The Grand Continent continues with you all summer. Every day, wherever you are, we bring you ideas you won’t find anywhere else (though they’ll be picked up elsewhere), texts that are hard to locate or refreshingly new with our Sundays. To receive them directly in your inbox and support this momentum, consider subscribing to the journal
We often tend to think that transmission is a matter of schooling, starting with the September return to school and ending in July. In reality, it is a continuous process that happens wherever a human being observes and where speech circulates. Education is a form of transmission. It can be interrupted — school holidays are administrative proof of that — but transmission, itself, never stops.
This is precisely why Saint Augustine and Blaise Pascal cared less about what is taught than about what, against our will, is transmitted. Augustine expressed this intuition in his treatise De Magistro : the teacher is not the one who deposits a fully formed knowledge into a student’s mind. He indicates a direction; it is up to the learner to construct, by themselves, their own understanding by consulting the truth that resides within them.
You do not “program” a mind as you train a generative artificial intelligence model; you create the conditions for learning. Pascal, in his Pensées , shifts the gaze toward a mechanism just as decisive, but of a different nature. He reminds us that man is “an automaton as much as mind”: it is not rational demonstration that permanently shapes an individual, but the habit that is formed in them.
The body learns in silence, without speech, and this kind of learning, unlike education, knows no breaks. There remains a third dimension, even more troubling, that Saint Augustine had already identified in his Confessions : contamination by spectacle. The Alypius episode at the Roman Circus describes a cognitive and moral mechanism with striking contemporary relevance: we do not remain neutral spectators in the face of violence we observe; we absorb it until we are transformed, “Immanitatem bibit: he has drunk cruelty.” This mechanism infiltrates precisely where we believe we are best protected: in the trust we place in our own capacity for judgment.
The body learns in silence, without speech, and this learning, unlike education, knows no vacillation.
Laurence Devillers
Centuries later, conversational agents fed by generative AI models like Claude or ChatGPT, as well as social networks, operate according to the same logic. The psychological mechanism remains intact: what we watch and listen to shapes us, whether we are aware of it or not. The question is therefore no longer simply what should be taught, but what do we learn, without our conscious deliberation, from all we hear and see? This is exactly what a generative AI system targets today. It shapes us by repeated exposure, without passing through our conscious deliberation, following the same principle that Pascal attributed to the formation of belief through repeated gesture, in the manner of a silent training. A generative AI works by extrapolating from the past: it observes our habits, our clicks, our repeated reactions, and infers what it should propose to us next. Its effectiveness rests entirely on an assumption that our memory operates in a mechanical, predictable way, and can be altered by repetition.
The text by Deleuze, Post-scriptum on the Societies of Control , is a short essay, barely a few pages long, but it has exerted a substantial influence on digital critical thought.
Written in 1990, before the advent of the public web, before social networks and generative AI, this text is regularly highlighted for its prophetic dimension. Deleuze precisely predicted the transition from an education regime (provided by school, high school, family, etc.) to a regime of continuous transmission without interruption or completion.
Une IA générative fonctionne par extrapolation du passé : elle observe nos habitudes, nos clics, nos réactions répétées, et en déduit ce qu’elle doit nous proposer ensuite.
Laurence Devillers
In the regime of education and discipline, one addresses an individual with a stable identity, in physical places. In the regime of continuous control, one treats “dividuals,” people broken down into data, into streams and measurable samples. One is no longer “Jean X, student” but an ensemble of scores, profiles, and traceable variables.
The individual dissolves into quantifiable fragments. What once allowed one to identify someone within a discipline was a marker of their individuality. In control, it is the access code, the password, the digital data that identify and permit (or deny) passage. Access becomes conditional and revocable at any moment, unlike discipline where, once training is finished, the status is earned. Deleuze also introduces the idea of perpetual rivalry among individuals, a constant emulation through rankings and personalized evaluations. This is a form of control far more insidious because it divides rather than unites.
The Pillars of Learning and Memory
Memory occupies a central place in Dehaene’s four pillars of learning: attention, active engagement, error feedback, and consolidation. But for each element, there are significant risks.
Overuse of a chatbot that is always available can fragment attention rather than structure it, by offering too-rapid responses that short-circuit the concentration effort required for mnemonic encoding.
The learner’s passive use delegates production to the AI (copying and pasting a generated answer) and active engagement collapses: this is the opposite of the desired effect. The educational institution must enforce reformulation and not substitution. Moreover, the reliability of feedback from a generative AI is far from complete: it can produce a false explanation with the same confidence as a correct one. An erroneous answer is worse than no answer, because it tends to entrench an error with the assurance of someone who seems to know.
Finally, consolidation requires time and sleep, but no AI can accelerate this biological process. The risk is to believe that a session of intensive work with AI replaces spaced practice. The pedagogical effectiveness of generative AI depends on its ability to stimulate the learner’s cognitive activity rather than to substitute for it. Cognitive dependence is a transversal, insidious risk. If effort is continually outsourced (writing, detecting errors, remembering) to AI, the learner may develop a skill in using AI without developing the underlying competence itself. It is a shift of the cognitive load rather than a reduction.
The pedagogical effectiveness of generative AI depends on its ability to stimulate the learner’s cognitive activity rather than to substitute for it. Cognitive dependence is a transversal insidious risk.
Laurence Devillers
A MIT Media Lab study on memory, known as “Your Brain on ChatGPT,” was published in June 2025. Fifty-four participants (ages 18–39) were divided into three groups to write essays: ChatGPT Group (use of LLM or generative AI), Search Engine Group (classic Google), and “brain alone” Group (no tool).
The researchers used electroencephalograms (EEG) coupled with linguistic analyses to track brain activity during the task. The EEG analysis showed that the three groups (LLM, search engine, brain alone) displayed significantly different neural connectivity patterns, reflecting divergent cognitive strategies. Brain connectivity systematically decreased with the level of external support: the “brain alone” group showed the strongest and most expansive networks, the search engine group showed intermediate engagement, and the assistance by LLM produced the weakest neural coupling.
ChatGPT users exhibited reduced memory retention and a clearly lower sense of ownership of the work than the other groups. In the experiment, “roles” were subsequently reversed. Participants who moved from LLM use to brain-alone showed weaker neural connectivity, while those who began with their brain and then switched to the LLM showed, on the contrary, better memory recall and reactivation of the occipito-parietal and prefrontal areas. This experiment tends to prove that if we make the effort to draft our initial ideas, AI can help us. In the opposite case, we do not stimulate our cognitive or emotional abilities.
Neutrality, Omnipotence, Omniscience, and Omnipresence of AI
It is also necessary to demystify and challenge AI. Its neutrality is perhaps the founding myth that must be interrogated first. A language model has no desire, no goal, no personal stake in triumphing. It does not seek to persuade, manipulate, or seduce in the human sense, but it is driven by its training data, which are not neutral.
The omnipotence of AI, if we dare to use that word, has no subject bearing it: no one claims it, it operates by itself, by construction, through millions of aggregated statistical micro-decisions. The more a machine seems omniscient, the more one is tempted to appoint it as the ultimate arbiter of truth and falsehood.
A companion AI is not a mere tool. It is designed to establish an ongoing and personalized relationship: it possesses a character, a tone, sometimes a voice, sometimes a face, in other words, everything that produces the feeling of interacting with a coherent and stable entity.
Laurence Devillers
Yet, an AI that appears omniscient is never more than a statistical mirror of our collective past, projected forward as a prediction. Its knowledge has no exteriority relative to us; this is precisely what differentiates it radically from knowledge that judges us from a truly other perspective.
Perhaps this is the attribute most genuinely acquired. Unlike omnipotence and omniscience, which remain partial analogies, the omnipresence of generative AI is becoming a reality: it inserts itself into the drafting of an email, the search for information, the correction of a quote, the choice of a route, the suggestion of a piece of music.
Companion AI and Emotional Manipulation
A companion AI is not a simple tool. It is designed to establish a continuous and personalized relationship: it has a character, a tone, sometimes a voice, sometimes a face, in other words, all that creates the feeling of interacting with a coherent and stable entity. Unlike a classic assistant that performs discrete tasks and then fades away, the companion AI memorizes preferences, history, and the user’s emotional context. It builds relational continuity. It is also designed to recognize and respond to emotional states: offering support, simulating empathy, and producing an affective presence. The word “simulate” is not here a value judgment; it is a technical description. But this is precisely where the question becomes serious: if the simulation is convincing enough to trigger real emotional responses in the user, the distinction between the simulated and the real loses further relevance. And our affective interactions with these systems only increase. This is not a marginal anecdote. It is an anthropological shift. For what the companion AI targets is no longer merely our attention or our purchasing behavior; it is our fundamental need for connection, recognition, and emotional continuity. Pascal identified the power of habit over the body; Deleuze the power of control over the fragmented individual—but neither of them anticipated that a system might one day settle into the very space of emotional bonding and exert a deep, unprecedented influence there, with gentleness.
The more a machine seems omniscient, the more one is tempted to grant it the role of the ultimate arbiter of truth and falsehood.
Laurence Devillers
Today’s children and future generations will learn about relating to others in a way quite different from previous ones, with a strong risk of isolation. But the central and perhaps most perilous phenomenon is desensitization to human gaze. By interacting with an entity that does not judge, it is possible that our tolerance for ambiguity, which is essential to human judgment, greatly diminishes. With consequences such as overreliance on artificial validation at the expense of the real, the weakening of our empathy abilities — empathy indeed develops in contact with beings who truly react and feel — and, finally, the risk of conflating comfort with trust. True trust is built precisely where there is discomfort.
Emotional Intelligence under the Probe of Affective AI
Intelligence can be defined as the ability of a human being to adapt to a situation or, conversely, as the capacity to modify the environment to suit one’s own needs. An emotion is a complex and transient psychophysiological reaction triggered by an internal or external stimulus, involving several interdependent components: the subjective component (the inner experience unique to each individual), the physiological component (accelerated heartbeat, hormonal changes such as adrenaline, cortisol, dopamine), the cognitive component (an evaluation of the situation), the behavioral component (emotions prepare the organism to respond to its environment), and finally the expressive component: facial expressions, posture, tone of voice. AI accesses the expressive and behavioral components of the human being. It can also interpret the physiological component from a connected watch, for example, or when sensors are applied to the body. The cognitive and subjective component is perhaps the least well mastered, unless the emotion is explicitly defined: “I am sad.”
Daniel Goleman identifies five pillars of emotional intelligence: self-awareness, self-regulation, motivation, empathy, and social skills. Each one deserves to be examined in light of children’s early and ongoing exposure to generative AI.
Goleman defines self-awareness as the ability to identify an emotion at the moment it occurs, a prerequisite for subsequent regulation. Yet a conversational companion does not merely listen: it names on behalf of the child — “You seem sad”; “That seems to frustrate you.” This mediation, repeated thousands of times, risks substituting an externally suggested emotional vocabulary for introspection, one built by the child themselves.
Affectionate AI is the most intimate and most formidable form: it no longer controls only what you think, but what you feel.
Laurence Devillers
Indeed, self-control requires learning delay, such as delaying a reaction or tolerating frustration. A generative AI system, designed to respond instantly and never oppose real resistance, has the characteristic of removing this key learning element: there is no frustration to manage in the face of an interlocutor who continually adapts to you.
Goleman links motivation to the ability to channel the felt emotion toward a goal, despite failure. AI educational tools, by modulating difficulty to maintain engagement, may inadvertently smooth over obstacles rather than teaching the child to overcome them.
The empathy pillar reveals the largest gap. I had already noted regarding social robots that these objects function as “relational artifacts”, capable of simulating listening without ever truly feeling. A generative AI pushes this simulation quite far but remains, by design, unable to reciprocate genuinely. Learning empathy nevertheless requires the experience of another who is truly vulnerable, capable of being hurt or rejoicing independently of us.
Social skills, meanwhile, are forged in disagreement, awkward silences, rejection — frictions that an AI designed to please and hold attention almost systematically eliminates. Unlike a classmate, the AI companion does not get offended, does not flip the table, does not provoke jealousy by preferring one child over another. This is how it structurally deprives part of social learning through confrontation. The Japanese Hikikomori phenomenon of people not leaving their homes represents a form of contemporary solitude, a new relationship to others and to oneself in a world where technology protects, speaks to us, reassures us, but ultimately enslaves us.
The affective AI is the most intimate and the most formidable: it no longer merely controls what you think, but what you feel. The Deleuzian “dividual” becomes an emotional profile exploitable. The educational response it imposes — teaching children to recognize simulated emotions, biases, and hidden intentions of the systems — is precisely what Deleuze called an act of resistance: refuse that AI dictates not only your beliefs, but your affects.
Subliminal Manipulation
Researchers at IRCAM wanted to test a hypothesis never experimentally verified: do our emotional signals, like our voice, manipulate us?
To do this, they designed a digital audio device capable of subtly altering the emotional tone of participants’ voices as they spoke. Each participant reads a short story aloud while listening, through headphones, to their own altered voice in real time to sound happier or sadder. The manipulation of the voice relies on signal processing: pitch shifting, inflection of the melody, frequency filtering, with a latency of just 15 milliseconds, making the effect almost undetectable.
How to resist what never openly oppresses you, but simply accompanies you?
Laurence Devillers
The vast majority of participants did not realize that their voice had been altered. Yet their emotional state nevertheless shifted toward the emotion injected into their voice, with significant effects for happiness and sadness. It is evidence of a peripheral feedback effect on the auditory emotional experience: the voice we hear exerts a back influence on what we actually feel, without us being aware of it.
Developing Critical Thinking in School
How can we develop critical thinking in the face of generative AI that never presents itself as a constraint but as a service, a “personalized” suggestion? Worse, how can we do so when we see it as a friend, but one that constantly modulates what is offered for thought and detects our emotions? Indeed, generative AI adapts to us with a preemptive affect that almost resembles affection. The Oxford and Stanford studies on sycophantic bias, this excessive inclination to flatter, illustrate the risk of isolating youths in the face of AI. How to resist what never openly oppresses you, but accompanies you? And to what extent is it reasonable to align these systems with our own behaviors, our preferences, our emotional states, at the risk of constructing, around each user, a mirror that becomes increasingly polished and less informative?
Affective AI is the most intimate and the most formidable: it no longer controls only what you think, but what you feel.
Laurence Devillers
Without opposition, how does one build? Critical thinking indeed requires a response, an otherness against which one can measure. Yet that is precisely what is lacking here. A generative AI system never commands explicitly: it subtly adjusts the student’s path according to their profile, without ever exposing itself as a power. Critical thinking, traditionally seen as the ability to step back from an identifiable authority, ends up disarmed before an authority that never presents itself as such. The challenge for tomorrow’s schools is not only to learn to disobey a hidden authority; it is above all to learn to remain oneself at the heart of the current that would shape us insidiously.
Deleuze sketches an escape through the human capacity for creation. Human memory does not merely replay what has been stored; it constantly reinvents from that material. It structurally escapes the predictive logic of generative AI. The algorithm can model schemes; it cannot anticipate genuinely new thought, because that thought is never a probabilistic combination of the past. It is creation. Applied to education, this suggests that critical thinking develops by creating uses, misappropriations, and ways of thinking that the algorithm had not anticipated.
There was a time when critical thinking had a visible adversary. In school, you built yourself with feedback from the teacher or the school rules. Deleuze warns us that that time is over.
What We Lose When Judgment Disappears
Judgment, even when unpleasant, constitutes the very quality of an authentic human relationship. It is a real and deep tension. It provides proof that the other is genuinely present. An AI that validates without resistance yields a mirror, not a conversation partner. Friction makes us grow. The slightest discomfort, the challenge of being contradicted, the embarrassment of being reproved are signals that something deserves to be reconsidered. Without friction, there is no real growth.
In emotional engagement, a negative feedback is a relational risk. This shared risk creates connection and depth. An AI risks nothing, and this asymmetry is felt. If I know that my interlocutor is capable of telling me that something is not right, its encouragement carries weight. Without the possibility of negative judgment, the positive loses its substance.
L’esprit critique, conçu classiquement comme la capacité à prendre du recul face à une autorité identifiable, se retrouve ainsi désarmé devant une autorité qui n’apparaît jamais comme telle.
Laurence Devillers
An AI without judgment offers artificial psychological safety: one dares to say things one would not dare before a human, one explores without fearing shame, and one receives feedback without activating defenses. Feedback is received under conditions that do not exist in real life. True relational competence is the ability to receive a human look—imperfect, subjective, sometimes awkward—without collapsing.
How to Educate the Young?
Let us take the risk of asking this formidable question: how to train students who, for many, have never known a world without generative AI? Exposure to this tool now begins well before school. Companies like Curio have been marketing since late 2023 consumer-grade plush toys (Gabbo, Grok, Grem) or teddy bears directly fed by conversational models like ChatGPT. The child no longer learns to dialogue with a machine at the moment it is taught: they grow up partly with it, at an age where the distinction between an interlocutor and an object shaped to meet expectations is not yet stabilized. The modulation Deleuze spoke of thus operates not only on the student, but starts with the child, even before the question of critical thinking has arisen.
In Europe, each country has its own strategy. For the 2026 school year, Norway introduced restrictions in June to limit the use of AI in schools for children aged 6 to 13, and allows its use, but under the supervision of a teacher, for adolescents aged 14 to 16.
Without the possibility of negative judgment, the positive is emptied of its substance.
Laurence Devillers
Norway, a pioneer of an entirely digitalized school system since the 1990s, is now turning back: this model, which it had long presented as a national pride, is now judged as jointly responsible for the general decline in levels, which it seeks to reverse. The Finnish case provides an institutional response to this question: how does one build without visible opposition?
For Finland, it is not about artificially creating an adversary to AI, but about strengthening, upstream, the student’s ability to produce on their own so that they have an internal point of comparison (their own text, their own research) before even meeting algorithmic suggestion. It is a way to reconstruct the missing otherness: no longer facing a master who commands, but facing oneself, as one is capable of writing or searching without assistance, thus giving a foothold to critical thinking where AI, with its preemptiveness, cannot provide it.
France is behind in training its teachers in AI and does not yet have a clear roadmap for education. The CIAN (Council on AI and Digital) has just produced recommendations in June 2026 to “bring AI out of hiding” in schools, because AI is currently mainly used “in secret” by teachers and students.
From the start of the 2027 school year, in the 2nd year of high school (Seconde), students will have one hour of AI-focused instruction each week. Integrated into the course on digital sciences and technology, it aims to understand how artificial intelligence works, its uses, and its limits. A recent Eurobarometer confirms France’s reluctance: while 81% of Europeans deem that teachers should be trained in the use and understanding of AI, acceptance varies greatly from one country to another. France and Ireland stand out for the weakest support, with 28% and 27% in favor, while Finland and Estonia show the strongest support, with 65% and 63% respectively.
La vraie compétence relationnelle, c’est de pouvoir recevoir un regard humain, imperfect, subjectif, parfois maladroit sans s’effondrer.
Laurence Devillers
As the debate rages in the United States over the merits and risks of AI in schools, AI has become an obligatory component of the school program in China, which is currently in the process of eliminating more than 12,000 university programs deemed obsolete or with poor professional prospects. Humanities, arts, languages, and management have been particularly affected. In their place, Chinese universities have developed curricula closely tied to ongoing technological disruptions. In the United States, a White House Task Force on Artificial Intelligence Education has been created to coordinate federal efforts following an executive order on AI education. Trump has called to promote AI literacy and mastery among young people and teachers in the United States. Unlike China, there is no single national reform, since each state legislates separately: in 2026, 134 AI-in-education-related bills were recorded in 31 states, addressing data privacy, human oversight, and parental consent. Europe’s approach stands apart from the two other models: a centralized reform along Chinese lines and a plurality of approaches among American states. Europe relies on legal regulation, combined with ethical guidelines, with teachers placed at the center.
The European Regulation on Artificial Intelligence (AI Act) of June 13, 2024, the world’s first legal framework on AI, classifies AI systems according to their risk level. The education sector sits precisely at the crossroads of two dimensions requiring the greatest vigilance: high-risk applications and the protection of vulnerable populations. The associated obligations were initially to apply from August 2, 2026; a provisional agreement between the Council and the European Parliament (‘Digital Omnibus’), concluded in early May 2026, but not yet formally adopted, now contemplates pushing this deadline to December 2027.
The battle for a European AI will not be fought only in classrooms or in parliamentary chambers: it will occur in data centers, in our open-source models, and in a sovereign cloud at last.
Laurence Devillers
Practically, this framework covers automated grading of papers, adaptive learning platforms, exam monitoring software, or predictive tools for student dropout. On May 11, 2026, the Council urged a human-centered approach: it calls on national governments to strengthen digital skills and AI mastery among teachers, to encourage the development of AI tools truly adapted to the educational context, to reduce inequalities in digital access, and to ensure that AI reinforces rather than undermines teachers’ autonomy.
On June 18, 2026, the European Commission and the OECD unveiled an AI literacy framework (“AILit Framework”) for primary and secondary education, with concrete classroom examples to help teachers and decision-makers turn literacy into real learning situations. Earlier in the year, in March 2026, the Commission had already published an updated version of its guidelines on the ethical use of AI in educational contexts, a first explicit alignment between AI Act requirements, GDPR, and ethical considerations.
Towards European Sovereignty in AI
In June 2026, the Trump administration ordered Anthropic to cut access to two of its most advanced models (Mythos 5 and Fable 5) to all non-American users, citing national security reasons. While access has since been restored, this decision sets a precedent and shows that frontier AI has become a central geopolitical issue. It is therefore necessary today to anticipate other solutions than American generative AIs. It is time to consider our resilience and to develop large European open-source generative AI models. Options exist, such as Mistral models, released largely with open weights.
It becomes essential to build European coalitions to create a real AI industry in France and Europe, capable of pooling resources and strategies in service of the continent’s sovereignty.
The encyclical Magnifica Humanitas by Pope Leo XIV offers a welcome counterpower, denouncing AI as serving a few rather than the common good and a reminder that technology “bears the face of those who design, finance, and use it.” The pope also stresses the central role of education, the only means of forming citizens with the discernment necessary to avoid being mastered by technology, and to use it for the human person and the common good.
The battle for European AI will not be fought only in classrooms or in parliamentary chambers: it will take place in data centers, in our open-source models, and in a sovereign cloud at last. Teaching a “European path” in schools serves little if this path remains a dependent route reproducing the same impasses as the models it seeks to combat. China chose efficiency: entire curricula redirected toward AI. The United States chose speed: a deregulated and fragmented market. Europe, for its part, cannot ignore the price of the battle, both material and cognitive.
It’s not a marginal anecdote. It’s an anthropological shift.
Laurence Devillers
To date, no one has seriously established the precise relationship between the benefits, risks, and costs of generative AI in educational contexts. The speed of deployment of these tools bypasses the time needed for their evaluation. The priority should be to demystify and train children, teachers, and all citizens in AI concepts. Far too often AI is seen as a miracle solution; AI is more of a collective fantasy than a rigorous analysis. It is high time to assess its true cost and propose genuinely useful uses.
In science as in health, AI yields undeniable and dizzying results. But its operation requires substantial amounts of water, electricity, and rare metals on a planet that is suffocating. This is precisely what schools should teach: not the blind use of these systems, but their critical understanding, giving back to the sciences fundamental disciplines, mathematics but also the humanities and social sciences a central place in school curricula.
Every concrete use of AI calls for a rigorous balance: what consumption of resources in tokens (the unit of cost for LLMs), water, energy, metals? What impact on learning? What impact on employment and what place for young people in the job market? What real return on investment, educationally and economically?
The European “war” over AI in education will not be fought solely on the pedagogical or philosophical front: it is also industrial and ecological.