Test the Grand Continent. The ideas and analyses you won’t find anywhere else. But you’ll see them everywhere soon. In your inbox starting at 8 euros per month
As debates rage about the relevance of the colossal investments needed for France and Europe to take part in the AI race, it is worth widening the perspective. Mastery of AI is indeed an existential question that probes French and European cognitive sovereignty and goes far beyond purely economic concerns.
What is cognitive sovereignty?
The answer varies depending on whether we are talking about AI, brain–computer interfaces and the analysis of brain activity, or the algorithms of social networks, online games, and betting markets calibrated to capture attention. If we focus on AI and its governance, cognitive sovereignty can be defined as “the state’s ability to possess its own intelligence-producing infrastructures as well as the capacity to protect its population from any act of cognitive warfare aimed at degrading its rationality.”
In practical terms, cognitive sovereignty first requires border AI models, data centers, and chips. It also requires training the minds of its population, especially their critical thinking, to emancipate them from any dependence on AI. It finally demands being able to monitor, detect, and counter any act of cognitive warfare targeting the population’s rationality, notably through manipulation of algorithms and AI models. It is no accident that NATO’s Allied Transformation Command has designated cognition as the sixth dimension of modern conflict alongside air, land, sea, space, and cyberspace.
This sovereignty comes at a substantial price. Initiatives like Operation Prometheus require investments of such magnitude that it is legitimate to ask whether they “are worth it.” Answering anything other than an affirmative and not committing quickly to this path would, however, amount to continuing on the current road, the one that leads to vassalage.
What is the economic cost of cognitive vassalage?
As the authors of Operation Prometheus note, “continuing to do nothing amounts to massively shorting all the American tech giants, every day.” If we remain in a state of vassalage, the rather conservative hypothesis of a 10% replacement or profound transformation of the payroll by AI would lead us, within two to three years, to permanently pay five points of GDP per year as pure rent to the United States. And this is only the beginning because cognitive vassalage also means giving up mastery over the nature, ergonomics, interface, and features of AI models that will anyway be used far more massively by the European population.
Not controlling AI means not being able to act on its very deleterious effects that are starting to be documented in several so‑called junior jobs. This is all the more serious because these jobs are essential for training experts: this is where they confront the real world and acquire the experience they need. And these experts, who play a key role in many organizations and thus in the functioning of the economy, are precisely those on whom we will have to rely to monitor AI, since they will possess the required expertise to engage with it on equal terms.
Those who do not master AI production cannot shape it as they wish. Yet AI is here for good, not by fate, but because incentives to use it are too strong and opting out would seriously jeopardize our competitiveness, our defense, and our security. Using it without producing it is thus to endure the choices of others. And given the impact of those design, operation, and development choices on our cognition, enduring them means living in a state of vassalage.
The role of nature, ergonomics, and the interface of AI models in education is becoming clearer. Some designs would lead students to shed part of their cognition, to the point of becoming much less proficient when AI is withdrawn. In the same way, some designs would sometimes render human–AI collaboration suboptimal: on certain tasks, humans would perform better alone than with AI, just as the AI would perform better without humans on other tasks.
In very simple terms, we can consider “deleterious” those designs that offer the user maximal fluency and “comfort.” They indeed encourage the user to shed their cognition in favor of AI, to develop a habit of not thinking and, above all, of not exerting effort to think. This notion of effort is central: just as physical effort is essential to the health of the body — which the advent of the internal combustion engine has not changed — cognitive effort is essential to the brain’s health. AI that lets you reap the fruit of cognitive work without any effort thus kills cognition little by little.
Conversely, “Cognitive Forcing Functions” refer to designs that reintroduce cognitive friction at the heart of the interaction with AI: formulating one’s own answer before consulting the AI’s, answering questions proving one has actually read the submitted documents, etc. These designs are beneficial because they stimulate cognitive effort while leveraging AI’s contributions.
But American tech chiefs, under enormous financial pressure due to their colossal investments, will overwhelmingly opt for “deleterious” designs, which are far more commercially profitable than beneficial but friction-generating designs. They will do so for the same reason that fast-food chains thrived on hyper-caloric menus rather than healthy salads.
It is worth noting, however, that when public debate turned to the very harmful effects of this “junk food” on health, public perception gradually evolved. Some fast-food outlets then had to modify their offerings and highlight their nutritional quality, in other words, their “taste friction.” The commercial superiority of “comfort” over “friction” is thus not a foregone conclusion. One must still be able to develop and promote beneficial designs that embrace friction…
Cognitive vassalage is the guarantee that one day one will have to pay to counterbalance the effects of deleterious designs on the education and university system, as well as to see productivity and the competitiveness of organizations that use AI in a suboptimal manner erode.
Open Weight and the Open Revolt of American Lawyers
In the United States, the extreme litigiousness of society and the pharaonic sums regularly at stake have made the elite lawyer profession extremely competitive and, in some cases, highly remunerative.
This is partly what has allowed major American firms to invest heavily in their cognitive sovereignty. Latham & Watkins, Kirkland & Ellis (500 million dollars) or Morgan & Morgan (1 billion dollars) have thus devoted substantial sums to developing their own legal AI platforms. Kirkland & Ellis and Morgan & Morgan are designing end‑to‑end proprietary agentic platforms, while Latham & Watkins has acquired so‑called “open‑weight” models — that is, weights that are open, giving access to a pre‑trained neural architecture — which it has subsequently fine‑tuned at scale on legal corpora. Morgan & Morgan would not only deploy its MX2 platform but also license it to other American firms, thus building a software intellectual property business while amortizing development costs.
Notably, these firms have made this choice even though the tech giants are, like them, American. Admittedly, their approach has nothing to do with training European frontier models, but it invites us to ask: how can Europeans, who must take defense, security, cultural, and espionage issues into account, judge that such investments are not worthwhile? If American law firms deem it worthwhile to invest billions in their cognitive sovereignty, we can gauge the scale of investments needed for a continent like Europe.
Is the cognitive vassal doomed to speak the language of its cognitive sovereign?
As Wittgenstein once wrote: “The limits of my language are the limits of my world.” Language is intimately tied to sovereignty. The heroine of the Denis Villeneuve film Arrival, a renowned linguist, says that “language is the foundation of all civilization and the first weapon deployed in a conflict.” For a long time people believed theories like the Sapir-Whorf hypothesis, which attributed to the properties of the mother tongue a significant impact on the cognition that develops afterward. This hypothesis, highly deterministic, held that certain languages prepared the mind especially well for certain kinds of complex reasoning while others constrained these abilities.We have largely moved beyond that: contemporary neuroscience speaks of “linguistic relativity” and regards the mother tongue as a factor that partly shapes the mind, without a hierarchy among languages or the idea that some would “prepare better” for complex reasoning. Nevertheless, these studies show that language directly impacts cognition. And LLMs are, in a sense, a language — a written language not under the Coupole of the French Academy, but in the offices of American tech chiefs.
Lera Boroditsky studied how English uses horizontal metaphors to talk about time (looking forward to the next game, moving the meeting back), while Mandarin uses vertical metaphors (shàng, “above,” for past events, and “below” for future events). According to her, these differences confer no “advantage,” but predispose native English speakers to organize a calendar horizontally and native Mandarin speakers to organize it vertically.
Jonathan Winawer, for his part, observed that Russian, whose lexicon distinguishes light blue from dark blue as two distinct colors rather than as two shades of the same color, predisposes its native speakers to be more sensitive to blue nuances: they distinguish them much faster than English speakers.
Finally, Stephen C. Levinson, then a researcher at the Max Planck Institute for Psycholinguistics, published in 2003 an article and a monograph entitled “Space in Language and Cognition: Explorations of Cognitive Diversity.” There he studies the influence of a language’s spatial lexicon on spatial perception, as well as on how native speakers represent and navigate space.
LLM designers have wide latitude to introduce modifications that could affect users’ cognition. And those modifications are all the more pernicious because they are subtle. It is not about renaming the Gulf of Mexico or Lake Ontario, but about introducing nuances in a model that locally alter users’ perception of a concept. Opting for cognitive vassalage is thus also giving up control over the language through which an increasing portion of the population interacts, and leaving it exposed to a form of manipulation that is intrinsically difficult to detect in time.
European languages, other than English and Chinese, could also be directly attacked. Nothing subtle here: it would simply involve exercising domination by imposing the dominant language and marginalizing the language of the vassals. AI is not equally effective in all languages; it is, for example, less effective in Polish than in English . This gap today is essentially explained by semantic questions, the structure of languages, the volume of well-organized data available online in a given language, and tokenization — the splitting of each word into computational units by the model. This splitting is optimized for English, so a language with complex declensions like Polish is cut into a much larger number of fragments to say the same thing, which degrades performance.
The Trump administration — which ordered in June 2026 Anthropic to suspend non‑Americans’ access to its latest models, Fable and Mythos, before restoring it — has recently pressured Canada to sharply reduce its support for Francophone audiovisual programs. The United States had until then no official language: on March 1, 2025, Donald Trump signed a decree granting this status to English. Why stop there? What would prevent a future administration from officially or unofficially ordering tech giants to make their models markedly less effective in other languages? Without making them unusable, they could sow doubt that would push many users to express themselves only in English to obtain “the best performance.”
What is the security and human cost of cognitive vassalage?
Unlike American lawyers, Europeans must take security issues very seriously, as evidenced by the October 12, 2026 cyberattack on the refrigeration facilities of many French supermarket chains. While details are not yet fully known, there are already Iranian‑made attacks prepared with Anthropic’s Fable model, and others launched autonomously by AI agents. Cyber, like cognition, is, indeed, one of the dimensions of conflict identified by NATO: controlling one’s cyberspace becomes as essential as controlling one’s airspace.
What would happen tomorrow if a country launched a massive cyberattack using an American AI, not against supermarket fridges but against our critical infrastructures—nuclear power plants, hospitals—and the White House ordered the company developing this AI, Anthropic for example, to suspend France’s access to the model it would need to defend itself?
And if a U.S. administration made frontier-model access a bargaining chip to compel France or another European country to buy American military equipment or to sign a treaty highly favorable to the United States? No one has forgotten the images of Ursula von der Leyen visiting Donald Trump’s golf course in Scotland to sign the deal meant to end tariff escalations. How far would the American president have pushed his demands if Europe had a vital need for the best American AIs to defend itself?
Finally, how can we ignore all the recent findings about how Mark Zuckerberg makes trade‑offs regarding features, design, and ergonomics of Meta platforms such as Facebook and Instagram? Even after his own researchers demonstrated to him—supported by internal studies—that certain designs exposed minors to predation by unknown adults and gravely harmed the mental health of many users, especially young girls, he categorically refused any changes. European vassalage to American social networks thus allowed a CEO to make trade-offs, in the name of his own economic interests alone, that directly harmed the health of many young Europeans.
These initial human costs are compounded by the brutalization of public debate. Lawsuits against Mark Zuckerberg also revealed that Facebook researchers questioned European policymakers who complained about being “forced to appear increasingly violent and increasingly radical” to be heard by voters on social networks. These elected officials had pleaded in vain with Facebook to modify its algorithms so as to stop rewarding hateful speech, in order to offer a quality public debate.
The absence of choice is actually a choice
In just a few years, frontier AI has progressed from solving olympiad-level math problems for high school students to the mass production of meaningful research results, to the extent that French Fields medalist Hugo Duminil-Copin could assert that “AI has annihilated his field of research.” And this is happening while some leading mathematicians seem barely to begin to glimpse what is happening — not because they are poorly informed, but because everything has progressed at an extraordinary pace.
In this context, “taking one’s time” to decide whether to reclaim cognitive sovereignty amounts to choosing vassalage. While we deliberate, models keep advancing and the bill keeps rising. Young people as well as European companies already use AI extensively, and the job market is already feeling the effects. The default route leads to total dependence on Sino‑American AI hegemony — a dependence whose real cost will only become apparent in the wake of a major crisis.
Yet first steps toward cognitive sovereignty have been taken. With the release of Mistral Large 4 on October 6, 2026, which places France in the third position worldwide for AI, Mistral shows that, in this race, anything remains possible.