Teacher Mediation and the Contingent Promise of Generative AI in Mathematics Education

Synthesis: This eight-week qualitative multiple-case study examined how ten secondary mathematics teachers in two contrasting public schools used ChatGPT-4 to scaffold learning, and what happened to classroom authority and equity when they did. Five teacher-mediated practices were identified: provoking inquiry with AI-generated problems, sequencing hint-based scaffolds, facilitating critical comparison of AI output, customizing prompts for diverse learners, and using AI as a reflective dialogue partner. Where teachers enacted these practices, AI output became an artifact for critique rather than an answer key, and student activity shifted toward conjecture and metacognitive questioning. Where mediation was absent, authority migrated toward the algorithm and teachers became validators rather than first sources of knowledge. An exploratory comparison of 300 model responses found the free-tier model inaccurate in 32.7% of instances against 12% for the premium model (χ2(1)=17.5, p<0.001). The paper concludes that GenAI amplifies existing pedagogical conditions, and that its equitable promise is contingent on critical mediation and systemic support.

Key Findings

• Five teacher-mediated scaffolding practices were identified: AI-generated problem posing, sequenced hint scaffolds, critical comparison of AI output, differentiated prompting, and reflective dialogue.
• Students using AI as a reflective dialogue partner moved from answer retrieval to self-regulated reasoning; one student reported the AI "didn't answer, it made me figure it out myself."
• An exploratory comparison of 300 model responses found the free-tier model inaccurate in 49 of its 150 responses (32.7%) against 18 instances (12%) for the premium model (χ2(1)=17.5, p<0.001).
• Power shifted toward the algorithm: students consulted the AI before their teacher and returned for validation. One teacher described moving from "oracle" to "editor."
• Coding reached saturation after 80% of the data; double coding of 25% of the data yielded 89% average agreement.
• One teacher said the school acceptable use policy was 10 years old and did not mention AI; another described "conducting an experiment without informed consent."

Study design and context

This interpretivist multiple-case study followed an eight-week GenAI intervention in two public secondary schools in a large urban district, purposively selected for contrasting technology integration. Participants were 10 mathematics teachers (six female, four male) with 2 to 15 years of experience, and 24 students (13 female, 11 male) aged 15 to 17. ChatGPT-4 was integrated into lessons on problem solving and algebraic reasoning through a prompt-based framework emphasizing Socratic dialogue. After a two-hour professional development session, all classes used two common tasks, in week 3 and week 6, supported by five core prompt stems. Data comprised teacher interviews, student focus groups, classroom observations, and student artifacts.

What critical mediation looked like

Five teacher-mediated Scaffolding practices recurred. Teachers used GenAI to pose open-ended problems: one asked for a version of x² − 7x + 10 = k with an unknown parameter k, which precluded a single mechanical answer. They pre-prompted the AI to release tiered hints while withholding the solution: "I never let it give the solution." They required students to place AI output beside their own work and argue which method was better. They customized prompts so a struggling student received visual analogies while a peer received a cubic extension, with students choosing their own path. They also had students interrogate the AI with "why" and "what if" questions; one reflected, "The AI didn't answer, it made me figure it out myself." The authors present these practices as an integrated model of critical mediation that develops Metacognition rather than Cognitive Offloading.

The algorithmic divide and shifting authority

The second research question concerned power and ethics. GenAI created a triadic relationship among teacher, student, and algorithm: students consulted the AI first and brought its output to the teacher for validation. One teacher said, "I'm no longer the oracle; I'm the editor." An exploratory comparison of model tiers produced the study's sharpest equity finding: with 50 mathematics prompts submitted three times to each model, the free-tier model produced inaccurate responses in 49 of its 150 (32.7%) versus 18 of 150 (12%) for the premium model (χ2(1)=17.5, p<0.001). Teachers in the under-resourced school called the free model "borderline useless for mathematics." The paper frames this as an algorithmic divide extending the classic Digital Divide beyond device access to tool quality, and reports a AI Governance gap between opaque platforms and outdated institutional policy.

What the findings mean

The authors argue that GenAI is not a neutral tool but a culturally situated artifact whose value is contingent on critical mediation, drawing together sociocultural theory and Critical Pedagogy. GenAI amplified existing pedagogical conditions: where teachers curated and interrogated outputs, students engaged in conjecture and argument; where mediation was absent, the tool risked an algorithmic banking model in which students passively received answers, as Productive Failure research would predict. The paper positions the Teaching as transformed rather than supplanted, into an epistemic guide whose critical mediation is a professional practice rather than a technical add-on. Because the technology amplifies whatever conditions it enters, the authors conclude the central challenge of AI in education is pedagogical and political rather than technological.

What this means for practice

• Treat AI output as a provisional artifact for critique: require students to place it beside their own work and justify which method is better.
• Sequence AI assistance as hints that fade, pre-prompting the tool to withhold solutions so support declines as competence grows.
• Teach and model critical evaluation explicitly, since episodes where the AI errs become the strongest openings for mathematical argument.
• Advocate for parity in tool quality and current data governance, given the ten-year-old acceptable use policy teachers described.

Limitations

• The qualitative multiple-case design and small sample (10 teachers, 24 students in two schools) limit statistical generalizability, and equity findings rest on a comparative case study.
• The study used a premium model (ChatGPT-4), which may overstate benefits achievable with widely available, lower-performance tools and thereby intensify the equity concerns central to the inquiry.
• The model-tier comparison was exploratory: prompts were not randomized, so the observed accuracy gap warrants larger-scale investigation.

Citation

Canonigo, Allan Mesa. (2026). Teacher Mediation and the Contingent Promise of Generative AI in Mathematics Education. Journal of Computer Assisted Learning, 42, e70307. https://doi.org/10.1002/jcal.70307