The PISA score gap between a privileged student and a disadvantaged student reaches 113 points in France. In Germany, this same gap stands at 86 points. The difference is explained neither by the average level of teachers, nor by the volume of public spending — France spends more per student than the OECD average. It is explained by the way the school institution treats social heterogeneity from the earliest years. The arrival of artificial intelligence in classrooms does not render this question obsolete. It makes it urgent.
Two readings of this fact coexist today, each supported by serious arguments. The first contends that pedagogical personalization enabled by AI can correct what uniform schooling has never managed to resolve: the mismatch between a program designed for an average student and the reality of each child. The second responds that technological personalization, left without a strong institutional framework, only accelerates the fragmentation of trajectories according to parental income. Data allows us to partly settle the matter — and Estonian results provide an insight that neither camp exploits sufficiently.
The Essential Points
- The PISA gap between privileged and disadvantaged students is 113 points in France, versus 86 in Germany; France prepares 73% of its school-age human potential, Estonia 88% (OECD PISA).
- Estonia did not bet on technology first: it universalized public kindergarten and massively trained its teachers before deploying digital tools.
- Tutoring AI reduces learning inequalities when it integrates into a school that holds its students; it amplifies them when it substitutes for this presence.
- The real tipping point for the generation born after 2020 is not the technological tool but the political decision on early childhood support.
Babeau Is Right on the Diagnosis, Wrong on the Remedy
Olivier Babeau has been developing for several years a coherent critique of the French school model: a school designed for industrial reproducibility, which values memorization over comprehension, conformism over initiative, and which resists any individualization of pathways in the name of formal egalitarianism. The argument deserves to be taken seriously. National grading, rigid curricula, evaluation by average rather than by progress — these mechanisms were constructed to produce homogeneous cohorts, not to accompany individuals.
AI does indeed change something in this equation. Adaptive tutoring systems, whose most serious evaluations show real gains in foundational learning, allow a student to progress at their own pace, to repeat an exercise without the gaze of a peer, to receive immediate feedback without waiting for collective correction. Experiments conducted in India with Mindspark, evaluated under controlled conditions, showed learning gains on the order of 0.37 standard deviations in mathematics over a few months. This is not marginal.
But the leap that Babeau makes — from “AI improves individual learning” to “traditional curricula are obsolete” — does not withstand the data. Because the decisive question is not: does AI help a student learn? It is: does AI help in the same way a student whose parents can supervise use and a student left to themselves? The answer is no, and the numbers show it.
What Estonia Teaches France
Estonia prepares 88% of its school-age human potential. France, 73%. This indicator, developed by the World Bank, measures the probability that a child born today survives to school age, is enrolled in school, and acquires foundational learning. The fifteen-point gap between the two countries is not anecdotal — it represents, at the scale of a generation, a considerable mass of unrealized human potential.
Estonia is often cited as a model for digitizing education. This is true, but it is only half the story. The real Estonian lever precedes the digital. The country universalized access to kindergarten from age three, invested massively in initial and ongoing teacher training — who are among the best trained in Europe — and reduced resource inequalities between schools before deploying technological tools. Digital technology was grafted onto a solid institution, not substituted for it.
France operated in the reverse order. It distributed tablets, experimented with adaptive platforms, launched successive digital plans, all while maintaining structural inequalities in kindergarten access in the most disadvantaged territories and notoriously insufficient teacher training in differentiated pedagogical practices. The result is predictable: technology amplified existing gaps rather than reducing them. Families who knew how to use it benefited; others received an additional tool without the institutional instruction manual.
This lesson aligns with what we observe more broadly on the link between institutions and economic performance: the tool is worth only what the framework carrying it is worth. The pioneers of all-digital who today forbid social networks before fifteen understand this from their own experience: unframed technology produces effects opposite to those expected.
The 113 PISA Points Reflect a Political Choice, Not a Fatality
The 113-point gap between privileged and disadvantaged students in France is not a natural datum. It is the result of a series of institutional choices accumulated over several decades.
The first is the persistence of the school assignment model combined with strong residential segregation. When the most privileged students concentrate in the same schools, peer effects mechanically amplify entry inequalities. The OECD documents this phenomenon consistently: France is one of the developed countries where social origin best predicts school results, more so than in most Nordic countries or than post-reform Germany.
The second is the insufficiency of early childhood support. Cognitive inequalities deepen massively between ages zero and six. Children who arrive in first grade with limited vocabulary, insufficient exposure to written language, without emotional regulation learning, incur a delay that primary school almost never closes. France maintains a theoretically universal kindergarten but whose quality varies considerably depending on territories, and whose class sizes remain high in the most disadvantaged areas.
The third is the treatment of heterogeneity in class. Unlike other systems that explicitly train teachers in differentiated pedagogy, the French model largely relies on uniform front presentation, with correction by average that renders both the most struggling and most advanced students invisible.
These are the three choices that AI can, depending on conditions, attenuate or worsen.
AI Amplifies What the Institution Chooses to Do
The question is not whether AI is good or bad for education. This is a poorly posed question. AI amplifies existing dynamics. In a school that already individualizes, that has trained its teachers to support different pathways, that frames the use of tools, technology can free up teaching time, allow finer monitoring, reduce certain effects of uniform grading. In a school that has not done this preliminary work, it becomes a tool of differential autonomization: students capable of self-discipline, supported by parents who know the right tools, progress; others drop out faster, without the safety net that the classroom collective represented.
This asymmetry is documented in the first available data on AI tutors in countries with high school inequality. An MIT report on Khan Academy Khanmigo deployments in low-income American schools showed that the tool reduced learning inequalities when teachers were trained in its integration and maintained an active support role. In classes where the tool was deployed without prior training, results for the most fragile students stagnated while those of the most autonomous students progressed.
Baptiste Larseneur formulates this challenge with precision: institutional schooling is not an obstacle to technological personalization, it is its condition. This is not a reactionary position. It is a coherent reading of what the data produces when you remove the collective framework from an equation that has not yet produced equality without it.
The Liberal Argument Has a Limit That Markets Don’t Resolve
Babeau’s argument has a more radical variant, which he does not fully carry but which others defend: if AI allows individualized learning at low marginal cost, why maintain the public school monopoly? Diversification of pathways, recognition of competencies acquired outside the system, validation by employers rather than diplomas — so many directions that the educational platform economy tends to push.
The argument runs into a structural problem that markets do not spontaneously resolve: asymmetric information about learning quality. When a family chooses an educational platform for their child, they do not have the instruments to evaluate the real pedagogical effectiveness of the tool. Labels, third-party evaluations, independent certifications either do not exist or are captured by the actors they are supposed to rate. This is precisely the argument Jean Tirole develops on regulating markets with asymmetric information: the absence of a trusted institution does not produce optimal allocation, it produces the market for lemons.
For the generation born after 2020, this point will be decisive. If France does not invest now in teacher training for AI pedagogical integration, in rigorous evaluation of deployed tools, and in extending quality early childhood support in under-resourced territories, it risks looking back in ten years at a PISA gap that will have risen from 113 to 130 points. This is not catastrophic projection — it is the trajectory observed in countries that deployed technology without reforming the institution. Irreversibility plays here: cognitive inequalities that form between ages zero and six are the hardest to correct afterward. Each cohort that passes without quality early childhood support is a definitive loss, not a recoverable delay.
The tipping point is known. It is not technological. It is budgetary and political: to what extent does France accept concentrating its educational resources on the early years and most disadvantaged territories, instead of distributing them equally to all? This is the reform that Nordic countries made in the 1970s and 1980s, and that Estonia adapted in the 1990s. This is not an outdated model — it is an entire remaining priority.
On the links between institutional choices and long-term economic performance, Philippe Aghion’s reflection on Schumpeterian growth provides useful insight: investment in human capital is the condition of creative destruction, not its adversary. An economy that automates without training loses the workforce capable of managing, improving, and redirecting machines. ECB data on AI and employment point in the same direction: it is the best-trained workers who benefit most from AI, not those who needed it least.
What Estonia Should Inspire in France
Estonia does not prove that technology saves schools. It proves that institutions can be reformed without being abandoned, and that technology becomes a lever when it arrives second.
France has the resources to correct the three identified deficits. Money is not lacking globally — what is lacking is allocation. French educational spending is concentrated on secondary and higher education, where inequalities are already set. Shifting resources toward primary and pre-primary, massively training kindergarten teachers in language and emotional regulation practices, seriously framing the deployment of digital tools with randomized evaluations rather than uncontrolled experiments — none of these actions requires waiting for the next technology.
Babeau raises a real question when he interrogates curriculum rigidity. But the answer to this rigidity is not the dissolution of the school institution in favor of unframed personalized learning. It is a reformed institution, more flexible in its pedagogical methods, firmer in its demands for equal access to quality support. The tension between these two positions is not metaphysical. It is resolved through specific budgetary choices, and time is beginning to run short for the generation entering kindergarten this year.
Sources
- Observatory of Inequalities — PISA France data
- OECD, PISA 2022 Results, Programme for International Student Assessment — data on socioeconomic gaps by country
- World Bank, Human Capital Index 2024 — indicator of school-age human potential preparation (France 73%, Estonia 88%)
- Muralidharan, K. & Sundararaman, V., The Impact of Diagnostic Feedback to Teachers on Student Learning, evaluation of the Mindspark program in India, J-PAL
- Institut Montaigne, Vaincre les inégalités scolaires — report on the concentration of French educational resources
- Khan Academy / MIT, preliminary data on Khanmigo deployments in low-income schools, 2023-2024
- Aghion, P. & Howitt, P., The Economics of Growth, MIT Press — Schumpeterian framework applied to human capital investment