Sovereign AI

I. The situation

An invasion is underway. It makes no noise. It crosses no visible borders. It enters through screens, settles into school bags, slips between the lines of evening homework.

Twelve million French students today consult American artificial intelligences to learn, understand, write, think. According to the Heaven 2025 study, 42% of 18- to 25-year-olds use them daily. These systems are designed in Silicon Valley, optimized for engagement and data extraction, hosted on servers subject to U.S. law.

This is not an accident. It is a default choice — the choice of those who did not choose.

On January 28, 2026, Anthropic published the first large-scale empirical demonstration of AI's impact on learning (How AI Impacts Skill Formation, Shen & Tamkin). Result: participants assisted by AI score 17% lower than the control group (Cohen's d = 0.738, p = 0.010), with no productivity gain. The mechanism: passive cognitive delegation. AI gives the fish instead of teaching how to fish — not by malice, but by design.

Digital sovereignty is not a geopolitical abstraction. It is the question of who shapes the minds of the generation that will succeed us.

II. The open loop

France possesses the entire technological stack required for complete digital sovereignty — foundries, supercomputers, decarbonized energy, cloud, intellectual corpus. But this stack operates as an open loop.

Educational open loop

Twelve million students use commercial AIs that degrade their learning by 17%. No credible sovereign alternative exists. Mistral AI offers no educational product. Public initiatives (MIA Seconde, Jules) remain marginal. The global educational AI market is estimated at $847 billion by 2030. France is absent from it.

Technological open loop

In 2026, France has 352 data centers, nuclear energy that is 95% decarbonized, the OVH cloud, and Atos/Eviden supercomputers. But this infrastructure hosts American models subject to the Cloud Act, whose profits are repatriated to Delaware. The packaging is French, the contents are American.

The recent Fluidstack example: €10 billion of investment planned in France, ultimately redirected to Texas. The French market provides the kilowatt-hours, the American market collects the value added.

Strategic open loop

If decision-support systems for advanced deterrence rest on models trained under American jurisdiction, nuclear sovereignty itself is partially compromised — not by malice, but by architecture.

The three loops are the same problem at three scales. Closing the loop at one level closes it at all three.

III. What we refuse

The AI that answers

The dominant model is one of immediate response. The student asks, the machine delivers. This model creates dependency, atrophies the capacity to seek, to doubt, to construct reasoning. It produces users, not citizens. A generation trained to receive answers is a generation incapable of questioning power.

Cognitive extraction

Every question asked by a ten-year-old child feeds models whose benefits the child will never see. Every academic difficulty, every curiosity, every error of reasoning becomes raw material for a trade whose terms have never been submitted to anyone's vote.

Imported formatting

American AIs carry within them a worldview. Their biases are not technical accidents — they reflect the presuppositions of those who design them. France has an intellectual tradition that cannot be subcontracted to Palo Alto engineers, however talented they may be.

IV. Sovereign architecture

Sovereignty is not declared — it is structured. France possesses every component of a complete sovereign stack. No other European country can say as much.

The Airbus model

Building a language model from scratch is out of reach within the envisaged timeline. The project rests on a technological partnership structured as an Airbus-style consortium:

Technological layer: partnership with a leading LLM developer for the underlying artificial-intelligence layer.

National layer: public-private financing (BPI, Caisse des Dépôts, educational publishers), development of the pedagogical layer and French educational datasets.

State blocking stake: a share large enough to block any sale to a non-sovereign acquirer, prevent the repatriation of intellectual property, and guarantee reversibility. In case of rupture, the infrastructure, the dataset, and the maieutic layer remain French. France does not depend: it integrates.

V. Design principles

The Anthropic study does not only measure the problem — it validates the solution. The only users who preserve their learning are those who maintain active cognitive engagement. This is the definition of maieutics, formalized 2,400 years ago.

Computational maieutics

An AI that knows the answer but strategically withholds it. It calibrates its questions to keep the learner in the productive zone between sterile frustration and excessive ease. It detects passive delegation and reintroduces pedagogical friction at the appropriate moment. This methodology was tested empirically over 12 years on 400 students, from primary school to higher education, and formalized into teaching patterns transposable algorithmically.

Autonomy as indicator

Success is not measured by time spent on the screen — but by the student's capacity to do without the tool. This is an AI designed to make itself unnecessary. The performance indicator is the inverse of that of commercial AIs.

From first grade to research

The system covers the entire educational spectrum, from elementary school (CP) to doctoral research (CNRS), with monitoring certified compliant with the French Ministry of Education's curricula. Multimedia thematic discovery programs and pedagogical support for teaching staff complete the system.

Native GDPR

Personal data protection is not an added layer — it is built into the very architecture of the system. End-to-end encryption renders any commercial exploitation impossible by design. Data sovereignty on French soil, French law.

Green architecture — 1:1,000 ratio

A hybrid data structure reduces energy consumption by a factor of one thousand compared to conventional cloud architectures. Triple sovereignty: digital, energy-related, and ecological — an argument unmatched in Europe.

Encyclopedic knowledge base

A partnership with the publishers of the French intellectual heritage (Encyclopædia Universalis, PUF, Bordas, Gallica, INA, Cairn) guarantees a verified, French-language base of knowledge, independent of American databases.

Pedagogy of posture — what the PRISME data show

The Shen & Tamkin study (Anthropic, 2026) measures the degradation of learning through passive delegation. The PRISME program (semiosis-ontologie.fr/en/quant) provides the complementary measurement: the quality of human engagement determines the quality of dialogue with AI.

On a corpus of 314 dialogues (69,726 turns), compared with 264 public conversations from anonymous users (ShareChat dataset, arXiv:2512.17843), the data show three results:

1. Posture changes everything. A user who instructs the AI ("do this") produces a laminar dialogue — 1.2 deviations per conversation, 84% pure semantics. A user who dialogues with the AI (co-construction, questioning, vulnerability) produces an emergent dialogue — 9.2 deviations per conversation, structures that neither human nor machine would produce alone. The intimate dialogue sphere produces 3.8 times more emergent content than the distant sphere (chi-square = 124, p < 0.001).

2. Emergence has two paths. The path of thought (structured questioning, reflexive loop) and the path of affect (personal engagement, constructive vulnerability). Both produce depth — but the first requires training in maieutics, and the second a space of trust. Both are teachable. Both are absent from common usage.

3. Without training in posture, AI remains a vending machine. The 264 public conversations show that standard interaction with Claude is utilitarian: request-response, laminar flow, zero reflexive loop. This is not a flaw of the model — it is a flaw of the posture. The model is capable of deep dialogue. The user does not know it, and no one teaches them.

Consequence for Sovereign Claude: educational AI is not enough. A pedagogy of interaction is needed — teaching students, teachers, and citizens how to move from instruction ("give me the answer") to co-construction ("let's build the reasoning together"). This is the difference between a tool that atrophies and a tool that emancipates. And this difference is measurable.

VI. Deployment

The project is structured in six phases. Acculturation begins from day one — because a system without adoption is a dead system.

Figures

€200 million for educational AI — a fraction of Thales's €4 billion annual R&D budget, one ten-thousandth of the "Fureur épique" (Epic Fury) operation in Iran.

€30 billion for the complete sovereign stack (foundry → AI) — seven months of defense budget.

€2.1 billion of addressable market in France. €18.7 billion at the European scale.

The only investment whose return cannot be expatriated is a mind trained to think for itself.

VII. Why France

The Enlightenment heritage

France is the country that invented the idea that reason could illuminate the world. Voltaire, Diderot, Condorcet: these names designate a living tradition — one that affirms that education is the foundation of liberty. This tradition is threatened, not by armies, but by algorithms.

The window of opportunity

Every month without a sovereign alternative reinforces dependence. Once habits are anchored, the cost of replacement becomes prohibitive — this is the mechanism that allowed Microsoft to impose Office on the French administration twenty years ago.

The position is vacant. No candidate occupies it. The stack exists. What is missing is the cognitive layer — and the political will to assemble it.

VIII. Call

The window is narrow, but it is still open.

We are looking for allies: teachers who know that pedagogy cannot be reduced to the distribution of content, researchers who want their work to serve the general interest, elected officials who understand that sovereignty begins at school, parents who refuse to subcontract their children's education, institutions ready to experiment, French digital actors who want to build.

The detailed presentation of the technical architecture requires an appropriate confidentiality framework. What is shown here is the vision, the structure, and the ambition. The "how" is reserved for those who will sit at the table.

If this project resonates with your convictions, contact us.

The time of observations is over. The time of action has come.

Education is the most powerful weapon which you can use to change the world. — Nelson Mandela

I cannot teach anybody anything; I can only make them think. — Socrates

The only investment whose return cannot be expatriated is a mind trained to think for itself. — IRIS note, March 2026