Responsible AI use in mathematics requires students to judge whether a generated explanation is valid and whether they understand the reasoning behind it. In a study of students preparing fraction-related portfolios, Mbazima Amos Ngoveni (2026) found reports of useful support alongside concerns about inaccurate or incomplete outputs.
Some students described checking and adapting AI responses. The study could not establish how well they performed those checks, which leaves educators with a practical question: how do we teach the verification we expect students to do?
That question becomes particularly difficult when students are still learning the material. Someone seeking help with fractions may have limited knowledge for evaluating the explanation they receive. Telling that student to verify an answer sets an expectation, but it also creates a teaching responsibility. We need to help them understand what verification involves.
Responsible AI use in mathematics depends on subject knowledge
Mathematical checking involves examining whether a method is appropriate, whether individual steps are justified, and whether the result makes sense. An answer can look familiar without being understood. A student might recognise the format of a worked solution while missing a conceptual mistake within it.
This makes verification a disciplinary practice. Students need opportunities to use mathematical knowledge to evaluate explanations. A broad warning about hallucinations gives them a reason to be cautious; deciding whether a particular fraction calculation is valid requires more specific understanding.
The paper connects verification with students’ continuing responsibility for the reasoning in their submissions. This is also a central concern in AI and assessment. Teachers need evidence that allows them to interpret what students understand, including when generated explanations have contributed to the work.
Summary of the Study
Ngoveni (2026) examined the reported experiences of 27 Higher Certificate students in a mathematics module at a South African open and distance learning institution. Students completed individual assessment activities on fractions, culminating in a portfolio.
General guidance on responsible AI use was available through an online lesson, a recording, and an institutional academic integrity course. However, AI use was not explicitly addressed in the portfolio instructions, and disclosure was not required.
After submitting their assessments, students voluntarily answered a single anonymous, open-ended questionnaire item. Their responses were examined using reflexive thematic analysis, an interpretive approach to identifying patterns of meaning.
Four overlapping themes emerged. Students described support with understanding tasks and concepts, assistance with organising and expressing written responses, conditional trust and evaluation, and experiences involving non-use, resistance, or unclear engagement.
Some found explanations and examples helpful. Others considered outputs inaccurate, too general, or insufficiently explained. Ten participants expressed reliability concerns or described safeguards; five specifically reported checking, adapting, or limiting their reliance on generated material. These accounts show considerable variation in how students perceived and approached AI assistance.
The study did not examine prompts, outputs, or submitted portfolios to corroborate students’ accounts. It therefore provides no direct evidence that AI improved mathematical understanding or that reported checking was rigorous.
The small sample from one module, the single questionnaire item, and the possibility that students presented their practices favourably also limit interpretation. I find the author’s distinction between reported responsibility and demonstrated responsibility useful: recognising the need to check an explanation is one thing; being able to establish its mathematical validity requires further evidence.
Practical insights: teach verification and ask students to show it
The author recommends subject-specific guidance and assessments that require students to explain, justify, and verify their reasoning. My suggested classroom adaptation would be to model one check in detail before asking students to carry it out independently.
A teacher could work through an AI-generated fraction solution and explain how to evaluate the method, identify an unjustified step, and test the result. This is a teaching proposal, not an intervention evaluated in the study.
Students could then submit a short verification note with a solution. The note might identify a step they checked, explain the mathematical principle used, and show what they concluded. If they corrected the response, they could explain the correction.
If they accepted it, they would still need to justify that decision. This would give teachers more information than a statement that the student had checked the answer.
Such work connects with evaluative judgement in the AI age. Judging quality requires criteria and practice. Students should have opportunities to discuss why an explanation is sound, where it is incomplete, and how an apparently plausible method can fail.
The findings also suggest a need to clarify assistance boundaries within each assignment. Some participants valued AI for organising language and expressing their own ideas.
Yet the researcher could not determine where language support ended and substantive generation began. Assignment instructions can explain which forms of assistance fit the learning outcome and which reasoning students must demonstrate themselves.
Finally, non-use deserves careful interpretation. Some students avoided AI or distrusted it; others gave accounts too brief to interpret confidently. We should leave room for students to use course materials and other supports while developing the same mathematical capabilities. The assessment question remains: what evidence shows that this student can explain and justify the reasoning in the work?
Reference
Ngoveni, M. A. (2026). Problem solving and responsible GenAI use in assessment: Students’ reported experiences in open and distance learning. Frontiers in Education, 11, Article 1889499. https://doi.org/10.3389/feduc.2026.1889499




