Artificial intelligence in primary education: a systematic literature review 2020–2025

 📜🔗LINK to the source

1️⃣ ANALYTICAL SUMMARY

A rapid deployment, still fragile conditions. The review documents the rise of AI in primary education between 2020 and 2025, with a clear acceleration (1 study in 2020, 41 in 2025, p. 703). The studied populations are mainly teachers (43%), then students (34%), mixed groups (19%), and very marginally, future teachers (4%, p. 702-703). The identified issues are recurrent: unequal digital skills of teachers (AI-TPACK), infrastructure inequalities, and ethical concerns regarding transparency, biases, and data governance. The common thread: AI is neither saviour nor threatening in itself, but depends on the human and institutional conditions of its integration.

What research validates as operational contributions. AI produces its best effects when it personalises feedback and training, supports questioning and computational thinking, and enhances teaching decisions through learning analytics. Reported gains: reading fluency, problem-solving, motivation, participation from diverse audiences. The most robust and transferable result is the principle of « human-in-the-loop » (human in the loop): sustainable effects emerge when the professional interprets and moderates what the AI produces. The review concludes that equity must be a design choice from the outset (equity-by-design), not a fix added later.

2️⃣ KEY POINTS OF THE DOCUMENT

1️⃣ The principle of « human in the loop » is the most consistent success factor. Across the studies, sustainable outcomes appear when the teacher interprets and regulates the outputs of the AI rather than delegating them (p. 706, 708, 714). A direct comparison shows that human tutors achieve double the learning gains of AI agents on arithmetic tasks (Anton et al., 2025, p. 706).

2️⃣ ‘Unplugged’ and low-cost AI can reduce inequalities. In Brazil, the AIED programme ‘Unplugged’ has reached over 8,000 schools and 160,000 students, with comparable writing gains in rural and urban areas, without reliance on the Internet (Portela et al., p. 706). This is a strong counter-argument to the idea that equity equates to technical sophistication.

3️⃣ Teaching competence is about orchestration, not technical mastery. The review formalises three levels: the “AI fluency with safeguards” (epistemic humility: knowing where AI goes wrong), the “class orchestration” (knowing when to delegate, when to take back control) and the “design for equity” (p. 714).

4️⃣ National frameworks structure appropriation. Documented example: China has defined a national curriculum of 90 AI concepts and 63 learning indicators (Song et al., 2023, p. 706). Validated measurement instruments (AI-TPACK scale, AILST literacy scale) allow for diagnosing skills and targeting training (p. 709, 711).

5️⃣ The recurring limits are ethical and developmental, not just technical. Large language models generate plausible but sometimes incorrect reasoning, unsuitable for the cognitive maturity of children (Getenet, 2024, p. 708, 712); issues of consent, data minimisation, and transparency remain unresolved, particularly with “smart” sensors and classes (p. 710, 712).

3️⃣ ACTION TRACKS FOR LOCAL ACTORS

1️⃣ Establish a “human in the loop” rule before any deployment of AI tools to an audience: explicitly define who interprets, validates, and corrects automated outputs (validated principle p. 706-714). Transferable to any feedback or user profiling system.

2️⃣ Prioritise low-cost, offline solutions that are context-appropriate rather than “high-end” systems where resources and connectivity are limited (Unplugged model, p. 706-707). A principle directly relevant in social care and community health.

3️⃣ Making training a prerequisite, not an optional accompaniment : the review documents that a structured professional development system (e.g. 75-hour training based on the TPACK framework, Sun et al., 2023, p. 707-708) significantly improves the competence and confidence of professionals.

4️⃣ Diagnosing skills before training using validated grids (type AI-TPACK, AILST) to target real needs rather than training "at random" (p. 709, 711, 715).

5️⃣ Integrating consent and data protection from the design stage (consent-by-design, minimisation, documentation of data flows), involving the public and families in decisions about data usage (p. 710, 712, 716).

6️⃣ Explicitly teaching critical thinking in the face of AI : teaching the public to identify erroneous outputs and to contextualise automated outputs, to transform dependence into reflective learning (p. 706, 712). Identified unmet need: the review notes that the preparation of (future) professionals remains very underexplored (4% of studies, p. 703) — a blind spot to be filled locally through tutoring and supported practice.

4️⃣ ADDITIONAL REFERENCES

White Paper: Educating about AI from school: an emergency plan - Version 1  - https://www.pratiquesensante.com/blog/actualites-17/livre-blanc-eduquer-a-l-ia-des-l-ecole-un-plan-d-urgence-version-1-6063

1️⃣ Ministry of National Education — The framework for the use of artificial intelligence in education (June 2025). Official French framework governing the pedagogical and administrative uses of generative AIs (GDPR, SREN law, European regulation on AI). Complementary to the governance/data/consent aspect of the review. 🔗 https://www.education.gouv.fr/sites/default/files/2025-06/l-ia-en-ducation---cadre-d-usage-440685.pdf

2️⃣ IGÉSR — Artificial intelligence in schools (May 2025). Field assessment (survey of 4,943 teachers, primary and secondary levels), 10 recommendations focused on the lack of training and equity. Direct echo to the review's findings on professional preparation and inequalities. 🔗 https://www.education.gouv.fr/igesr/l-intelligence-artificielle-dans-les-etablissements-scolaires-465627

3️⃣ UNESCO — Competency frameworks in AI for students and teachers (2024-2025). International frameworks, available in French, structuring literacy and competencies in AI. Complementary to the core “AI literacy / AI-TPACK” of the review. 🔗 https://www.unesco.org/fr/digital-education/artificial-intelligence

5️⃣ FREQUENTLY ASKED QUESTIONS (FAQ)

1️⃣ Does AI really improve learning outcomes?

Yes, but under certain conditions: reported gains in reading, problem-solving, motivation, when the intervention is age-appropriate, guided by a professional, and adjusted to the context (p. 705-706).

2️⃣ Can AI replace the teacher or the assistant?

No. Human tutors still surpass AI in empathy and multimodal signals (gesture, gaze); in arithmetic, they achieve double the gains (Anton et al., 2025, p. 706, 712).

3️⃣ Is expensive equipment necessary for AI to be useful?

No. Low-cost and offline approaches work at scale and reduce inequalities (Portela et al., p. 706-707).

4️⃣ What are the main ethical risks?

Algorithmic bias, opacity, privacy violations, inappropriate consent for children, and plausible but false outputs from generative models (p. 706, 708, 710, 712).

5️⃣ How to prepare professionals?

Through structured professional development combining AI-TPACK, assessment literacy, and data ethics; supported practice remains the best predictor of confidence (Batubara et al., 2025; Sun et al., 2023, p. 708, 713).

6️⃣ Is AI suitable for audiences with specific needs?

It shows real potential (sign language games, no-code robotics, learning profiling), provided there is an institutional framework and ethical safeguards (p. 706, 709-710).

7️⃣ How to know if a tool is reliable before adopting it?

Check: pedagogical alignment, transparency of operation, developmental calibration, and the existence of human supervision. Without these conditions, the review recommends caution (p. 712).

6️⃣ REWRITING IN EASY LANGUAGE

What is this document about?

This document studies artificial intelligence (AI) in primary school.

AI is a computer tool that helps with learning.

The researchers read 94 studies conducted between 2020 and 2025.

What they found good:

• AI can help children read better and calculate better.
• AI can inspire a desire to learn.
• AI works even without the Internet and without expensive equipment.
• This helps schools that have little money.

The most important point is:

• AI works well when an adult is present.
• The adult checks what the machine says.
• The adult corrects when the machine makes a mistake.
• The machine does not replace the teacher.

What poses a problem is:

• Many adults are not adequately trained.
• AI can make mistakes. Sometimes it gives false answers.
• AI can raise privacy issues.
• It is necessary to protect children's data.

What needs to be done is:

• Train adults before using AI.
• Choose simple and inexpensive tools.
• Always keep an adult who decides.
• Teach children not to believe everything.

7️⃣ CROSS-ANALYSIS — VALUES OF HEALTH PRACTICES

• Literacy: yes — the review places AI literacy at the heart of the subject and cites frameworks and graded scales (AILST) suitable for various levels (p. 709-711).
• Empowerment: beneficiaries are involved mainly through co-design mechanisms (co-design of educational units) and the call to involve students and families in usage standards (p. 707, 716).
• Participation: several co-construction mechanisms are described (teacher-researcher co-design, school-university collaboration), but remain minority in the corpus (p. 707, 709).
• Community health: the collective dimension appears through « family-school-society » models and projects rooted in rural community (Xia et al., 2025 ; Wang et al., 2025, p. 709-711).
• Ethics : yes — cultural and social biases explicitly identified (reproduction of gender, origin, status biases) and treated as a condition, not an option (p. 710, 712).
• Human rights : the approach claims equity and inclusion as structuring principles (equity-by-design), with attention to audiences with specific and multilingual needs (p. 710, 716).
• Intersectorality : recommended partnerships between teachers, researchers, decision-makers, businesses, and communities (p. 710, 715).
• Partnership : collaboration models are formalised (research-practice, school-university, family-school-society linkage), without standardisation between contexts (p. 709-710, 716).
• Combating discrimination : yes — the document warns of the risk that AI « deepens the divides » (unequal access, biases) and the need for culturally adapted and low-bandwidth solutions (p. 710, 714).

8️⃣ EVALUATION OF RESOURCE RELIABILITY

Scientific relevance — high. Transparent and replicable methodology (PRISMA 2020), significant corpus (94 studies from 831 records), double coding with strong inter-rater agreement (Cohen's κ = 0.82, p. 701), very current data (up to 2025). Limitations honestly acknowledged by the authors: restriction to English, overrepresentation of high-income countries, small samples and short durations limiting causal robustness, absence of meta-analysis (narrative/thematic synthesis only, p. 700-702).

Operational relevance — indirect. The document does not provide a ready-made tool; its value is strategic (discernment grid, validated points of caution). For actors outside education, transposition requires adaptation work.

⚠️ Internal inconsistencies noted (reported, not corrected):

• Divergent citation dates. Portela et al. is cited as "2024" in the text (p. 706, 707, 714) but dated "2025" in the bibliography. Dahal et al. appears as "2025" (p. 709) and "2024" (p. 711), with the reference bearing 2025 for a FIE 2024 conference.
• Floating attribution of the Chinese curriculum of 90 concepts : "Song et al., 2023" (p. 706) then "Song et al., 2025" (p. 708) — two Song references coexist, the attribution of the specific framework varies.
• Notable chronological tension. The article is "received in February 2025" (p. 697) but includes 41 studies from 2025 (Fig. 2, p. 703), some of which were published in November 2025 (Pyżalski). This inclusion implies an undated revision after the initial submission.
• Non-additive counts : ~60 + ~70 + ~55 + 36 studies exceed the 94 of the corpus — this is explicitly justified by the overlap of categories (footnote, p. 705), therefore consistent but not to be added.
• Incomplete DOI (« 10.15388/infedu.2025.24 », without article suffix) and spelling variants of names (Kızıltas/Kiziltas, Pyżalski/Pyzalski).

Overall : reliable and rigorous, of which the few bibliographic inconsistencies do not undermine the substantive conclusions.

9️⃣ STRATEGIC HASHTAGS

#healthpractices #AIandEducation #DigitalLiteracy #DigitalEquity #EthicsAI #ProfessionalDevelopment #DigitalInclusion #HumanInTheLoop