France devotes approximately 6.7 to 6.8% of its GDP to education (full national scope), which is more than the OECD average, including on the restricted scope comparable to the OECD where the figure stands at 5.4% versus 4.9% on average—and yet obtains results among the most unequal in the developed world. Artificial intelligence applied to education is emerging as the personalization tool that schools have been lacking for decades. But in a system where school of origin remains one of the best predictors of academic destiny, the technology risks arriving where it is already least useful.

The Essential Points

  • France ranks among OECD countries where social background most heavily influences school results, according to the Montaigne Institute.
  • Educational AI tools promise individualization of instruction impossible to deploy manually by a single teacher facing thirty students.
  • Schools best equipped with digital resources and trained teachers will be the first to adopt these tools.
  • The unequal deployment of a technology designed to reduce gaps could widen them further, unless there is a deliberately inverted resource allocation policy.
  • Experiments exist in France and abroad that demonstrate what piloted adoption can produce when it targets the most vulnerable populations first.

The Diagnosis That Nobody Really Disputes

The diagnosis enjoys surface consensus. Institutional actors acknowledge the inequalities, publish reports, commission studies, without this shared recognition producing any real reorientation of resources. Urgency is admitted in principle, but budget decisions and internal resistance perpetuate existing balances. Disadvantaged schools receive supplementary programs, rarely permanent structural resources. Rebalancing stumbles on the lack of willingness to bear the political cost of a genuinely inverted allocation.

The Montaigne Institute published in 2024 a detailed analysis of French education establishing a clear finding: the inequalities between schools are among the most marked in developed countries. A student enrolled in a well-off downtown high school and a student enrolled in a disadvantaged suburb high school do not study in the same educational country. They do not have the same teachers, not the same teacher replacement rate, not the same access to electives that open doors to preparatory classes. Geographic assignment, largely determined by place of residence, does the work of selection before the student has even opened a textbook.

Since the 2000s, PISA surveys have documented this at regular intervals: France is one of the rare countries where the performance gap between the lower quarter and upper quarter of students tends to widen rather than narrow. Average performance has declined. The gap has grown. This dual movement characterizes a system that lets the most vulnerable fall behind without raising the rest upward.

The reasons are well known and have long been debated: a school map that is not very constraining for those who know how to circumvent it, differentiated attractiveness of teaching positions by region, unequally accessible extracurricular resources. France tends to let adjustments happen by default rather than actively pilot rebalancing. Education is no exception to this logic.

Educational AI: Promises and Current Foundations

A teacher facing thirty students constantly plays on an average. He adapts his pace to the majority, lets the fastest grow bored and the slowest fall behind. Educational AI promises to escape this structural constraint by continuously adjusting the level, pace, and type of exercise for each student. What educators call pedagogical differentiation, which requires considerable preparation time, would become automatic.

Systems that already exist give an idea of what this can produce. Khanmigo, developed by Khan Academy, offers students Socratic dialogue rather than dry correction: it asks questions to lead the student to find their own error. In Estonia, adaptive tools have been integrated for several years into mathematics instruction. Studies on Mindspark, a tool used in India, showed significant learning gains for initially lower-level students over just a few months.

Current language models enable natural language interaction, explanations reformulated on demand, detection of recurring errors. The technology is capable of individualizing instruction, imperfectly but genuinely. The decisive question concerns the conditions under which it takes root and which audiences it reaches first.

Adoption Always Follows the Same Geography

Educational technologies have a history. It is rarely flattering from an equity perspective. The interactive whiteboard arrived first in schools where student parents funded active associations. ENT—digital learning environments—took a decade to standardize across educational regions. Access to online resources during the 2020 health crisis was largely conditioned by family equipment: the 2021 DEPP survey showed that a significant share of disadvantaged background students lacked access to a personal computer during school closure periods.

Educational AI will follow the same logic if nothing stops it. Schools with stable leadership, experienced teachers, and a network of partnerships with technology actors will be the first to experiment, train their teams, adjust practices. Schools in permanent recruitment tension, with often newly certified teachers and high turnover, will have neither the time nor the resources to absorb an additional pedagogical change.

This mechanism is well documented in economic literature on health or infrastructure innovations. As with automation in industry, the productivity gains from new technology benefit first to actors already in positions of strength. In education, this means that educational AI could accelerate progress for students who already needed it least, leaving untouched the situation of those for whom it would have been most useful.

Experiments Showing That Another Path Exists

Several countries have made the opposite choice: deploying tools first where needs are most acute, with proportionate support.

In the United Kingdom, the Education Endowment Foundation, established in 2011 to fund programs for disadvantaged students, has undertaken recent work on AI in schools classified as disadvantaged, imposing rigorous evaluation of effects before any scaling up. Specific programs on AI are very recent, posterior to 2022. Results are mixed depending on the tools, but the method is sound: test first where impact would be strongest, measure, adjust.

In France, the Ministry of Education has engaged since 2023 in reflection on AI at school, with experiments in several regions. Among existing programs, MIA, for Adaptive Interactive Modules, is a remediation tool for tenth-grade students in French and mathematics, developed by EvidenceB for the ministry. It’s a start. But the pilot regions are not necessarily those where equity stakes are strongest: volunteer regions tend to be those that already have a culture of pedagogical innovation.

Private actors are also beginning to explicitly target vulnerable populations. The Vikidia association, tutoring platforms like Schoolmouv or Lagonav, and startups like Lalilo for reading in first grade are progressively integrating adaptive components. Lalilo, deployed in several thousand French classrooms, many in priority education areas, adjusts decoding exercises in real time according to gaps detected for each student. Teacher feedback is encouraging, even if independent evaluations remain scarce.

Training Teachers Before Rolling Out Tools

This asymmetry of use between schools does not correct itself spontaneously over time. The more a tool is used in a favorable context, the more the practices surrounding it become sophisticated, widening the gap with contexts where use remains superficial. Ongoing training must therefore precede deployment rather than accompany it at the margins. Conditioning access to a tool on a minimum threshold of team preparation is not an additional bureaucratic burden: it is the only way to prevent the tool from becoming a marker of school rather than a lever for equity.

Technology solves nothing by itself. An adaptive tool poorly integrated into teaching practice becomes an additional distraction. Worse, it can reinforce the tendency to delegate monitoring of the most difficult students to the machine, at the precise moment when these students need the presence and judgment of a trained adult.

Initial and ongoing teacher training is the real bottleneck. In practice, French teachers follow on average a limited number of training days per year, a lag documented by the 2022 Senate report. The regulatory requirement for primary education nonetheless reaches 18 hours per year, as in Finland, while Singapore offers over 100 annual hours. The French lag thus stems from training actually taken. Successful integration of educational AI in a classroom requires the teacher to understand what the tool does, its limitations, how to interpret the data it produces, and how it articulates with his or her own interventions.

Without this training layer, the risk is real of creating two parallel uses: in well-equipped schools, a thoughtful use where AI augments the teacher; in tension-ridden schools, default use where AI substitutes for insufficient teacher presence. Free access to a tool is insufficient to guarantee equitable use of it: this is a lesson that higher education has already learned at its expense.

Public Policy Must Precede the Market

France has an advantage few countries possess: a unified public education service with genuine regulatory capacity over tools deployed in classrooms. It can impose standards for protecting student data, requirements for independent evaluation of tools before adoption, and especially an inverted resource allocation logic: REP and REP+ schools first, with associated training resources.

A few conditions make this possible. The first is detailed mapping of current uses: who uses what, where, with what measured effects. DEPP has statistical tools to conduct this work. The second is a labeling mechanism for educational AI tools, on the model of what the Ministry of Education has begun to structure for digital textbooks, conditioning classroom entry on impact evaluation on heterogeneous populations. The third is deliberate investment in training teachers in the most fragile schools, not as an emergency measure but as a structural priority.

Educational AI can reduce French school inequalities. It can also worsen them. The difference between these two scenarios does not hinge on the technology itself: it depends on the choices public authorities will make in the next two or three years, before adoption habits take hold and usage gaps become as difficult to remedy as the result gaps they will feed.

Sources

  1. Montaigne Institute, School: Where to Focus Our Efforts, https://www.institutmontaigne.org/publications/ecole-ou-concentrer-nos-efforts
  2. OECD, Education at a Glance 2023, https://www.oecd.org/fr/education/regards-sur-l-education.htm
  3. DEPP, Survey on Schooling Conditions During the Health Crisis, 2021, Directorate for Evaluation, Planning and Performance, Ministry of Education
  4. Muralidharan, K. et al., Disrupting Education? Experimental Evidence on Technology-Aided Instruction in India, American Economic Review, 2019, Mindspark study
  5. Education Endowment Foundation, AI programs in disadvantaged environments in the United Kingdom, https://educationendowmentfoundation.org.uk/impact-report
  6. Ministry of Education, MIA Tenth Grade – Adaptive Interactive Modules, https://eduscol.education.gouv.fr/6693/mia-seconde-une-approche-personnalisee-de-la-remediation-en-francais-et-en-mathematiques
  7. DEPP – Education Spending 2024 (6.8% of GDP), https://www.education.gouv.fr/depp/en-2024-1971-milliards-d-euros-consacres-l-education-soit-68-du-pib-452853
  8. DEPP – OECD Scope 2021 (5.4% vs 4.9% OECD average), https://www.education.gouv.fr/le-financement-de-l-education-en-2023-415375
  9. Inequality Observatory – PISA 2022 and Social Inequalities, https://www.inegalites.fr/pisa
  10. PISA 2022 – Café Pédagogique / OECD, https://www.cafepedagogique.net/2023/12/05/pisa-2022-la-chute/
  11. Khanmigo Official Website – Khan Academy, https://www.khanmigo.ai/
  12. Ei Mindspark – Official India Website, https://mindspark.in/
  13. Senate – Report on Ongoing Teacher Training (2022), https://www.senat.fr/rap/r22-869/r22-869_mono.html
  14. DEPP – Domestic Education Spending 2023 (189.9 Bn€, 6.7% GDP), https://www.education.gouv.fr/en-2023-1899-milliards-d-euros-consacres-l-education-soit-67-du-pib-415774
  15. OECD Education at a Glance 2023 – France Country Note, https://gpseducation.oecd.org/Content/EAGCountryNotes/EAG2023_CN_FRA_pdf.pdf
  16. CAE – Note on Public Education Spending, https://cae-eco.fr/education-comment-mieux-orienter-la-depense-publique-communique-de-presse
  17. Montaigne Institute – School Inequalities: Taking Action at the Root, https://www.institutmontaigne.org/publications/inegalites-scolaires-agir-la-racine
  18. American Economic Review – Muralidharan et al. (2019) – Mindspark, https://www.aeaweb.org/articles?id=10.1257%2Faer.20171112
  19. Légifrance – Decree n°2008-775 / n°2017-444 (primary education service obligations), https://www.legifrance.gouv.fr/loda/id/JORFTEXT000019278548
  20. Senate – Report n°869 (2022-2023) – Ongoing Teacher Training, https://www.senat.fr/notice-rapport/2022/r22-869-notice.html
  21. Wikipedia / EEF – Education Endowment Foundation, https://en.wikipedia.org/wiki/Education_Endowment_Foundation
  22. DEPP – Educational Digital and Confinement Data (2021), https://www.education.gouv.fr/la-depp-et-l-impact-de-la-crise-sanitaire-305177