AI literacy-related domains and AI-TPACK readiness among preservice mathematics teachers: A factor-informed structural equation modelling study

Synthesis: AI literacy-related domains and AI-Technological Pedagogical Content Knowledge (TPACK) readiness among preservice mathematics teachers: A factor-informed structural equation modeling study

Key Findings

• In an exploratory CFA/SEM study of 130 preservice mathematics teachers at a South African public university, the AI-TPACK readiness items were essentially unidimensional; one weak design-confidence item was removed, and the refined seven-item measurement model fitted better than the original eight-item specification.
• The removed item — confidence in designing AI-supported geometry lessons — had an EFA loading of .294 and communality of .087, indicating it shared too little variance with the broader readiness factor; the refined measurement model fit well, χ2(474) = 601.36, p < .001, CFI = .991, TLI = .990, RMSEA = .046, SRMR = .082.
• The primary gender-controlled latent SEM (n=129) showed good approximate fit (CFI = .981, TLI = .984, RMSEA = .049, SRMR = .082; χ2(602) = 789.92, p < .001) and explained 53.0% of the variance in AI-TPACK readiness.
• Positive associations with readiness were observed for prior AI use, critical-ethical appraisal, and support/enablers; the support/enablers path had the largest standardized coefficient (β = .725, p = .024) but must be interpreted cautiously because the construct had marginal AVE (.478) and overlapped with information-source engagement (r = .748).
• Prior AI use (β = .327, p < .001) and critical-ethical appraisal (β = .251, p = .016) were the other significant positive paths; AI awareness, information-source engagement, tool familiarity, contextual barriers, prior AI training, and gender did not retain significant independent paths in the joint model.
• Year level was significant in the primary model (β = .199, p = .041) but less stable in sensitivity analysis, so claims about progression across cohorts remain provisional; a merged "AI-literacy core" model explained less variance (R² = .460) than the refined model.
• Discriminant-validity evidence for the broader AI-literacy-related domains was mixed — five constructs passed the Fornell-Larcker check, but information-source engagement and support/enablers showed the strongest overlap (r = .748, then r = .708 with awareness) — so those domains were treated as theoretically informed predictors rather than fully validated independent latent variables.
• Scale reliabilities were acceptable (Cronbach's alpha: AI awareness .868, tool familiarity .841, readiness eight-item .844, contextual barriers .770, critical-ethical appraisal .755, support/enablers .745, information-source engagement .738).
• Overall, readiness was associated with direct AI experience and critical-ethical judgment, while the contribution of support/enablers remains provisional.

Study Design & Method

The study is described as factor-informed: AI-literacy-related domains were treated as theoretically informed and empirically tested predictors of AI-TPACK readiness rather than as fully validated independent constructs. Exploratory factor analysis using polychoric correlations supported the unidimensionality of the readiness items, after which a structural equation model was estimated with gender as a control (n = 129 because one participant selected "prefer not to say" for gender). The refined structural model (χ2(602) = 789.92, CFI = .981, TLI = .984, RMSEA = .049, SRMR = .082, R² = .530) outperformed the original eight-item specification, which showed weaker approximate fit (CFI = .957, TLI = .962, RMSEA = .075, SRMR = .094, R² = .539); sensitivity checks confirmed the stability of the support/enablers, prior-AI-use, and critical-ethical-appraisal paths while year level became non-significant. The sample of 130 preservice mathematics teachers comes from a public university in South Africa, making the study a contribution to AI-TPACK research in a Global South teacher education context.

What this means for practice

• Faculty developers. Prioritize direct, hands-on AI use over awareness-raising: prior AI use was a significant predictor of readiness (β = .327, p < .001), while AI awareness, tool familiarity, and prior AI training retained no significant independent paths in the joint model.
• Faculty developers. Teach critical-ethical appraisal of AI tools as a core part of preparation, because it was the other significant positive predictor (β = .251, p = .016), and build programs that combine direct AI experience with critical evaluation.
• Faculty developers. Do not treat seniority as readiness: year level was significant in the primary model (β = .199, p = .041) but became non-significant in sensitivity analysis, so progression across cohorts remains provisional.
• Faculty developers. Use the refined short readiness instrument for efficient diagnostics, but report it as one unidimensional readiness score rather than separate AI-literacy sub-scores, since information-source engagement and support/enablers failed discriminant validity (r = .748) and information-source engagement had weak average variance extracted.
• Faculty developers. Do not conclude that support and enabling conditions are irrelevant: the support/enablers path was the largest (β = .725, p = .024) but its marginal AVE (.478) means the contribution of the teacher education setting itself cannot be isolated, and the non-significant contextual-barriers path only shows that infrastructure added no independent variance within this model.

Limitations

• The design is cross-sectional and based entirely on self-report, so structural paths should not be interpreted causally.
• The sample came from one university context and was modest for a complex latent-variable model, and the EFA-informed refinement, CFA, discriminant-validity diagnostics and SEM were all conducted on the same N = 130 dataset rather than split into development and validation samples; lavaan also produced near-singular variance-covariance warnings in some models.
• Measurement evidence was not uniformly strong — information-source engagement had weak AVE and support/enablers was marginal, with mixed discriminant validity — so the broader AI-literacy domains should not be treated as fully validated dimensions.
• The study did not measure cognitive load, classroom performance, observed AI use or longitudinal development, leaving open whether the identified predictors translate into actual pedagogical practice.

Citation

Mosia, M., Nannim, F. A., & Egara, F. (2026). AI literacy-related domains and AI-TPACK readiness among preservice mathematics teachers: A factor-informed structural equation modelling study.