AI integration in mathematics education: A systematic review of implementation challenges and pedagogical implications (2017-2025)

Synthesis: Manto et al. (2026) systematically review 54 Scopus-indexed studies (2017-2025, PRISMA, minimum quality score of 6 of 9) on AI integration in mathematics education and find the field's problem is not tool access but pedagogical capacity: constraints in technological–pedagogical integration and teacher readiness appear in 38 of 54 studies, ahead of equity, access and ethical risks (34) and limits to deep mathematical reasoning (29). Seven recurring challenges are mapped against five pedagogical responses, and both are reframed as a three-dimensional framework — cognitive–epistemic, instructional–interactional, systemic–ethical — plus a structural model of 14 proposed links between the two. The review's most consequential claim runs against the assumption that capable AI makes teachers less necessary: mediation becomes more critical, not less, because AI supplies answers while mathematics instruction is supposed to make students produce them.

Key Findings

1. Teacher readiness, not tool access, is the most frequently reported constraint. Technological–pedagogical integration and teacher readiness appear in 38 of the 54 studies, ahead of equity, access and ethical risks (34), constraints on deep mathematical reasoning (29), assessment validity and integrity (26), AI overreliance and declining cognitive effort (24), the conflict between AI-generated answers and deep learning (20), and reduced mathematical discourse and collaboration (19).
2. The pedagogical literature is dominated by teacher-mediated integration. Studies addressing adaptive and personalized learning pathways (37), AI-supported conceptual understanding (36), inquiry-driven and problem-centered pedagogy (36), teacher-mediated integration (32) and expanded computational and multimodal mathematical practices (23) all converge on deliberate instructional orchestration rather than autonomous tutoring.
3. The trajectory runs through five phases. Technology-assisted instruction with ICT and robotics (2017–2019), intelligent and adaptive systems (2020–2021), pre-generative expansion with early AI writing tools (2022), the generative AI disruption of 2023, and generative AI integration across teaching, learning and assessment (2024–2025).
4. Output is recent and concentrated. Of the 54 included studies, 37% were published in 2025 and 22% in 2024; MDPI (30%), Springer (28%) and Elsevier (19%) published most of them, and the United States and China each account for 16% of the contributions.
5. AI supports answers better than reasoning. Across the corpus, AI handles proof construction, multi-step problem-solving and abstraction poorly, its explanations often lack the depth of formal mathematical reasoning, and in Calculus instruction AI-generated answers were sometimes incorrect and needed teacher intervention.
6. Overreliance changes when learners engage, not only how much. Students defer initial processing to AI and then intervene at specific moments — checking or refining outputs — shifting the temporal structure of effort rather than simply reducing it, which the review frames as displacement of the conditions for extended reasoning.
7. Equity risks are redistributive, not just persistent. Robots and programming tools remain unavailable in some schools on cost and technology grounds, privacy and algorithmic bias enter through AI-driven feedback, and the review argues AI may reconfigure existing inequalities rather than merely reproduce them.

Teacher readiness is the headline constraint

The frequency counts reorder the usual priorities. Teacher readiness and technological–pedagogical integration (38 of 54) outrank every other challenge, including overreliance (24) and academic integrity (26) — the two that dominate public discussion of AI in mathematics classrooms. The review reads this as evidence that integration failures are systemic: gaps in teachers' technological–pedagogical knowledge, tensions in professional identity as instructional authority shifts, and institutional and curricular misalignment. AI-generated tasks, examples and materials do save preparation time, but they do not reliably match intended learning objectives even under careful prompting, which increases demands on both mathematical expertise and technological fluency (TPACK). The review also reports that effective use of AI-generated materials develops over time through experience and reflective practice rather than arriving with the tool. For teacher role and educational development, the implication is that AI adoption is a pedagogical-capacity problem and not a procurement problem.

Mediation grows more important as the tools improve

The review's structural model links each challenge to a pedagogical response, and the teacher-mediated links carry the most weight. H4a and H4b propose that overreliance is mitigated by inquiry-driven pedagogy and teacher-mediated integration; H3 proposes that assessment validity and integrity are maintained by teacher mediation; H7a and H7b propose that disrupted collaboration and discourse are restored through inquiry and computational practices. Evidence in the corpus points the same way: groups with active teacher participation showed higher motivation and more productive tool use than groups working with AI alone. The finding extends Desirable Difficulties logic to AI-mediated classrooms — the review notes that AI-supported environments can sustain exploration, scaffolding and error-based feedback when teachers treat AI inaccuracies as material for guided inquiry rather than as failures to be hidden. Teachers' own judgment remains the mechanism through which AI errors are caught and corrected.

Assessment, integrity and the opacity of AI-assisted work

Assessment is where the review's challenge and response categories most directly collide. Twenty-six studies raise validity and integrity problems: AI-assisted responses make it hard to establish the authenticity of student work, especially in take-home and open-ended tasks, and an epistemic opacity in AI-assisted answers blurs the boundary between student understanding and AI contribution. Because AI can generate complete solutions, scaffold reasoning or simulate problem-solving, the boundary of what counts as legitimate mathematical work becomes institutionally unclear. The review notes that teachers' contextual knowledge of learners still yields richer interpretations of mathematical understanding than automated systems, but that this advantage erodes when AI mediates substantial portions of the learning process. Its proposed response is squarely instructional: maintain assessment validity and academic standards through teacher judgment (H3) rather than through detection, with sustained critical evaluation of whether AI tools reflect genuine learning.

Equity, discourse and the social conditions of learning

The dimension the review labels systemic–ethical gathers the challenges that no single classroom can solve. Thirty-four studies raise equity, access and ethical risks; nineteen report reduced mathematical discourse and collaboration as AI-mediated environments individualize engagement and displace peer argumentation. The mechanisms are concrete: robots and programming tools are unavailable in some schools, digital literacy varies, and AI-driven feedback systems introduce privacy and algorithmic bias risks. The review also flags that AI can simulate both teacher and peer roles, altering classroom relationships without necessarily preserving their function, and that unvalidated AI errors go uncorrected without expert oversight. Its proposed responses are adaptive and personalized pathways (H6a) and teacher mediation (H6b), which places the burden of equitable AI use on instructional design and institutional policy rather than on the tools themselves.

What this means for practice

• Instructors. Treat AI output as a text to be interrogated, not a solution to be accepted: the corpus shows students defer initial processing to AI and re-engage only to check or refine outputs, so assign the adjudication — which step is wrong, which representation is misleading — explicitly.
• Instructors. Keep your participation visible in AI-mediated tasks; the review reports higher motivation and more productive tool use in groups where the teacher actively mediates, and locates the mitigation of overreliance and integrity risk in that mediation rather than in tool restrictions.
• Curriculum designers and teacher educators. Address pedagogical integration before purchasing decisions — teacher readiness and technological–pedagogical knowledge is the most frequently reported constraint (38 of 54 studies), and effective use of AI-generated materials develops over time through reflective practice.
• Assessment designers. Redesign for evidence of process; when AI can produce the answer, validity depends on capturing the reasoning a student supplies, and the review's integrity findings concentrate in take-home and open-ended formats.
• Administrators. Fund equitable access alongside pedagogy: unavailable robotics and programming tools, uneven digital literacy and privacy exposure in AI feedback systems are the conditions under which AI reconfigures rather than reduces existing inequalities.

Limitations

• The review searches Scopus only, and its search string deliberately includes broader terms (intelligent tutoring systems, adaptive learning systems, educational technology) to capture pre-2021 work, so some included studies are educational technology rather than AI.
• Inclusion required English-language, peer-reviewed articles accessible through a university subscription, so relevant non-English and paywalled-adjacent work is outside the sample.
• Quality appraisal (minimum 6 of 9, ICC 0.764 across three independent assessors) excluded 260 of the 314 eligible studies, so the synthesis rests on the highest-scoring quarter of the screened literature.
• The challenge and pedagogical counts are frequencies of studies addressing each theme, not effect sizes or pooled estimates; the review reports no meta-analytic synthesis of outcomes.
• The three-dimensional framework and its 14 hypotheses are proposed rather than tested, and the authors call for the dimensions to be operationalized into measurable constructs with reliability and validity evidence.

Citation

Manto, A., Visitacion, M. A., Yee, A. M., Roble, J., Capuyan, S. N., Hermoso, F., Milano, M. L., Gonzales, R., Hallarte, D. K., & Gonzales, G. (2026). AI integration in mathematics education: A systematic review of implementation challenges and pedagogical implications (2017-2025). Social Sciences & Humanities Open, 14, 103147.