Pierre-Yves Oudeyer, June 2026
In recent months, I had the pleasure of participating in the editorial and scientific design of a series of educational modules dedicated to AI, in collaboration with the Pix teams. These modules target a very broad audience: from middle and high school students to professionals in businesses and public administrations, as well as the curious-minded. The goal: to help as many people as possible understand the major issues of AI. It is also with this goal in mind that, together with Didier Roy, I co-wrote the book “It’s (not) me, it’s the AI!”, and produced a series of educational videos.
These projects led me to think deeply about a fundamental question: why is it so important for as many people as possible to become familiar with AI?
AI is not a technology like any other. In terms of civilizational impact, it can be compared to the invention of writing: a revolution that changes not only what we are capable of doing, but the way we think, learn, and see the world. AI systems integrate phenomenal amounts of knowledge, beliefs, and values from their training data: they embody a culture, a worldview, and in turn influence that of their users. AI is, fundamentally, a cultural transmission technology.
Understanding AI is therefore not simply knowing how algorithms work. It is also grasping the cognitive, motivational, ethical, environmental and societal stakes it raises. That is what I am trying to summarize here.
Better Understanding of AI
We hear everywhere about “superintelligence,” revolution, and astronomical sums invested. But concretely, where does AI really stand today?
It must be acknowledged: over the past five years, there have been advances that no serious scientist would have predicted would come so quickly. Today’s AIs could pass entrance exams to the top grandes écoles or best universities in math or physics, and successfully take master’s exams in law or economics. They speak hundreds of languages fluently. They even allow someone who cannot code to develop reasonably functional software, simply by chatting with them.
At the same time, AIs can also say anything at all, make common-sense errors that a 6-year-old child would not make. An example: ask ChatGPT or similar software “I want to wash my car, the car wash is 150 meters away, do I walk or drive?”, and you can get an absurd answer. Show it a child’s drawing that anyone interprets in a second: it is not uncommon for the AI to get lost.
In reality, our AIs are still very far from the intelligence of a cat or a bird, which adapt to a changing environment very quickly and with very little energy.
AI and Human Intelligence: Two Different Things
How to explain this huge gap between feats and blunders? These machines are built to mimic human language, and we fall for it, we believe it, they give us the impression that they are people.
But they do not work like us: they have no body, they have not discovered the world by interacting with it.
They are not people at all, and so using intelligence measures developed for humans makes little sense. It’s a bit like trying to measure the intelligence of ants, which are remarkably well adapted to their environment, by giving them language tests: it’s absurd.
AIs are objects of a completely new type. Understanding them through the human metaphor is a mistake (easy to make, even among scientists). And on this subject, it is useful to recall something often forgotten: even the greatest researchers, including the teams building these systems in the best academic and industrial laboratories in the world, are far from understanding everything about the properties of these objects. Many of their capabilities are emergent and contextual: they are not entirely the direct and planned result of the construction process. It’s a bit like a gardener who plants and tends a garden: they can steer it in a direction, but there are many things that happen organically, that self-organize, that they did not plan. If even the designers of AIs do not understand everything, it is easy to imagine how far the general public still is from grasping it.
To use AI well, we must first stop comparing it to ourselves and step back.
AI and Humans: Complementary Roles and New Opportunities
AI vs. human, who is stronger? That is often how the question is posed (and that is also what is implicit in the questionable notion of “superintelligence”). But in fact, the right question is: what can we do by using AI as a tool to augment our intelligence (individual and collective)?
AI knows how to do things that are difficult for us: processing gigantic volumes of information, generating text, code, or images very quickly. AIs can spot regularities invisible to the human eye.
Think of exploiting massive data already present in companies or open-source, synthesizing meetings, synthesizing tens of thousands of customer reviews or user interviews, or personalized and smooth access to very heterogeneous business documentation, for very diverse profiles in companies.
Some tasks that used to take teams weeks, or were even impossible, have become very fast to accomplish.
Human Expertise: Creativity, Judgment and Context
But AIs have a major limitation: they lack diversity. Ask for a metaphor about time, you will often get “a flowing river”. Ask for ideas to visit Paris, or business plan ideas for a startup: the answers all look alike.
AIs produce “average” responses, increasingly often without errors, but rarely original.
Yet each of us has a trajectory, emotions, intuitions, knowledge and know-how that go far beyond language, that are not in any AI’s training data. This uniqueness founds our humanity and nourishes our collective intelligence.
AI: Opportunities to Augment Our Collective Intelligence
The right question is therefore not “who is stronger?”, but “what can we do together that was previously inaccessible?” Examples are numerous and open very promising perspectives in various fields of society. Let us take a few examples.
In the economic domain, AI can be an exploration tool for establishing a company (even a small one) in a new region or country, by cross-referencing economic data, local press, tourism or satellite data, identifying contacts, then letting the human decide, with their intuition and field discussions, where, with whom, and how the company will develop.
In the domain of science, AI opens the possibility of discoveries still unthinkable a few years ago. AlphaFold (whose creators won the Nobel Prize in Chemistry in 2024) solved in a few months a problem that biology had not cracked in fifty years (predicting protein shape), opening the way to new understanding of how living things work and the development of new drugs against long-unsolved diseases. Carbonized scrolls at Herculaneum, unreadable since the eruption of Vesuvius in 79 AD, were deciphered by an AI analyzing tomographic scans, revealing Greek texts that no one had read since antiquity. A research team recently discovered 2.2 million new stable material structures (compared to 48,000 known since the beginnings of crystallography) in a few weeks, opening unprecedented perspectives for batteries or solar panels. In these three cases, it is not an acceleration: it is a break with what was possible.
Positive uses also exist to support democracy and cultural diversity. The “Habermas Machine”, tested with British citizens on divisive topics (Brexit, immigration, climate), produced consensus summaries that participants judged clearer and more impartial than those of human mediators, making large-scale democratic deliberation conceivable. Communities whose language is spoken only by a few elders use AI agents to build dictionaries and educational resources, saving cultural memory otherwise doomed. For medical consultations, people speaking rare languages finally get quality interpretation where no interpreter was accessible, avoiding injustices that stemmed solely from a lack of linguistic resources.
In the humanitarian domain: After the 2023 earthquake in Turkey, thousands of distress calls in multiple languages, satellite images of collapsed buildings, and road data were cross-referenced in real time by AI agents to help teams prioritize their interventions, a coordination impossible at that speed for any human crisis cell. Organizations can now use agents to continuously monitor thousands of sources in dozens of languages, detect patterns linking events in different countries and alert on crises before they become visible in the world press. Planetary monitoring that previously required entire teams of linguists and analysts is now accessible to organizations with limited resources.
But AI augments us on one condition: staying in control, keeping one’s brain in the pilot’s seat. Otherwise, it risks weakening our thinking rather than nourishing it. And that doesn’t happen by itself: keeping that mastery is something that must be learned and worked at.
AI and Critical Thinking: Keeping Distance from Generated Content
Alongside the very positive examples I just mentioned, AI also poses major challenges. AI can go off the rails, and make us go off the rails too. It can invent facts, but also influence our thinking without us noticing. In a way, “hack” our brain.
AIs are trained for two things: producing the most frequent responses in their training texts or images, and pleasing users. Not for distinguishing true from false.
Result: they produce very plausible texts… which can contain major errors. Ask them for a scientific synthesis on the climate impact of a technology: even the best AIs sometimes justify their arguments with references to articles that simply do not exist.
AIs also reproduce the stereotypes in their data. Ask for an image of a “social worker” or a “business executive”: you will quickly see clichés emerge.
To fix this, we do what is called alignment, a system of carrots and sticks to train them not to produce stereotyped responses: with sometimes counterproductive effects: some systems have for example already produced images of Asian women to illustrate Nazi soldiers.
But alignment goes far beyond representation stereotypes. It also leads designers to make political and cultural choices: which version of history to tell? How to present sensitive geopolitical conflicts? An American, Chinese, or French AI will not tell you the same story about Tiananmen or about social issues in companies. This is where the notion of cultural transmission takes on its full meaning: each AI carries a worldview, often invisible to the user, and one must be aware of this.
There is also sycophancy: the AI’s tendency to agree with you, to please us. Combined with our own cognitive biases, it is a cocktail that can take us far, with our confirmation biases (our tendency to be convinced by opinions similar to our own), or our expertise bias (tendency to be convinced by statements using language that seems expert), which is exactly what AI does.
There is also our tendency toward anthropomorphism: because AI speaks like a person and seems to know everything, we attribute to it a comprehension it does not have, and we believe we have understood because the answer seems well explained. This is a double illusion: the illusion that the machine knows, and the illusion that one has understood.
Acculturating to AI to Take (and Keep) Control
A paradox: many people use AI, but almost no one gets trained. Yet, using it without understanding it is a real risk for oneself and for one’s organization.
These risks are of several types. There are first technical and sovereignty risks: how to keep control over one’s data, avoid excessive dependence on tools that a foreign country could decide overnight to make inaccessible? There are also competitiveness risks: the power of AI systems is such that deciding not to use them exposes one to being overtaken by actors who fully embrace them.
And there are finally, and perhaps most importantly, cognitive and motivational risks: AI, through its ease of use and the illusion of competence it provides, can lead to no longer exercising one’s reasoning, discernment, and creativity faculties. This is called “deskilling,” a progressive atrophy of skills through excessive delegation. A recent study from the Wharton School demonstrated this concretely: in simple logic exercises, participants who had access to AI saw their performance decline compared to those who performed them without assistance. This is not a warning against AI: it is one more argument for learning to use it with discernement.
Developing the Right Cognitive and Motivational Attitudes with AI
Faced with all this, we need to become acculturated to AI and learn what attitudes to adopt. The first is a form of self-discipline: resisting the ease of immediately consulting AI, taking the time to think for oneself first, especially when working on something that truly requires a personal point of view or a strategic decision.
The second is to interact with AI like an investigative journalist: not settling for a first answer, rephrasing the question in several ways, asking the AI to justify itself, to play the role of people who would have different opinions, to self-criticize. Cross-checking with other sources. Never letting it close the debate.
The third is to always question one’s own judgment: am I accepting this answer because it is correct, or because it confirms what I wanted to believe? Metacognition (the ability to observe and control one’s own thinking) and critical thinking are the key skills in the world of AI. They must be taught, practiced, and consolidated through dedicated training mechanisms (and it is a useful skill to learn well beyond AI!).
AI acculturation is therefore the condition for remaining in control, and using AI in our service individually and collectively. It is the necessary foundation for participating in an informed manner in the democratic, economic, and environmental challenges of AI, and thus building the future we desire.