AI-Enhanced Visualization in Secondary Mathematics: A Systematic Review and Meta-Analysis of Achievement and Conceptual Understanding, with a Focus on Functions and Graphs

This project is a systematic review and meta-analysis of whether artificial intelligence (AI) that works with mathematical visual representations improves secondary school students’ mathematics achievement and conceptual understanding. Graphs, diagrams, dynamic geometry and augmented reality are central to learning secondary mathematics. AI can now generate or adapt such visuals, or interpret and respond to students’ work on them, as when intelligent tutoring systems check plotted graphs or generative AI produces diagrams. Existing reviews of intelligent tutoring systems and educational technology do not separate the contribution of the AI from that of the visualization itself. This review addresses that gap. We include experimental, quasi-experimental and crossover studies with a comparison group, conducted with students in lower or upper secondary education (ISCED 2–3). Interventions are classified by how the AI acts on the visual: Tier 1: the AI generates or adapts the visual. Tier 2: the AI evaluates the learner’s work on the visual. Tier 3: the AI is used alongside a visual tool; these studies are used for sensitivity analysis only. We answer two questions separately. The first is the effect of AI-enhanced visualization compared with instruction without it. The second is what the visual component adds within AI tutors, using comparisons of the same tutor with and without its visual. Functions and graphs form a pre-specified subgroup. ERIC, Web of Science and Lens.org were searched in October 2026 (2,278 records). Two reviewers screened records and assessed full texts and risk of bias (RoB 2, ROBINS-I), with disagreements reconciled. AI tools assisted both reviewers, and every decision was verified against the source. This registration is made after searching, screening and risk-of-bias assessment, but before any effect sizes are pooled. It fixes the synthesis plan in advance. Because few eligible studies exist, the primary synthesis is a structured narrative synthesis (SWiM). Pooled random-effects estimates (REML with Hartung–Knapp–Sidik–Jonkman intervals and prediction intervals) are exploratory, and studies at critical risk of bias enter only sensitivity analyses. Certainty of evidence is rated with GRADE. Expected outputs are: a peer-reviewed article; the review’s data, decisions and analysis code on OSF; a clear account of how strong the current evidence is and where better-designed studies are needed.