Artificial intelligence has swept into primary mathematics classrooms with a bold promise: personalised tutoring for every child, at a scale no human teacher could ever match. Intelligent tutoring systems sequence practice problems and adjust difficulty in real time, adaptive platforms reshape entire learning pathways from streams of performance data, and conversational chatbots hold natural-language dialogues that can probe a learner’s reasoning. Yet a new conceptual analysis published in Frontiers in Artificial Intelligence argues that none of this technological sophistication matters much without something far less glamorous: teachers who are deliberately prepared, through open distance e-learning, to orchestrate these tools. The paper, authored by Zanele Dibane of the University of South Africa, proposes that well-designed distance professional development is an enabling condition for AI’s educational benefits, not a guarantee of them, and it builds a testable framework showing exactly how that preparation is supposed to work.
The core insight driving the analysis is that AI does not operate autonomously in classrooms. Existing research explains parts of the puzzle: the AI-TPACK framework describes the integrated knowledge teachers need to use adaptive systems ethically, orchestration research shows how teachers coordinate activity across individual, group and whole-class planes, and human-centred AI models argue that systems should augment rather than replace human control, keeping teachers firmly in the loop. But these literatures do not connect. Knowledge models say nothing about how that knowledge grows at a distance; orchestration studies generally assume a teacher who already has the requisite capacity, developed mainly in well-resourced settings with older learners; and distance education research rarely specifies how programme design would change AI-related teaching practice. Dibane’s central question is therefore mechanistic: through which mechanisms, and under which boundary conditions, can distance-based professional development be expected to enable teacher-facilitated human-AI collaboration that supports young learners’ mathematical problem-solving?
To answer it, the study combined an integrative review of peer-reviewed and selected grey literature with theory synthesis. Structured database searches in Scopus and Google Scholar, supplemented by an AI-assisted discovery tool whose every suggestion was verified in original published form, screened 57 records down to 23 included studies, with eight further sources added through targeted searches. Sources were classified as direct evidence from primary mathematics, analogical evidence from other educational levels, theoretical works, or grey literature, and claims resting only on analogical or grey sources were phrased conditionally. The synthesis proceeded by extracting constructs and mechanisms from each source, coding them against four theoretical perspectives, assigning them to levels of analysis, and formulating relationships as propositions that were each checked against disconfirming evidence before being retained. The primary context is South African primary schooling, spanning Grades 1 to 7, with multilingual classrooms, uneven connectivity and unreliable electricity supply shaping both teachers’ participation and learners’ access.
The framework’s theoretical engine is a multilevel meta-model that assigns distinct functions to four perspectives plus one addition. At the programme level, transactional distance theory explains how distance education works: the psychological and communicative space between instructors and learners varies with dialogue, structure and learner autonomy, so well-designed programmes reduce that space through mentoring, flexible sequencing and professional community. At the teacher level, Intelligent-TPACK defines the capacity that professional development targets, comprising knowledge of how adaptive systems work, how AI can personalise instruction, the ability to judge AI-generated mathematics for conceptual soundness, and an explicit ethics component covering fairness, transparency, privacy and accountability. At the interaction level, Vygotsky’s Zone of Proximal Development supplies criteria for scaffolding quality, and at the classroom level, social constructivism specifies the discourse and collaborative structures into which AI-supported individual work must be embedded. Polya’s four-phase problem-solving model, understanding the problem, devising a plan, carrying it out and looking back, serves as the analytic lens at the learner level.
The scaffolding criteria are where the analysis gets technically sharp. Drawing on the scaffolding literature, the framework insists that assistance counts as scaffolding only if it is contingent on a diagnosis of the learner’s current understanding, calibrated in amount and type, gradually faded, and oriented toward transferring responsibility to the learner. AI systems estimate a learner’s position from proxies such as step correctness, error patterns, hint requests and response latency, but these are performance measures, not direct readings of a child’s Zone of Proximal Development, which is inherently social, dialogic and shaped by emotional and cultural context. The framework therefore requires teachers to validate the AI’s estimate by triangulating it with their own formative evidence, override task assignment where the two diverge, configure fading rules so that help declines as learners succeed, and treat AI errors, such as a generative chatbot producing fluent but incorrect mathematics, as hypotheses to be checked and even as teaching opportunities. On this account, it is the teacher-AI system together, not the AI alone, that may function as the more knowledgeable other.
From this synthesis, Dibane proposes a three-layer framework separating inputs, mechanisms, practices and outcomes, along with seven testable propositions. Layer one contains distance professional development as the input and AI-TPACK as the teacher capacity it builds. Layer two contains teacher orchestration, the practice of selecting and configuring tools, validating AI estimates, deciding when to intervene or fade assistance, managing AI error, and designing social learning around AI use, whose proximal product is scaffolding quality. Layer three contains learner problem-solving processes and outcomes, including mathematical reasoning, self-regulation and, crucially, AI dependence as an unintended outcome predicted to rise when scaffolding substitutes for rather than supports the learner’s own thinking. The propositions state that programme design quality and dosage relate positively to AI-TPACK growth; that AI-TPACK shapes orchestration; that orchestration mediates the relationship between knowledge and scaffolding quality; and that scaffolding quality, rather than AI access alone, predicts learners’ reasoning and lower dependence.
The propositions also specify moderators and boundary conditions. Infrastructure readiness moderates both professional development participation and learner outcomes; AI type moderates the verification burden, since generative tools require more case-by-case checking than model-based systems whose errors are systematic and detectable; learner age and the alignment between home language and the language of learning and teaching moderate scaffolding effects, a particular concern in South Africa where the language of instruction changes in the Intermediate Phase; and school support moderates how knowledge translates into practice. A sixth proposition holds that orchestration decisions are of higher quality when teachers triangulate system-generated analytics with their own professional observation of learners’ explanations, participation and language use, because each channel captures what the other misses. The seventh is the starkest: governance conditions protecting children’s rights are necessary for implementation, and where they are unmet, the framework does not recommend AI use at all, regardless of how promising the pedagogy.
That governance condition carries eight specific requirements, from parental consent and child assent, data minimisation and defined retention periods, to prohibitions on consequential automated profiling without human review, explainability and the right to contest system inferences, child safeguarding with supervised and curriculum-restricted conversational tools, bans on commercial use of learner data, accessibility standards for learners with disabilities, and linguistic fairness in multilingual classrooms. The empirical backdrop makes the caution understandable. Meta-analyses of intelligent tutoring in schools report strikingly mixed effects: one synthesis of K-12 mathematics studies found average effects ranging from just 0.01 to 0.09 standard deviations, with smaller benefits for low achievers, while another reported a median gain of 0.66 standard deviations that shrank dramatically on standardised tests. A field experiment in secondary mathematics found that unrestricted access to a GPT-4-based tutor improved performance during practice but reduced subsequent unassisted performance, whereas a version with learning-oriented guardrails largely avoided this harm. The framework predicts exactly this heterogeneity: if AI benefits depend on orchestration and contextual conditions, effects should vary as those conditions vary.
The analysis closes with implications for five stakeholder groups and an honest account of its own limits. Teacher education institutions should build distance programmes organised around all five AI-TPACK components and five design principles: sustained engagement, practice-based learning tasks enacted in teachers’ own classrooms, collaborative community, contextual responsiveness with low-bandwidth and offline-capable materials, and congruent modelling in which the programme itself uses AI the way teachers are expected to use it with learners. Policymakers should set AI competency standards and fund preparation rather than buying technology alone; school leaders should protect collaborative learning time and invest in connectivity, devices and stable electricity; researchers should test the propositions through longitudinal multilevel and comparative studies; and learners should be positioned as rights holders with voice, not merely beneficiaries. Dibane acknowledges that the framework is a conceptual synthesis, not an empirically validated model, and that its transferability beyond the South African context is an open question. But the theoretical bottom line is clear: AI is neither a substitute for teachers nor a peripheral add-on, but a powerful yet bounded instructional tool whose value depends on the very human expertise it is sometimes assumed to replace, and the real measure of success is whether children taught by well-prepared, critically engaged teachers learn more mathematics.
Subject of Research: An ODeL-based teacher professional development framework for teacher-facilitated human–AI collaboration in primary mathematics problem-solving: a conceptual analysis
Article Title: An ODeL-based teacher professional development framework for teacher-facilitated human–AI collaboration in primary mathematics problem-solving: a conceptual analysis
Article References: Dibane, Z. (2026). An ODeL-based teacher professional development framework for teacher-facilitated human–AI collaboration in primary mathematics problem-solving: a conceptual analysis. Frontiers in Artificial Intelligence, 9, Article 1970792. https://doi.org/10.3389/frai.2026.1970792
Image Credits: AI Generated
DOI: 10.3389/frai.2026.1970792
Keywords: ODeL-based, teacher, professional, development, framework, teacher-facilitated, human, collaboration, primary, mathematics, problem-solving, conceptual
News Source: Denise Maddox. (October 23, 2026). Distance Training Could Decide Whether AI Actually Helps Children Learn Math. Scienmag.
