When Should Teachers Control AI Generation for Mathematics Visuals?

Synthesis: Generative AI can help teachers rapidly create classroom-ready visual materials, particularly in mathematics where diagrams and visual representations must be pedagogically meaningful and instructionally correct. This paper investigates when and how teachers should control AI generation of mathematical visuals. The study identifies key decision points where teacher input is essential: ensuring pedagogical correctness of visual representations, alignment with curriculum goals, and appropriateness for student levels. The authors propose a framework for teacher-AI collaboration in visual material creation that balances AI efficiency with pedagogical control.

Connections to AIED

This work connects to Teaching by showing how teachers remain essential as pedagogical validators of AI-generated content. It intersects with STEM Education since mathematics visual generation is a core need in STEM teaching.

The findings also relate to Generative AI and A principled way to think about AI in education: guidance for educators and policy makers based on goals, models, suggesting that principled AI use in education requires human oversight for pedagogically sensitive outputs. The teacher control framework connects to Evolution of AI in Education: Agentic Workflows where AI agents assist but humans direct.

What this means for practice

• Instructors. Judge a visual tool by where correction happens, not by how polished the output looks. In this study, manual verification of the 90 images generated under pre-generation control found only 11 with the correct number of objects (12.2% accuracy), rising to 70 of 117 under mid-generation control (59.83%) and only 2 of 93 incorrect under post-generation control (97.85%).
• Instructors. Prefer workflows that let you edit the generated image directly when a visual must encode an exact quantity or structure; post-generation editing removed the hidden transformations that made errors hard to localize and cost bounded and predictable to fix.
• Instructors. Budget verification time for prompt-only generation in mathematics. Seventeen of the 24 teachers reported that highly automated generation produced visually appealing but numerically incorrect images, and 11 abandoned otherwise promising images once they spotted an inconsistency.
• Instructors. Brief students and colleagues on the account of authorship this creates: teachers in the study said post-generation editing increased their sense of accountability for the final representation, so human-AI collaboration here means owning the artifact, not approving it.
• Administrators. Choose tooling that keeps structure inspectable rather than optimizing for one-click creation. The paper's core claim is that control timing shifts risk between interpreting intent, translating intermediate structure, and managing local refinement — a procurement question, not just a preference.

Limitations

• The study is a single-session, controlled, remote Zoom experiment with 24 primary mathematics teachers recruited via Prolific, roughly 70 minutes total with about 12 minutes per condition, so it captures first-use experience rather than adoption into real lesson-planning workflows.
• Tasks were limited to generating visuals for primary-level arithmetic, which the authors say constrains generalizability; interaction preferences and correctness sensitivities may differ at other grade levels or in other subjects.
• The study measures teacher perceptions, observed workflows, and image accuracy, not downstream outcomes — the authors state they cannot determine whether the observed differences produce higher-quality instructional materials or improved student understanding.
• Verification depended on teachers' own manual inspection, which places sustained cognitive demands on them, and the researchers had no automated correctness check to fall back on.

Citation

Li, Z., Wang, J., & Wang, A. Y. (2026). When should teachers control AI generation for mathematics visuals?. Proceedings of the Thirteenth ACM Conference on Learning @ Scale (L@S 2026), Seoul, Republic of Korea.