A new theoretical framework from a South African education researcher argues that artificial intelligence tools in primary mathematics classrooms should be designed to protect and extend children's independent reasoning, rather than simply automating task completion.
Published in the journal Frontiers in Education, the Human-Centred AI Pedagogy Framework was developed by Z D Zanele Dibane of the University of South Africa's Institute for Open and Distance Learning. The paper synthesises insights from sociocultural theory, human-centred AI, social cognitive theory, the TPACK (Technological Pedagogical Content Knowledge) model, and Ubuntu-inspired relational ethics.
According to Dibane, the central problem is that AI can help a child complete a maths problem without ensuring the child understands the underlying reasoning. "The distinction between successful task completion and mathematical understanding is central to this paper," Dibane writes. "AI should be evaluated not only by what children accomplish with it, but also by the reasoning they develop through its use."
The framework introduces the concept of "decision drift" — a situation where, as teachers reduce direct guidance, mathematical decisions may be taken over by the AI system rather than being handed to the child in a supported way. The proposed alternative is "agency handover," which means increasing children's supported responsibility for strategy, representation, justification, and help-seeking, while teachers retain safeguarding and instructional accountability.
Dibane draws on evidence from older learners to illustrate the risks. A field experiment involving nearly 1,000 high school students, cited in the paper, found that unrestricted generative AI support improved performance on assisted practice but reduced subsequent unaided performance. Another study of 1,625 university students across 10 courses identified a tendency to rely on AI assistance rather than learn from it. Dibane cautions that these findings cannot simply be transferred to primary children, whose development, language resources, and need for adult mediation differ, but argues they motivate investigation of the conditions under which AI might support learning without replacing thinking.
The framework is structured around five dimensions: prerequisites, enabling conditions, mediating processes, outcomes, and ethical boundaries. Two hypothetical lessons — one on Grade 3 place value and one on Grade 5 growing patterns — are used to illustrate how the framework would apply in practice, comparing AI-supported and non-AI alternatives.
Dibane also specifies important implementation conditions for South Africa, including multilingualism, unequal infrastructure, and teacher workload. The paper notes that critical accounts of educational AI question the emphasis placed on efficiency, prediction, and optimisation in adoption narratives, and that children may be positioned as recipients of algorithmically sequenced content if technical capability becomes the principal justification for use.
Five propositions are formulated to identify expected relationships, observable indicators, and competing explanations, providing a basis for classroom observation and comparative studies. However, Dibane acknowledges that the framework's explanatory and predictive validity remains to be tested. "Evidence from older learners motivates the analysis but does not establish effects in primary classrooms," the paper states.
For university students studying education, particularly those training to become primary school teachers, the framework offers a way to think critically about the role of AI in classrooms they will one day lead. It shifts the question from "how can AI be integrated?" to "how can AI-mediated learning environments protect and extend the human purposes of mathematics education?"
The paper is published as a Hypothesis and Theory article, meaning its claims remain open to empirical challenge. Dibane's contribution is described as an explanatory account of decision-making in an AI-mediated learning situation, rather than another model of technology adoption.