs as a cognitive barrier rests on exploratory respecification pending larger-sample validation.

Citation

Xie, M., & Luo, L. (2026). Assessing AI-TPACK readiness in mathematics teacher education: The role of self-efficacy and teaching beliefs. Computers and Education Open, 100375. https://doi.org/10.1016/j.caeo.2026.100375