Artificial intelligence in initial teacher training for pre-service primary school teachers in mathematics: a systematic review

Synthesis: Pinto, Rodrigues, Brito-Costa, Costa, Gonçalves, Pires, and Martins (2026) report a PRISMA 2020 systematic review of how artificial intelligence has been integrated into initial teacher training for pre-service primary school teachers in mathematics. Searching Scopus, Web of Science, and ERIC in June 2026 returned 341 records, of which 11 studies met the eligibility criteria, all published between 2024 and 2026. The interventions were mostly short-term and embedded within existing courses, and most moved from a preparatory phase into practical use of AI tools and then into reflection on the product. Monitoring leaned heavily on records of prompts and responses, while attitudes were usually measured once. The authors argue that brief, tool-focused training is not enough: pre-service teachers need sustained and continuously monitored training that builds AI competency on a base of mathematical and didactic knowledge, so that they can evaluate and adapt AI-generated content instead of accepting it.

Key Findings

• The review identified 341 records across Scopus (197), Web of Science (72), and ERIC (72), of which 11 studies met the eligibility criteria, all published between 2024 and 2026.
• Nine of the 11 interventions were short-term: a single session or a small number of sessions embedded in existing courses, while only two ran modules of five and eight weeks.
• Eight of 11 studies opened with a preparatory phase before practical use of AI tools, but prompting was treated as explicit training content in only some of them.
• Ethics was the thinnest dimension: only three studies touched on plagiarism, copyright and data privacy, or academic integrity, and none addressed transparency, algorithmic bias, or accountability.
• Monitoring relied most on records of prompts and AI responses, often combined with analysis of the outputs produced, such as formulated problems, solutions, and lesson plans.
• Participants' attitudes and perceptions were mostly collected at a single point in time, generally after the intervention; only one study used pre- and post-test questionnaires.
• Reported risks included uncritical acceptance of ChatGPT errors, increased delegation to the tool as tasks became more complex, and overreliance that participants themselves identified as a main challenge.

How the studies were selected and appraised

The review followed PRISMA 2020 and searched Scopus, Web of Science, and ERIC between 25 and 30 June 2026 with a string combining AI terms, teacher-training terms, and mathematics terms. Starting from 341 records, 76 duplicates and 14 items outside the predefined window were removed, leaving 251 records for screening. Of the 136 records taken to full-text assessment, 125 were excluded, most often because they did not involve pre-service primary teachers (63) or did not address AI-related competences in mathematics (40). The 11 survivors were appraised with the Mixed Methods Appraisal Tool, and inter-rater agreement was high: Cohen's kappa was 0.83 for study selection and 1.00 for both quality appraisal and data extraction. The authors flag the obvious limitation: a recent, small corpus drawn from mathematics-focused training restricts how far the conclusions travel to other subject areas. No numerical outcomes are pooled, so this is a synthesis of reported practice rather than effect sizes.

How the training was structured

Two typologies emerged. Nine interventions were short-term and embedded in existing mathematics or teaching-methods courses; three of these were single sessions and five spanned two to six sessions. Only two studies ran longer modules, of five and eight weeks. In eight studies a preparatory phase preceded tool use. Six of those phases centered on AI concepts and tool mechanics, two on mathematical content and didactic knowledge, and three studies skipped preparation and started directly with the task, which shows that no settled structure for introducing AI into initial teacher training exists yet. Every study included a practical component, but the role assigned to AI varied: problem solving and problem posing, constructing proofs, automated reasoning in outdoor learning, a chatbot built to support assessment, and chatbots that simulate students with mathematical misconceptions. Prompting mattered everywhere. Where preparation was thin, some pre-service teachers simply transcribed the task statement without exploring the tool. Ethics was addressed in only three studies, and then mostly as user conduct.

What this means for practice

• The review's central lesson is that brief exposure to an AI tool is not enough.
• Nine of the eleven studies described a single session or a handful of sessions folded into an existing course, so a workshop or a one-off assignment should be treated as a starting point rather than a finished program of Workplace Learning.
• Designers should instead plan for Teacher AI Competency to develop across the whole initial training sequence, from a preparatory phase through the curricular practicum and into the early years of practice.
• That preparatory phase is where the structure is currently thinnest.
• A short introduction to AI concepts can be combined with Prompt Engineering and with prior mathematical and didactic knowledge, so that pre-service teachers learn to judge and adapt AI output instead of transcribing a task statement or accepting a fluent answer at face value. Programs should also treat Ethics as core content, since only three of the eleven studies addressed it at all, and should monitor progress over time rather than sampling attitudes once at the end.

Limitations

• Monitoring drew on three kinds of data: the interaction process between pre-service teachers and AI, the outputs produced, and participants' perceptions and attitudes.
• Records of prompts and responses were the most frequently used source, read either as reconstructed interaction sequences or coded against categories for problem solving, prompting, or teaching strategies.
• Outputs analyzed included formulated problems, problem solutions, and lesson plans, which let researchers connect how the tool was used with what the task produced.
• Perceptions came from questionnaires, interviews, focus groups, and written reflections, but with one exception these were collected at a single point in time, after the intervention, and sometimes only as satisfaction ratings. The analysis of risks is equally instructive: pre-service teachers failed to notice conceptual errors produced by ChatGPT, accepted generated content uncritically, and delegated more problem solving to the tool as tasks became harder. The authors therefore call for training that runs continuously rather than as an isolated module, with monitoring instruments that follow pre-service teachers through their guided to autonomous progression, the curricular practicum, and the early years of practice.

Connected Concepts

• Meta-Analysis and Systematic Review — the review methodology behind the 11-study synthesis
• Teacher AI Competency — the competence base the authors argue training should build
• AI Literacy — AI literacy as a progressively developed competency
• Technological Pedagogical Content Knowledge (TPACK) — technological pedagogical content knowledge used to interpret lesson planning
• Prompt Engineering — treated as training content in only some interventions
• Generative AI — ChatGPT was the tool most frequently used across the studies
• Workplace Learning — initial teacher training as sustained professional preparation
• Scaffolding — progression from guided exploration to autonomous use
• Math Education
• Ethics

Connected Articles

• Preparing Pre-Service Teachers for Responsible Generative AI Use: Curriculum Implications for Ethics, Privacy, and AI Literacy — Preparing Pre-Service Teachers for Responsible Generative AI Use
• Teacher education for artificial intelligence literacy through a self-determination theory perspective — Teacher education for AI literacy through self-determination theory
• Taming the Black Box: Design Principles for Rule-Integrated LLM Tutoring Systems in Primary School Mathematical Problem Solving — Rule-Integrated LLM Tutoring for Primary School Mathematics

Citation

Pinto, F., Rodrigues, R. N., Brito-Costa, S., Costa, C., Gonçalves, S., Pires, N. A., & Martins, F. (2026). Artificial intelligence in initial teacher training for pre-service primary school teachers in mathematics: a systematic review.