<![CDATA[Artificial intelligence (AI) is increasingly embedded in mathematics instruction, yet the evidence base is fragmented across intelligent tutoring systems, adaptive platforms, learning analytics, robotics, and generative AI. This umbrella review synthesizes review-level evidence on learning effects, pedagogical roles, and implementation risks without treating technological novelty as evidence of effectiveness. OpenAlex was searched on 1 August 2026 using 12 reproducible title queries. The publication window was fixed at 1 January 2021–31 December 2025 to avoid partial-year bias. Systematic, scoping, mapping, and meta-analytic reviews explicitly identified as such in their titles were eligible. Of 542 records, 466 remained after deduplication; 51 reports were assessed and 11 reviews were included. Because primary-study overlap and outcome heterogeneity precluded defensible re-pooling, findings were synthesized narratively and appraised using a seven-item methodological-transparency rubric. Two meta-analyses reported small average effects on mathematics achievement (Hedges’ g = 0.351 for elementary learners and g = 0.343 for K–12 learners). Across reviews, intelligent tutoring and adaptive systems had the most mature evidence, especially for feedback, practice, and personalization. Evidence for generative AI was newer and concentrated on short-term performance, perceptions, task design, and demonstrations rather than durable conceptual learning. Recurring risks included inaccurate mathematical output, learner over-reliance, privacy and bias, weak teacher preparation, unequal infrastructure, and limited longitudinal evidence. AI should therefore be implemented as a teacher-governed instructional component, with curriculum alignment, independent verification of outputs, data safeguards, accessibility alternatives, and monitoring of differential outcomes.]]>
