• 1. RQ1. What baseline-adjusted achievement contrast was observed between teacher-mediated agentic AI-supported instruction and conventional instruction in Grade 10 functions and graphs?

• 2. RQ2. How did the modelled probability of reaching meaningful mathematical mastery thresholds differ between the two instructional conditions?

• 3. RQ3. How did the baseline-adjusted achievement contrast vary across learners’ baseline achievement levels?

• 4. RQ4. How did changes in learners’ mathematical interest or confidence align with changes in achievement?

The present study is informed by an integrated theoretical framework that combines Technological Pedagogical Content Knowledge (TPACK) with perspectives from mathematics education on representational learning. Rather than treating artificial intelligence as an autonomous instructional agent, the framework conceptualises AI as a teacher-mediated resource whose educational value depends on how teachers integrate technological capability with pedagogical reasoning and mathematical content knowledge. The intervention is therefore understood as a process of pedagogical orchestration rather than technological substitution. From the perspective of TPACK (Mishra & Koehler, 2006), effective technology integration occurs when teachers successfully coordinate knowledge of mathematics, pedagogy, and technology. The framework rejects the assumption that technological sophistication alone improves learning. Instead, AI-generated explanations, examples, visualisations, and formative questions become educationally meaningful only when teachers evaluate their mathematical accuracy, align them with curriculum goals, adapt them to learners’ needs, and integrate them into coherent instructional sequences. Teacher expertise, therefore, remains the central mechanism through which technology contributes to learning.

Theories of mathematical representation complement this pedagogical perspective. Learning functions and graphs requires learners to coordinate symbolic, graphical, numerical, and contextual representations rather than mastering isolated procedures (Leinhardt et al., 1990). Contemporary mathematics education research similarly demonstrates that conceptual understanding develops through purposeful engagement with multiple external representations, dynamic visualisations, and teacher-guided interpretation of mathematical relationships (Schoenherr & Schukajlow, 2024; Schoenherr et al., 2024). From this perspective, AI has educational value when it expands teachers’ capacity to generate diverse representations and multiple explanatory pathways that support conceptual reasoning rather than procedural imitation. The framework, therefore, proposes that teacher-mediated AI operates through an indirect instructional mechanism. Rather than interacting directly with learners, the AI system expands the range of curriculum-aligned explanations, worked examples, graphical demonstrations, formative questions, and misconception-sensitive prompts available to the teacher. Teachers then exercise professional judgement by selecting, adapting, sequencing, and contextualising these resources within classroom instruction. Learning improvements are therefore expected to emerge not because AI replaces teaching, but because it enriches the instructional repertoire through which teachers support mathematical understanding.

The study used a quasi-experimental pre-test/post-test comparison design to estimate the effect of teacher-mediated AI-supported instruction on learners’ achievement in Grade 10 functions and graphs. This design was appropriate because the intervention was implemented under intact classroom conditions, where the educational object was not a decontextualised software exposure but a teacher’s use of an AI system during the teaching of a topic. Preserving that instructional form meant that learners could not be treated as if they had been individually randomised to equivalent conditions. The design, therefore, required an analytic strategy that compared groups cautiously, adjusted for measured baseline differences, and stated clearly the assumptions under which the comparison could be interpreted causally. The analysis followed the logic of causal comparison under observed covariate adjustment. Rather than asking only whether two intact groups differed descriptively at post-test, the study estimated the expected difference in post-test achievement between comparable learners under the AI-supported and conventional instructional conditions. The causal reading is therefore conditional rather than automatic: it depends on the measured baseline variables, the temporal ordering of pre-test and post-test measures, overlap between groups, and the modelling assumptions made explicit below.

The analytic sample consisted of 115 Grade 10 mathematics learners with complete pre-test and post-test records for the focal achievement and interest measures. Fifty-eight learners were taught in the teacher-mediated AI-supported condition, while fifty-seven were taught in the conventional instruction condition. The learners came from the same socio-economic status and school-quintile context, and the intervention and comparison conditions covered the same mathematical topic: functions and graphs. The study involved two intact classes, one in each instructional condition, and both followed the same eight-week study schedule: one week of pre-intervention assessment, six weeks of classroom instruction, and one week of post-intervention assessment. Two different teachers delivered the two instructional conditions; both teachers were trained for the study and used the same lesson plan. These features strengthen procedural and contextual comparability, but they do not turn the study into an individually randomised experiment or separate the instructional condition from teacher and class context. For analysis, instructional condition was represented as a binary treatment indicator, with teacher-mediated AI-supported instruction treated as the focal condition and conventional instruction as the comparison condition. Gender was included as a measured learner characteristic because it was recorded before the post-test outcome and could reasonably be related to classroom participation, confidence, or prior opportunity. Baseline achievement was specified as a pre-intervention covariate to ensure that the models compared learners who were more alike in their starting points on the topic.

3.2.1 Instructional procedure

The intervention followed a structured instructional sequence designed to ensure that both instructional conditions received equivalent curriculum coverage, teaching time, and learning objectives. The study was implemented over eight weeks, comprising one week of pre-intervention assessment, six weeks of classroom instruction, and one week of post-intervention assessment. During the first week, all learners completed the Functions and Graphs Achievement Test (FGAT) and the Functions and Graphs Interest Inventory (FGII) to establish baseline achievement and interest before instruction. These baseline measures were subsequently incorporated into the Bayesian models as pre-intervention covariates.

The instructional phase extended across six consecutive weeks during the normal school timetable. Mathematics was taught for five 45-minute periods each week, totalling approximately 22.5 hours of instruction across the intervention. Both instructional conditions followed the same Grade 10 CAPS curriculum sequence, used identical learning objectives, and addressed the same mathematical concepts. The instructional progression comprised: (a) introduction to functions and relationships between symbolic, numerical, graphical, and contextual representations; (b) linear functions; (c) quadratic functions; (d) hyperbolic functions; (e) exponential functions; and (f ) revision, consolidation, and integrated problem-solving activities. One week after completing the instructional programme, learners in both conditions completed the FGAT and FGII post-tests. This interval reduced immediate recall effects while allowing learners time to consolidate their understanding before the outcome measures were administered.

3.3.1 Intervention boundary and instructional object

The first issue in describing Mothusi is one of the intervention boundary, because the platform was capable of more than the study asked it to do. In this phase, the teacher assigned to the AI-supported condition was the sole classroom user of Mothusi: learners did not log in, submit prompts, or receive responses from the system on personal devices, and the learner-facing functionality that existed elsewhere in the wider platform was not activated as part of the treatment. What the study evaluates, therefore, is not the unrestricted capability of the product, but a deliberately narrower instructional arrangement in which one teacher used the system to extend the planning, explanation, representation, and questioning resources available throughout the six-week classroom instruction phase of the functions-and-graphs unit. Within that boundary, Mothusi was neither an autonomous tutor nor a replacement for the mathematical judgement carried by the teacher. It was used as a teacher-mediated AI support system configured for Grade 10 CAPS functions and graphs, through which the teacher could request alternative explanations, worked examples, graph demonstrations, formative questions, and prompts organised around common learner misconceptions. The treatment object was thus a teacher-AI-curriculum configuration: Mothusi enlarged the repertoire from which the teacher could work, while the teacher determined what became instruction, in what sequence, and with what explanations.

3.3.2 System configuration and curriculum grounding

The implementation comprised an authenticated teacher workspace connected to OpenClaw, which maintained the Mothusi persona and session context, routed common graphing and introductory functions requests through preconfigured response and visualisation routines, and passed requests requiring model generation to the configured language model; during the principal May implementation, those model-mediated requests were routed to GPT-5.5. This agentic arrangement did not mean that the system acted independently in the classroom. It meant that the workspace could coordinate a bounded set of resources and tools in response to a teacher’s request, while the decision to use, change, or reject the resulting material remained with the teacher.

The system was curriculum-bound rather than an invitation to generate mathematics freely. Its local, page-indexed corpus contained the DBE Grade 10 Annual Teaching Plan as curriculum anchor, the Siyavula Grade 10 mathematics text as the primary explanatory source, selected Mind the Gap material for extension, a functions-and-graphs topic index, nine misconception patterns with corrective prompts, and a 24-item formative question bank. Model-mediated content requests were configured to consult this corpus before producing a substantive response. At the same time, SymPy supported algebraic checking, Matplotlib for static graph production, KaTeX for mathematical notation, and Desmos for interactive graphs. These technical provisions widened what the teacher could call upon; they did not, by themselves, establish the mathematical or pedagogical suitability of what entered a lesson.

Before implementation commenced, both participating teachers completed a structured professional development programme designed specifically for the study. The programme familiarised teachers with the lesson sequence, curriculum objectives, instructional protocol, and the intended pedagogical role of Mothusi within classroom instruction. Particular emphasis was placed on maintaining teacher authority over mathematical explanations by critically evaluating AI-generated outputs, verifying mathematical correctness, selecting appropriate representations, and adapting generated material to learners’ instructional needs. This preparation ensured that AI functioned as a teacher-support tool rather than as an autonomous instructional agent.

3.3.3 Teacher enactment of AI-generated support

At the point of use, the AI component was located in the generation and coordination of candidate instructional material rather than solely in the graphing display. A teacher request entered through the authenticated workspace; OpenClaw combined that request with the Mothusi persona and session context, consulted the curriculum-bounded corpus, and, where a generated response was required, routed the request to GPT-5.5 before returning a candidate explanation, example, question sequence, or graph-supported activity. SymPy, Matplotlib, KaTeX, and Desmos could support the mathematical checking and representation of that response, but these tools were not themselves the AI intervention. What distinguished Mothusi from conventional graphing software was the model-mediated production and coordination of curriculum-grounded teaching support in response to the teacher’s request.

What entered the classroom, however, still travelled through the teacher. During preparation, the teacher used these generated candidates to inspect possible explanations, examples, questions, and visual representations; during instruction, selected material could be displayed, revoiced, slowed down, or reorganised in response to the class. The teacher checked the mathematics, decided whether a representation served the immediate lesson, and placed questions around it so that the system’s capacity to produce an answer did not displace the learner’s mathematical work. The substantive intervention was therefore the teacher’s mediated use of AI-generated, curriculum-grounded support, not exposure to a platform interface and not independent learner interaction with a chatbot.

Throughout instruction, teacher mediation followed a consistent pedagogical sequence. Before each lesson, the teacher used Mothusi to generate candidate explanations, worked examples, graphical demonstrations, misconception-sensitive prompts, and formative questions aligned with the planned curriculum objectives. These outputs were critically reviewed for mathematical correctness, curriculum alignment, and pedagogical suitability before classroom use. During lessons, the teacher selected appropriate AI-generated resources, integrated them with direct instruction, facilitated classroom discussion, posed probing questions, responded to learner misconceptions, and determined the pace and sequencing of instructional activities. Learners therefore experienced AI only through teacher-mediated explanations, visual representations, and guided mathematical reasoning rather than through direct interaction with the system. Accordingly, references in this paper to the Mothusi intervention denote only this teacher-mediated configuration. They do not denote the learner login, independent learner chat, researcher dashboard, or other platform functions developed for separate phases of the wider programme.

The comparison condition was conventional instruction on the same functions-and-graphs topic. This condition should not be read as the absence of teaching; it represents the ordinary instructional arrangement against which the teacher-mediated AI-supported condition was compared. The contrast is therefore between two ways of teaching the topic: conventional instruction and instruction in which the teacher used Mothusi to extend the planning, explanation, representation, and questioning resources available during the topic. Because the comparison condition covered the same curriculum topic, followed the same lesson plan, and was implemented over the same eight-week study schedule (including six weeks of classroom instruction), the design reduces the risk that the observed difference reflects topic exposure or instructional time rather than instructional form. The shared socio-economic and school-quintile context also reduces one broad source of contextual imbalance. The remaining interpretive burden, however, cannot be carried by these design features alone. It is carried by baseline adjustment, explicit causal assumptions, and cautious interpretation of the quasi-experimental comparison, with robustness checks used only to examine model-specification fragility rather than to remove teacher, class, or implementation confounding.

3.4.1 Implementation fidelity

Several procedures were implemented to maximise intervention fidelity across instructional conditions. Both teachers followed identical lesson plans, addressed the same curriculum objectives, taught the same mathematical content during the same instructional period, and used the same assessment instruments. Throughout the intervention, the researcher monitored implementation through scheduled classroom observations and discussions with participating teachers to verify adherence to the instructional protocol. Where minor deviations occurred, these were addressed immediately to ensure consistency with the planned intervention. Although fidelity procedures strengthened procedural consistency, they could not eliminate the inherent teacher- and class-level confounding associated with the intact-class quasi-experimental design.

The primary outcome was the post-test total score on the Functions and Graphs Achievement Test (FGAT). The corresponding FGAT pre-test total was used as the baseline achievement covariate. The FGAT comprised 25 multiple-choice items aligned with the Grade 10 functions-and-graphs topic, each worth 2 marks on the classroom test form. For analysis, scores were recorded on a 0–100 scale. The achievement test was treated as the most direct measure of whether learners performed better after instruction on the target mathematical topic. Because the study concerns functions and graphs, the achievement outcome bears the main inferential burden of the paper. The secondary outcome was the Functions and Graphs Interest Inventory post-test total (FGII post-test), adjusted for the FGII pre-test total. The FGII was a 30-item four-category inventory with response options ranging from strong agreement to strong disagreement. Items addressed enjoyment, value, confidence or competence, and engagement with functions and graphs; negatively worded items were coded so that higher total scores indicated more favourable interest-and-confidence responses. Internal consistency evidence from the main sample was strong for the total score, with Cronbach’s alpha of 0.883. The FGII outcome was treated as secondary because the paper’s central claim is about mathematical achievement and mastery rather than affective change alone.

Ethical clearance for the larger study was granted by the General/Human Research Ethics Committee of the researchers’ institution. The study used learner-level educational data for research purposes and is reported without identifying the school, teacher, or individual learners. The analysis dataset contains only pseudonymous learner identifiers. The research team holds institutional consent and school-permission documents.

The dataset was prepared before modelling by retaining learners with complete information on instructional condition, gender, FGAT pre-test, FGAT post-test, FGII pre-test, and FGII post-test. The pre-test and post-test totals were treated as continuous scores, with higher values representing stronger achievement or more favourable interest-and-confidence responses, depending on the instrument. Gain scores were computed for descriptive reporting, but the causal estimates were based on post-test models that adjusted directly for the corresponding pre-test measure. Baseline FGAT and FGII scores were standardised before modelling. This step made the treatment coefficient interpretable at the average baseline level and made the interaction between baseline achievement and instructional condition easier to interpret. No post-test outcome information was used to define the instructional condition; the treatment indicator referred only to the instructional arrangement under which learners were taught.

The causal estimand was the average difference in post-intervention FGAT achievement expected under teacher-mediated AI-supported instruction compared with conventional instruction, after adjustment for measured baseline variables. In the potential-outcomes language associated with Rubin and Holland, this is the contrast between a learner’s potential post-test outcome under the AI-supported condition and that learner’s potential post-test outcome under the conventional condition. Because each learner is observed in only one condition, the alternative outcome is necessarily unobserved and must be estimated rather than directly measured. For this contrast to be interpreted causally, the analysis assumes temporal ordering, consistency, conditional exchangeability, overlap, limited cross-condition contamination, measurement validity, and approximate model adequacy. These assumptions require that baseline measures precede the outcome, that observed outcomes correspond to the instructional condition received, that comparable learners exist across instructional conditions, and that no major unmeasured confounder simultaneously determines instructional condition and post-test achievement after adjustment for measured baseline variables. In this study, conditional exchangeability is a strong assumption because the design used two intact classes, one per condition, taught by two different trained teachers, even though both classes followed the same eight-week study schedule, used the same lesson plan, and came from the same socio-economic and school-quintile context. The data cannot prove these assumptions. They define the conditions under which the estimates can be read as causal rather than as baseline-adjusted associations.

The primary FGAT analysis used a Bayesian baseline-adjusted Gaussian model. The post-test FGAT score was modelled as a function of instructional condition, standardised baseline FGAT, gender, and a treatment-by-baseline interaction. The treatment coefficient estimated the additional FGAT post-test points expected under the AI-supported instructional condition at the average baseline achievement level. The interaction term examined whether the treatment effect increased or decreased across the baseline achievement distribution. Weakly informative priors were used: the intercept was centred on the observed post-test mean with a standard deviation of 12; the treatment effect had a Normal(0, 15) prior; the baseline, gender, and interaction effects had Normal priors with standard deviations of 8, 5, and 6, respectively; and the residual standard deviation had a half-Normal(8) prior.

The secondary FGII model followed the same baseline-adjusted logic but did not include the treatment-by-baseline interaction. FGII post-test total was modelled as a function of treatment, standardised FGII pre-test, and gender. The FGII priors were also weakly informative: the intercept was centred on the observed FGII post-test mean with a standard deviation of 10; the treatment effect had a Normal(0, 8) prior; the baseline and gender coefficients had Normal priors with standard deviations of 5 and 4, respectively; and the residual standard deviation had a half-Normal(5) prior. The FGII analysis was specified as a secondary analysis rather than as the main test of the intervention.

Models were estimated in Python using PyMC, and posterior summaries were produced with ArviZ. Each reported Bayesian model used four Markov chains, 1,000 tuning draws, and 1,000 posterior draws per chain, with a target acceptance probability of 0.90. The analysis reports posterior means, posterior standard deviations, 95% credible intervals, and posterior probabilities in the direction of educational interest. Convergence and sampling quality were assessed using R-hat, effective sample sizes, and visual diagnostics.

To express the baseline-adjusted contrast in educationally interpretable terms, the analysis translated the FGAT model’s posterior predictions into probabilities of reaching achievement thresholds of 70, 80, 90, and 95. These thresholds were not treated as universal standards; rather, they were used as meaningful cut points for describing movement into stronger regions of functions-and-graphs performance. For each threshold, the model estimated the expected probability of crossing the threshold under the AI-supported condition and under conventional instruction, and then estimated the posterior distribution of the probability difference. The heterogeneity analysis estimated the baseline-adjusted contrast at selected points in the baseline FGAT distribution: the 10th, 25th, 50th, 75th, and 90th percentiles. This made it possible to ask whether the observed separation was concentrated among learners who began the topic with lower or higher initial achievement. Because the study used two intact classes, these estimates were interpreted as descriptive variation in the modelled contrast rather than as subgroup-specific causal effects.

Several additional checks tested whether the achievement pattern depended on one modelling choice. First, baseline comparability was examined using raw differences, standardised differences, and common-support checks for the FGAT pre-test. Second, frequentist ANCOVA models were fitted as cross-checks, including a specification matching the Bayesian covariates, a no-interaction specification, a specification adding baseline FGII, and a trimmed sample excluding extreme FGAT pre-test values. Third, posterior mean fitted values from the Bayesian model were used to examine residual size by condition. These checks do not replace the Bayesian estimand and do not address the teacher-and-class confounding created by the two-class design; their role is narrower: to test whether the descriptive achievement pattern is fragile to a single analytic specification.

The first empirical issue is whether the two groups began from broadly comparable achievement baselines before diverging after instruction. Table 1 shows that the conventional instruction group had a mean FGAT pre-test score of 61.70. In contrast, the teacher-mediated agentic AI-supported group had a mean of 62.02, a raw baseline difference of only 0.32 points. After the intervention, however, the mean post-test score was 63.81 in the conventional condition and 93.12 in the AI-supported condition, producing a raw difference of 29.31 points. The same pattern appears in gain scores: learners in the conventional condition gained 2.11 points on average, while learners in the AI-supported condition gained 31.10 points.

The baseline comparisons indicate that the instructional groups were similar on measured pre-intervention characteristics, supporting the use of baseline-adjusted comparisons while recognising that equivalence cannot be established in an intact-class quasi-experimental design. The standardised difference on the FGAT pre-test was 0.04, the standardised difference on the FGII pre-test was −0.18, and the standardised difference for the proportion of female learners was 0.19. The common-support check was also reassuring: the shared FGAT pre-test range was 46 to 75, which included all conventional learners and 98.3% of learners in the AI-supported condition. These checks justify a baseline-adjusted comparison while still limiting the causal interpretation to the two-class, two-teacher design. The Bayesian baseline-adjusted model shows that the achievement difference was not merely a raw-score pattern. As Table 2 reports, the estimated baseline-adjusted between-condition contrast was 29.12 additional FGAT post-test points for the teacher-mediated agentic AI-supported class, with a 95% credible interval from 27.31 to 30.92. The posterior probability that the fitted contrast was greater than zero was 1.00. The estimated treatment coefficient therefore, represents the expected difference in FGAT post-test performance between instructional conditions, after adjustment for measured baseline variables and gender, under the specified Bayesian model.

The second research question moves from average achievement to the educational size of the observed separation. A large mean difference is important, but teachers and mathematics educators also need to know whether learners are likely to move into higher-performing regions. Table 3 shows that the modelled probability of reaching 70 or above was effectively 1.00 in the teacher-mediated agentic AI class, compared with 0.125 in the conventional class. At the 80-point threshold, the modelled probability was 0.995 for the AI-supported condition and 0.001 for the conventional condition. Even at the more demanding 90-point threshold, the modelled probability was 0.732 in the AI-supported condition and approximately zero in the conventional condition.

The posterior estimates indicate consistently higher probabilities of reaching each predefined mastery threshold under the teacher-mediated agentic AI condition than under conventional instruction. The estimated probability differences were greatest for the 80-point and 90-point thresholds, where the posterior distributions showed clear separation between instructional conditions. Posterior probabilities that the threshold differences exceeded zero were 1.00 for all reported thresholds.

The third research question examines whether the achievement separation was concentrated among learners with particular baseline profiles. The treatment-by-baseline interaction was negative in the FGAT model, suggesting that the estimated contrast became somewhat smaller as baseline achievement increased. That pattern, however, should not be misread as evidence that only lower-achieving learners showed the separation. Table 4 shows that the estimated contrast remained strongly positive across the observed baseline distribution: 32.49 points at the 10th percentile of baseline achievement, 28.82 points at the median, and 27.25 points at the 90th percentile.

The estimated treatment contrast remained positive across the observed baseline achievement distribution. Although the estimated contrast declined modestly as baseline achievement increased, the posterior distributions consistently favoured the teacher-mediated AI condition across all reported baseline percentiles.

The fourth research question treats learners’ mathematical interest and confidence as secondary outcomes rather than as substitutes for achievement. Table 5 presents the Bayesian baseline-adjusted estimates for the Functions and Graphs Interest Inventory (FGII) together with the correlations between gains in achievement and gains in interest and confidence. The descriptive pattern remained favourable to the teacher-mediated AI-supported condition. The conventional group declined by 1.88 points on the FGII total, while the AI-supported group increased by 1.78 points. The Bayesian baseline-adjusted estimate for the FGII post-test total was 3.23 points in favour of the AI-supported condition, with a 95% credible interval from 2.25 to 4.19 and a posterior probability greater than zero of 1.00.

The relation between achievement gain and FGII gain was more limited than the group-level pattern might suggest. Across the full sample, the gain correlation was moderate and positive, r = .426, p < .001, n = 115. Within conditions, however, the correlations were near zero: r = −.015, p = .913, n = 57 in the conventional condition and r = .059, p = .662, n = 58 in the AI-supported condition. Collectively, the FGII analyses indicate higher post-intervention interest and confidence scores in the teacher-mediated AI condition, along with a moderate positive association between achievement and FGII gains across the full sample. Within-condition gain correlations, however, remained close to zero. Sampling diagnostics were adequate (R-hat = 1.00), but the estimates remain conditional on the quasi-experimental assumptions stated above.

The robustness checks had a narrower purpose than causal validation. They asked whether the central achievement estimate was fragile to a small set of modelling choices, not whether the intact-group comparison was free of confounding. The Bayesian model remained the primary model because it corresponded to the stated estimand and reported uncertainty as a credible interval. The frequentist ANCOVA models were used as specification checks: one matched the Bayesian covariates, one removed the treatment-by-baseline interaction, one added baseline FGII as an additional covariate, and one trimmed extreme FGAT pre-test values. As Table 6 shows, this sequence did not materially change the achievement estimate, which remained between 29.10 and 29.54 FGAT points. That stability reduces concern about a single narrow modelling artefact, but all checks still use the same intact-condition comparison.

Posterior fitted values indicated satisfactory model performance, with an overall RMSE of 4.74 FGAT points and mean residuals centred close to zero for both instructional conditions. Across all robustness analyses, the estimated treatment contrast remained stable, indicating that the principal findings were not materially altered by the alternative model specifications examined.

The anonymised participant-level dataset supporting the findings of this study has been deposited in Zenodo and is publicly available at https://doi.org/10.5281/zenodo.21525667. (Dlamini, et al., 2026).

The repository contains the anonymised learner-level data used in the analyses, including Functions and Graphs Achievement Test (FGAT) scores, Functions and Graphs Interest Inventory (FGII) scores, instructional condition, and gender. All personal identifiers have been removed to protect participant confidentiality.

The data are available under the terms of the Creative Commons Attribution 4.0 International (CC BY 4.0) licence.

The authors thank the participating schools, teachers, learners, and colleagues whose support contributed to the successful completion of this study.

•  Bloom BS: The 2 sigma problem: The search for methods of group instruction as effective as one-to-one tutoring. Educ. Res. 1984; 13(6): 4–16. Publisher Full Text
•  Cheung ACK, Slavin RE: The effectiveness of educational technology applications for enhancing mathematics achievement in K-12 classrooms: A meta-analysis. Educ. Res. Rev. 2013; 9: 88–113. Publisher Full Text
•  Chiu TKF, Xia Q, Zhou X, et al.: Systematic literature review on opportunities, challenges, and future research recommendations of artificial intelligence in education. Computers and Education: Artificial Intelligence. 2023; 4: 100118. Publisher Full Text
•  Darvishi A, Khosravi H, Sadiq S, et al.: Impact of AI assistance on student agency. Comput. Educ. 2024; 210: 104967. Publisher Full Text
•  Deng R, Jiang M, Yu X, et al.: Does ChatGPT enhance student learning? A systematic review and meta-analysis of experimental studies. Comput. Educ. 2025; 227: 105224. Publisher Full Text
•  Dlamini LV, Mosia M, Egara FO: Dataset for: Teacher-Mediated, Not Learner-Facing: A Bayesian Quasi-Experimental Study of Agentic AI Support in Grade 10 Functions and Graphs. [Data set]. Zenodo. 2026. Publisher Full Text
•  Drijvers P, Sinclair N: The role of digital technologies in mathematics education: purposes and perspectives. ZDM - Mathematics Education. 2024; 56(2): 239–248. Publisher Full Text
•  Engelbrecht J, Borba MC: Recent developments in using digital technology in mathematics education. ZDM - Mathematics Education. 2024; 56: 281–292. Publisher Full Text
•  Ertmer PA: Addressing first- and second-order barriers to change: Strategies for technology integration. Educ. Technol. Res. Dev. 1999; 47(4): 47–61. Publisher Full Text
•  Hew KF, Brush T: Integrating technology into K-12 teaching and learning: Current knowledge gaps and recommendations for future research. Educ. Technol. Res. Dev. 2007; 55(3): 223–252. Publisher Full Text
•  Hill JL: Bayesian nonparametric modeling for causal inference. J. Comput. Graph. Stat. 2011; 20(1): 217–240. Publisher Full Text
•  Hillmayr D, Ziernwald L, Reinhold F, et al.: The potential of digital tools to enhance mathematics and science learning in secondary schools: A context-specific meta-analysis. Comput. Educ. 2020; 153: 103897. Publisher Full Text
•  Holland PW: Statistics and causal inference. J. Am. Stat. Assoc. 1986; 81(396): 945–960. Publisher Full Text
•  Huang AYQ, Lu OHT, Yang SJH: Effects of artificial intelligence-enabled personalised recommendations on learners’ learning engagement, motivation, and outcomes in a flipped classroom. Comput. Educ. 2023; 194: 104684. Publisher Full Text
•  Hwang G-J, Xie H, Wah BW, et al.: Vision, challenges, roles and research issues of artificial intelligence in education. Computers and Education: Artificial Intelligence. 2020; 1: 100001. Publisher Full Text
•  Kasneci E, Sessler K, Kuchemann S, et al.: ChatGPT for good? On opportunities and challenges of large language models for education. Learn. Individ. Differ. 2023; 103: 102274. Publisher Full Text
•  Khosravi H, Shum SB, Chen G, et al.: Explainable artificial intelligence in education. Computers and Education: Artificial Intelligence. 2022; 3: 100074. Publisher Full Text
•  Kulik JA, Fletcher JD: Effectiveness of intelligent tutoring systems. Rev. Educ. Res. 2016; 86(1): 42–78. Publisher Full Text
•  Lai JWM, Bower M: How is the use of technology in education evaluated? A systematic review. Comput. Educ. 2019; 133: 27–42. Publisher Full Text
•  Leinhardt G, Zaslavsky O, Stein MK: Functions, graphs, and graphing: Tasks, learning, and teaching. Rev. Educ. Res. 1990; 60(1): 1–64. Publisher Full Text
•  Li Q, Ma X: A meta-analysis of the effects of computer technology on school students’ mathematics learning. Educ. Psychol. Rev. 2010; 22(3): 215–243. Publisher Full Text
•  Liu M, Zhang LJ, Biebricher C: Investigating students’ cognitive processes in generative AI-assisted digital multimodal composing and traditional writing. Comput. Educ. 2024; 211: 104977. Publisher Full Text
•  Lo CK, Hew KF, Jong MS-Y: The influence of ChatGPT on student engagement: A systematic review and future research agenda. Comput. Educ. 2024; 219: 105100. Publisher Full Text
•  Long D, Magerko B: What is AI literacy? Competencies and design considerations. Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems. 2020; 1–16. Publisher Full Text
•  Ma W, Adesope OO, Nesbit JC, et al.: Intelligent tutoring systems and learning outcomes: A meta-analysis. J. Educ. Psychol. 2014; 106(4): 901–918. Publisher Full Text
•  Medina Herrera LM, Juarez Ordonez S, Ruiz-Loza S: Enhancing mathematical education with spatial visualisation tools. Frontiers in Education. 2024; 9: Article 1229126. Publisher Full Text
•  Mishra P, Koehler MJ: Technological pedagogical content knowledge: A framework for teacher knowledge. Teachers College Record: The Voice of Scholarship in Education. 2006; 108(6): 1017–1054. Publisher Full Text
•  Ng DTK, Leung JKL, Chu SKW, et al.: Conceptualising AI literacy: An exploratory review. Computers and Education: Artificial Intelligence. 2021; 2: 100041. Publisher Full Text
•  Ouyang F, Jiao P: Artificial intelligence in education: The three paradigms. Computers and Education: Artificial Intelligence. 2021; 2: 100020. Publisher Full Text
•  Roll I, Wylie R: Evolution and revolution in artificial intelligence in education. Int. J. Artif. Intell. Educ. 2016; 26(2): 582–599. Publisher Full Text
•  Rosenbaum PR, Rubin DB: The central role of the propensity score in observational studies for causal effects. Biometrika. 1983; 70(1): 41–55. Publisher Full Text
•  Rubin DB: Estimating causal effects of treatments in randomised and nonrandomised studies. J. Educ. Psychol. 1974; 66(5): 688–701. Publisher Full Text
•  Schoenherr J, Schukajlow S: Characterising external visualisation in mathematics education research: A scoping review. ZDM. 2024; 56: 73–85. Publisher Full Text
•  Schoenherr J, Strohmaier AR, Schukajlow S: Learning with visualisations helps: A meta-analysis of visualisation interventions in mathematics education. Educ. Res. Rev. 2024; 45: 100639. Publisher Full Text
•  Steenbergen-Hu S, Cooper H: A meta-analysis of the effectiveness of intelligent tutoring systems on college students’ academic learning. J. Educ. Psychol. 2014; 106(2): 331–347. Publisher Full Text
•  Tamim RM, Bernard RM, Borokhovski E, et al.: What forty years of research says about the impact of technology on learning. Rev. Educ. Res. 2011; 81(1): 4–28. Publisher Full Text
•  ten Voorde A , Piroi M, Bos R: A taxonomy of didactic roles of dynamic visualisation in animated mathematics videos. Teaching Mathematics and its Applications: An International Journal of the IMA. 2025; 44(1): 47–67. Publisher Full Text
•  Tondeur J, van Braak J , Sang G, et al.: Preparing pre-service teachers to integrate technology in education: A synthesis of qualitative evidence. Comput. Educ. 2012; 59(1): 134–144. Publisher Full Text
•  VanLehn K: The relative effectiveness of human tutoring, intelligent tutoring systems, and other tutoring systems. Educ. Psychol. 2011; 46(4): 197–221. Publisher Full Text
•  Voogt J, Fisser P, Pareja Roblin N, et al.: Technological pedagogical content knowledge: A review of the literature. J. Comput. Assist. Learn. 2013; 29(2): 109–121. Publisher Full Text
•  Zawacki-Richter O, Marin VI, Bond M, et al.: Systematic review of research on artificial intelligence applications in higher education: Where are the educators?. Int. J. Educ. Technol. High. Educ. 2019; 16(1): Article 39. Publisher Full Text
•  Zhang Y, Wang P, Jia W, et al.: Dynamic visualisation by GeoGebra for mathematics learning: A meta-analysis of 20 years of research. J. Res. Technol. Educ. 2025; 57(2): 437–458. Publisher Full Text
•  Zhong B, Liu X: Evaluating AI literacy of secondary students: Framework and scale development. Comput. Educ. 2025; 227: 105230. Publisher Full Text