Artificial intelligences are accused of understanding nothing, of getting wrong questions a child can solve, and of consuming enormous resources to succeed when they do succeed. I propose to look at things differently: these machines fail where we succeed because they do not think along the same paths as we do, and it is along those paths that they find things we would not have found.
Counting the r’s in “strawberry” and Alice’s sisters
In 2024, a joke went around social media: ask a conversational assistant how many letter r’s the English word “strawberry” contains, it answers two, and it insists. The same year, a team from the Jülich Supercomputing Centre, the University of Bristol and the LAION laboratory, around Marianna Nezhurina and Jenia Jitsev, tested the best models of the moment on a primary-school problem: “Alice has N brothers and M sisters. How many sisters does Alice’s brother have?” The answer is M plus one, since Alice is herself her brother’s sister. Most models got it wrong, and what struck the researchers, even more than the error, was the assurance with which the machines justified it with reasoning made up out of whole cloth. François Chollet, one of the designers of the Keras software library, built in 2019 a series of small visual puzzles, the ARC (Abstraction and Reasoning Corpus), which most humans solve in a few seconds, by guessing the rule that transforms one grid of colours into another. For four years, language models scored close to zero on it.
Yann LeCun, a French researcher based in the United States, winner of the 2018 Turing Award, one of the inventors of the deep learning from which today’s models derive, and for a long time head of artificial intelligence research at Meta, sees in this type of failure the confirmation of a diagnosis he has been repeating for years and which he summed up in the Wall Street Journal in October 2024: these machines are dumber than a cat. A cat has a model of the physical world, a memory that lasts, a capacity to reason and to plan; a language model, he says, handles language like a virtuoso without that virtuosity proving anything, since one can manipulate language without being intelligent. I discussed this diagnosis in the article Ensourced Speech, where I take it seriously, because it is right on one point: these machines reason about sentences that speak of the world, not about the world. LeCun left Meta at the end of 2025 to build architectures that would learn the world through video rather than text, and his bet is a legitimate one.
This diagnosis also has something reassuring about it. To say that the machine is dumber than a cat is to tell ourselves that we are far more intelligent than it is, that the hierarchy is safe, and that these machines are, at bottom, just one more tool in our hands. I believe this reading misses what is happening to us, and that in the machine’s failures, if we take them seriously, we can read something other than its weakness.
Benjamin Beck’s gibbon and the string lying on the ground
The Dutch ethologist Frans de Waal published in 2016 a book entitled Are We Smart Enough to Know How Smart Animals Are? (W. W. Norton), and in it he recounts an experiment from the 1960s. Gibbons failed a test that chimpanzees and many monkeys pass without hesitating, pulling a string lying on the ground to bring in food that is out of reach. The conclusion was the cognitive mediocrity of gibbons, which was strange for primates so close to us. The American primatologist Benjamin Beck looked at their hands. Gibbons live in trees, move by hanging from their arms, and their hands are hooks, with a tiny thumb and elongated fingers, made for grasping branches and incapable of picking up an object from a flat surface. The ground is hardly part of their world. Beck raised the string to shoulder height, and the gibbons passed the test at once. De Waal takes from Jakob von Uexküll the word Umwelt, the world proper to each species, carved out by its organs and its ways of living, and he draws from it a rule of method: when an animal fails a test, one must first ask whether the test was designed for the animal’s world or for ours.
The letter r in “strawberry” is a string lying on the ground. A language model does not read letters, it reads units called tokens, fragments of words cut up according to their frequency during training, and the word “strawberry” reaches it in two or three pieces whose spelling it never sees. Asking it to count letters is asking a gibbon to pick up an object from the ground. The Alice problem comes down to something else, to the fact that the machine has no body and therefore no family: “sister” is for it a word connected to other words, never a person one has felt oneself to be the sister of. As for Chollet’s puzzles, where a rule must be guessed from two or three examples, they call on what we have been doing since birth by looking at the world, and what no one learns from a text.
The roboticist Hans Moravec had described in 1988, in Mind Children, the paradox that bears his name: it is easy to give a computer the level of an adult at intelligence tests or at checkers, and almost impossible to give it the abilities of a one-year-old in perception and motor skills. His explanation lies in the history of evolution. Recognising a face, grasping an object, understanding that a sister has a brother, these are skills that were inscribed, over a billion years of selection, in the largest parts of our brain, and that seem easy to us because they have become unconscious to us. Abstract reasoning, Moravec writes, is the thinnest veneer of human thought, a recent invention, which we find difficult because we have not yet mastered it. What we call common sense is the slowest and most costly fruit of evolution, and a machine that has no body inherits none of it, it inherits only text. Its stupidity is therefore first of all an artefact of measurement, since we judge it on tasks that are our own ground.
As early as 1947, when the first computers were barely out of the laboratories, Alan Turing argued in a lecture to the London Mathematical Society, on 20 February, that a machine required never to make mistakes cannot be intelligent. A machine that never makes mistakes, he explained, merely carries out what it has been asked to do. For it one day to propose something no one has asked of it, one must accept that it can also produce wrong answers: in a machine, the capacity to invent and the possibility of being wrong go together.
“That’s not how people play chess”
There remains the second part of the reproach, the most serious one to my mind: these machines consume enormously for childish results. Here too the fact is accurate. In December 2024, the o3 model from OpenAI solved 87.5% of Chollet’s puzzles in a configuration where, according to estimates by the ARC Prize foundation, each puzzle cost several thousand dollars of computation, whereas Chollet valued human work at five dollars per puzzle and the energy spent by a brain at a few cents. A year later, the cost had fallen to a dozen dollars per puzzle for a comparable score, and the best systems were above 90% on this first series; but in the spring of 2026, on the second series of puzzles, designed to remain easy for humans, the best systems reached 69%, and 13% on the third, where humans succeed at almost everything. This cost is real, and I described in the article Ecology and Artificial Intelligence what it concretely amounts to in electricity and water. It seems to me above all that this cost lies in the very nature of this cognition.
Richard Sutton, one of the founders of reinforcement learning, published in March 2019 a short text, “The Bitter Lesson”. From seventy years of research in artificial intelligence one can draw a single lesson, he writes: general methods that exploit computing power always win, and by far, over methods that try to put our own knowledge of the domain into the machine. He takes the example of chess. When Deep Blue beat Kasparov in 1997 through massive search of moves, the majority of computer chess researchers, who had spent their lives coding human understanding of the game, took it badly; they said that brute force had won this time, that it was not a general method, and that in any case that was not how people played chess. The lesson is bitter, Sutton writes, because it contradicts the approach the researchers preferred, the one that started from human knowledge. I would put it this way: we would like the machine to win by resembling us, and it wins by not resembling us.
We economise on computation because we have almost none; our brain runs on some twenty watts, and it makes up for this poverty with an immense store of knowledge, acquired through evolution and through experience, which allows it not to explore what it knows in advance to be futile. The machine does not have this knowledge, and it has computation. It replaces understanding with exploration, it tries thousands of paths we would not even have considered, and it keeps the ones that lead somewhere. What is, seen from our side, an insane expenditure for a result a child obtains in one second is, on the machine’s side, the only way to get to the end of a problem when you carry no world within you. Consumption and stupidity are two sides of the same fact: this cognition does not take our shortcuts, because it has no ground on which to take them.
Four grandmasters learn concepts no one had formulated
Now it is by not taking our shortcuts that it finds. I recounted in the article Biological intelligence, electronic intelligence move 37 of AlphaGo against Lee Sedol in 2016, that move which commentators first took for a mistake and which no human game contained. I would like to add here three more recent results, chosen because they show the same movement at three different scales, the game, mathematics, and the intimate workings of the model:
- The first comes from the game of chess. In March 2025, Lisa Schut, of the University of Oxford, and a team from Google DeepMind, including Demis Hassabis, published in the Proceedings of the National Academy of Sciences a study entitled “Bridging the Human-AI Knowledge Gap”. AlphaZero, the program that learned chess by playing against itself without any human game, represents the game in a mathematical space of very high dimension. These researchers searched it, without knowing in advance what they would find there, for concepts the program uses and which have no name in human chess theory. They extracted several, verified that they were teachable, then taught them, in the form of positions to solve, to four grandmasters, Vladimir Kramnik, Dommaraju Gukesh, Hou Yifan and Maxime Vachier-Lagrave. All four improved measurably on these positions. The authors’ conclusion is cautious: AlphaZero encodes knowledge that exceeds current human knowledge, but that is not beyond the reach of the human mind. Players among the strongest in the world learned something about chess from a machine that has never read a chess book, and what they learned existed in no book.
- The second comes from mathematics. In 1969, the German mathematician Volker Strassen had shown that two four-by-four matrices could be multiplied in 49 multiplications instead of the 64 of the classical method. For fifty-six years, no one did better in this setting. In May 2025, AlphaEvolve, a DeepMind system that evolves programs with the support of a language model, found a method in 48 multiplications for matrices with complex coefficients. One multiplication fewer seems derisory; but matrix multiplication is the basic operation of the computations of artificial intelligences themselves, and above all, this result came out of a system that mutates programs by the thousands and keeps only those that do better, a way of searching that is not a mathematician’s.
- The third comes from inside the model. I described in the article The machine that writes to itself the interpretability work by Anthropic published in March 2025: when you observe the circuits that activate inside Claude while it adds two numbers, you see several parallel paths of calculation, none of which resembles the column addition we learned at school, while the model, when asked, describes the school method. The machine learned from our texts how we recount an addition, and it adds otherwise. This gap is the trace of a cognition that found its own paths to a correct result, without having any means of knowing them itself.
Reasoning drafts no one can read
Since late 2024, language models have been “reasoning” before they answer, that is, they write a draft of intermediate steps, and they have been trained by reward so that this draft leads them more often to a correct answer. I retraced this history in the same article, “The machine that writes to itself”. This change is real. The Alice problem, which the models of 2024 failed, is solved far more often by reasoning models; Chollet’s puzzles, which they did not know how to approach, are solved in the majority. Writing steps means giving oneself computing time and a working memory, and that is enough to cross thresholds.
What interests me is that this reasoning, which resembles ours so much since it is written in human language, remains foreign to ours. Arun Jose presented at the NeurIPS conference in December 2025 a study on the legibility of the drafts of fourteen reasoning models. Several of them, including the open models R1 from DeepSeek and QwQ, produce in the course of reasoning illegible passages, sentences without head or tail, random words, characters from other alphabets, before returning to a perfectly coherent final answer. These passages serve a purpose, since when the model is forced to rely only on the legible parts of its draft, its rate of correct answers drops by more than half. The larger the model, within the same training regime, the more illegible its reasoning becomes. The author states it plainly: the model learns to reason in a way that is foreign to us. What is rewarded in training is what works, and what works has no reason to resemble what we would know how to read.
A team at Apple, around Parshin Shojaee, made a great stir in June 2025 with a study entitled “The Illusion of Thinking”, in which they showed that these reasoning models collapse beyond a certain level of complexity on classic puzzles, the towers of Hanoi for example, and that they “overthink” simple problems, continuing to explore after finding the answer. A reply came in the following days, from the pen of Alex Lawsen, who showed, in a text first co-signed, as a wink, with the model Claude Opus, that the collapse points coincided with the limit on the number of words the models can write, that the models moreover said so in their answers, and that some river-crossing problems posed in the study had mathematically no solution. It was, once again, the string lying on the ground. The two studies seem to me to be true, both of them: these machines think otherwise, and we keep evaluating them with tests made for us.
Daniel Andler’s crowbar
The philosopher Daniel Andler, one of the founding figures of cognitive science in France, published in 2023 with Gallimard Intelligence artificielle, intelligence humaine : la double énigme (“Artificial intelligence, human intelligence: the double enigma”). His thesis rests on one observation: however much artificial intelligence progresses, the distance separating it from its proclaimed goal, reproducing human intelligence, does not shrink. To understand this enigma, one must in his view confront a second one, that of human intelligence itself, which cannot be reduced to the capacity to solve problems. Andler calls “panproblemism” the temptation to bring everything back to problem-solving; human intelligence, he writes, is the way we face the situations we are in, whatever they may be, and an artefact does not relate to its environment as to a situation. He then proposes a move of vocabulary I find liberating: to take the expression “artificial intelligence”, that oxymoron of blurry concepts, as a new word designating a tangible thing, in the way we speak of a crowbar without wondering what it has of the crow or of the bar.
I share Andler’s observation about the distance that does not shrink, and I draw from it a consequence he perhaps does not draw himself. If this distance does not shrink while capabilities progress, it is because it does not correspond to a lag, which could be closed, but to a difference in nature. When you hold together the machine’s failures and its successes, you see a cognition that is not on its way towards ours, that is neither a weak version nor a future version of it, but another way of going from a question to an answer, without a body, without a situation, without common sense, with text and computation. I proposed in the article Trouble in Intelligence the word polyintelligence to name the cohabitation of forms of intelligence that no common scale measures, that of plants, of animals, of machines and our own.
Three ways of working with a cognition that is not ours
From this understanding I draw, for my own work and for those who, in the cultural field, wonder what to do with these machines, a few ways of doing things that I practise or that I propose, without claiming they are the only ones:
- Ask it what we would not have looked for. The value of this cognition does not lie in what it confirms for us, but in what it brings from its own side, and this contribution comes when it is given room. In the article Cultural investigation with artificial intelligence, I proposed entrusting to a machine hours of recorded conversations with the visitors of a cultural venue, without asking it any directive question, and looking at what it draws out of them. A reading done without our certainties produces connections our certainties forbade us. It is move 37, transposed to a theatre or a library.
- Take its errors as a revealer of our certainties. When the machine fails the Alice problem, it shows us that what we called reasoning was in reality a knowledge of the body, an experience of family, which we had never had to formulate. In each of the machine’s stupidities, one can thus read something we know without knowing it, and a team that takes the time to be astonished by it learns a great deal about itself. I described in Pedagogy of Artificial Intelligence how to make this discovery of the tool’s resistances the starting point of a workshop, rather than a lesson handed down in advance.
- Listen to its accidents as a material. I defended in The arte povera of artificial intelligence the strange forms produced by imperfect models, and in Artificial Intelligence and Creative Discovery the interest of a music no one would have composed. Artists have always made works out of the resistances of their tools. The paths of a cognition that does not have our shortcuts are one of those resistances, and I know of no stranger material to work with.
These three ways suppose that we give up measuring these machines by our own yardstick, on which one sees only lags and expenditures, and take them for what they are, a cognition of another type, which gets to the end of problems by paths we would not take, whose failures teach us about ourselves, and whose findings reach us without our having expected them.
A cognition beside ours, neither above nor below
The word stupidity keeps the machine below us, on that single scale Frans de Waal dismantled for animals. I believe the machine is beside us, neither above nor below, and this proximity is new in human history: for the first time, a cognition that is not ours speaks to us in our language. Gibbons never told us about the world as it appears to hands made for branches; the machine, for its part, answers us. It is in this sense that its stupidity is its strength.
