mathematics-education
Filtering by topic mathematics-education(2)Clear all filters
- PaperComputers and Education: Artificial Intelligence17 Jul 2026
AI literacy-related domains and AI-TPACK readiness among preservice mathematics teachers: A factor-informed structural equation modelling study
Moeketsi Mosia, Fadip Audu Nannim, Felix Egara
A study of 130 preservice mathematics teachers in South Africa found that AI-TPACK readiness was largely unidimensional, with prior AI use, critical-ethical appraisal, and support/enablers positively associated. The structural equation model explained 53% of variance in readiness, though discriminant validity among AI literacy domains was mixed.
Original abstract
In this factor-informed exploratory CFA/SEM study, AI-literacy-related domains were treated as theoretically informed and empirically tested predictors of AI-TPACK readiness rather than as fully validated independent latent variables. Artificial intelligence (AI) is increasingly entering mathematics education, making it important to understand how preservice teachers become ready to integrate AI-supported tools pedagogically. This study examined AI-TPACK readiness among 130 preservice mathematics teachers at a South African public university. Exploratory factor analysis using polychoric correlations indicated that the AI-TPACK readiness items were essentially unidimensional; one weak design-confidence item was removed. The refined seven-item measurement model fitted better than the original eight-item specification, although discriminant-validity evidence for the broader AI-literacy-related domains was mixed. The primary gender-controlled latent SEM (sample n = 129) showed good approximate fit, χ 2 (602) = 789.92, p < .001, CFI = .981, TLI = .984, RMSEA = .049, SRMR = .082, and explained 53.0% of the variance in AI-TPACK readiness. Positive associations were observed for prior AI use, critical-ethical appraisal, and support/enablers. The support/enablers path had the largest standardised coefficient, but should be interpreted cautiously because the construct had marginal AVE and overlapped with information-source engagement. Year level was significant in the primary model but less stable in sensitivity analysis. Overall, the findings suggest that readiness was associated with direct AI experience and critical-ethical judgement, while the contribution of support/enablers remains provisional. The study contributes a cautious empirical account of AI-TPACK readiness in a Global South teacher education context.
- PaperComputers and Education: Artificial Intelligence14 Jul 2026
Balancing AI responsibility with privacy, safety, and utility: Unlearning in large language models for mathematics education
Chenglu Li, Gökhan Gülfidan, Yinqi Zhang-Kopf
This study applies gradient-based LLM unlearning to reduce personally identifiable information (PII) and harmful content in math tutoring models while maintaining performance on math tasks. Results show substantial decreases in PII and harm rates without sacrificing utility, demonstrating a path toward responsible AI in education.
Original abstract
Online mathematics learning platforms are increasingly adopting large language models (LLMs) to provide scalable, on-demand support, but these models may reproduce private information from training data or generate harmful language. This raises concerns about responsibility in educational settings regarding the use of pre-trained models. LLM unlearning is an emerging area for reducing a model’s ability to produce specific unwanted content and remains underexplored in educational research. This study aims to investigate how LLM unlearning reduces the model's reliance on personally identifiable information (PII) and inappropriate content in the math tutoring context, while maintaining the model's utility on both single-label and multi-label downstream math tasks. We applied a gradient-based LLM unlearning approach to three different models, which were pre-trained on approximately 3 million data points from an Algebra I online discussion forum between students and professional tutors. PII and harmful content were detected on this training data and used for unlearning in two different orders (PII and harmful content unlearning). Then, the generated outputs from these two unlearning models were compared with those of the pre-trained model in terms of PII-containing output rate and harmful rate. Moreover, unlearned models were evaluated on two different math classification tasks. The results showed that the rates of PII-containing output rate and harmfulness substantially decreased compared to the pre-trained models, and the utility of the unlearned model was still maintained. These findings demonstrate how LLM unlearning can be applied to pre-trained models to behave them more responsibly, while maintaining strong model performance on math-related tasks.