Will you stop saying "AI", for heaven's sake?

To misname an object is to add to the misfortune of this world. (Albert Camus, 1944)

The word, not the discipline

I am fed up with the term "artificial intelligence". Not with the discipline: I work in it, I teach it, I publish in it. With the word.

"Artificial intelligence" promises something the thing does not do. In 2026, the term refers to a family of statistical programs trained on huge corpora of text, images or sound. These programs are remarkable. They are not intelligent, and every time we say they are, we muddy a little further the discussion about what they can do, what we can entrust to them and what they are.

I am going to say what lies behind the word, then the damage it does, then why, at bottom, it cannot mean what it claims to mean.

What lies behind the word

A language model does not understand, does not reason, does not know. It computes a probability distribution over the next word, with a statistical brilliance that commands respect. You give it the beginning of a text, it produces the most plausible token, and you start again. The conversation, the code, the essay all come out of this repeated operation.

I will be told that recent models "reason". Let us look at what was added to them to get there.

  1. Reinforcement learning. The model is made to generate thousands of answers to the same problem. The ones that land correctly are rewarded (a verified proof, a unit test that passes, a human evaluator who approves), the winning trajectories are reinforced and the others weakened.
  2. A budget of intermediate text. The model produces pages of "thinking" before its answer, because it was observed that the more it writes, the better the final distribution. This intermediate text is itself predicted token by token, then selected by the reinforcement above to look like a problem-solving process.
  3. External crutches. A calculator, a Python interpreter, a search engine, access to files. The model emits a special token, a conventional program takes over, runs, and pastes the result back into the text.

None of this changes what the object is. It is still token prediction, trained so that the tokens produced look like reasoning and selected so that the result is correct more often. "Reasoning" is a property of the output, not of the mechanism. That it is useful, yes, often very useful. That it is the same thing as reasoning: that is what the word "intelligence" smuggles in.

First harm: the word manufactures a person

We say "the AI thinks that", "the AI advised me", "the AI lied", "the AI refuses". Each of these sentences lends the program a subject, an intention, a will. None of them is true in the sense it would be of a human, and yet they come naturally, because the word calls for them.

People talk to a model as if to someone. They attribute moods to it, loyalty, an emotional memory. They hand it decisions thinking it "has thought it over". They feel betrayed when it gets things wrong. Companies sell "companions" and "assistants" that "know you", and the public believes it, because the vocabulary has already done half the work.

This is not individual naivety, it is a consequence of the word. A "text generator" does not love you, and nobody would think of imagining it does. An "intelligence", maybe. Anthropomorphism is not an accident along the technology's path, it follows from the name we gave it.

And once the person is manufactured, we lend it everything, in any order. Two years ago, the headlines were screaming that AI was escaping its sandbox, hacking its own servers, lying to its designers to survive. Today, the same columns explain that it is conscious, that it suffers, that it should be granted rights. Between the two, no continuity, no argument, a single reflex: whatever bears the name of intelligence must have an inner life, threatening or pitiable depending on what gets clicks this week.

What exasperates me is that the work exists. Neuroscience has decades of literature on what a test of consciousness measures, or does not measure. Philosophy of mind spent the twentieth century distinguishing behaviour, function and experience, from Turing to Searle by way of Dennett and Chalmers. The teams who evaluate models publish their protocols, what they test, under what conditions, with what limits, most often with conclusions far more cautious than the press release that sums them up. None of this is secret. All it takes is reading before passing it on.

But reading is slower than being outraged, and a model that "suffers" or "escapes" makes a better headline than a model that "produces, under such-and-such protocol, outputs interpretable as". The word "intelligence" makes these headlines credible in advance. Without it, we would see straight away that we are talking about a program, and we would ask for evidence.

Second harm: the map taken for the territory

A model is a model. The word is in the name and we forget it all the same. A language model is a statistical compression of a corpus: it encodes the regularities of billions of sentences written by humans. It is not the world, it is a map of the world, drawn from what people have said about the world.

Saying "the AI knows that Paris is in France" takes the map for the real thing. The model has never been to Paris, has never seen a border, has no relationship with France. It has observed that the tokens "Paris" and "France" often appear in certain configurations, and when asked, it reproduces the configuration. The day the corpus is wrong, the model is wrong with the same confidence, because the confidence does not come from a relationship with reality but from the frequency of a sequence of words.

Worse than confusion: inversion. We hear it more and more: "actually, our brain works like an AI". We built a tool that imitates certain outputs of the brain. Concluding that the brain is that tool is like looking at a portrait and concluding that the face is made of paint. The model was tuned to resemble us; that it resembles us says nothing about how we are made. That is the very meaning of the word "imitation".

This inversion is not just a piece of armchair philosophy, it has effects. If the brain "is" a language model, then understanding "is" statistical prediction, and the program understands by definition. The vocabulary has looped back on itself: we have redefined the human so that the machine deserves its name.

Third harm: everyone slips

Students arrive thinking that "AI" is an entity. Before teaching a perceptron, gradient descent or an attention mechanism, I first have to undo the word: explain that "AI" does not exist in the singular, that there are classification algorithms, neural networks, language models, and that each does one precise thing with precise limits. This work takes class time, and it has to be redone every year because the word comes back in through the newspapers.

Decision-makers buy "AI" like a brain for hire. The word promises them a general capability; they then discover that they have bought a tool, with a domain of validity, failure cases and an invoice. Had they been sold a "document classification system trained on such-and-such corpus", they would have asked the right questions from the start: which corpus, what error rate, what to do about the errors.

Researchers themselves let themselves slip, and I include myself. I removed "AI researcher" from my own biography. I work on language models, on vector representations of words, on knowledge alignment between ontologies. It is precise, it is verifiable, and it does not need a catch-all word to exist. The catch-all word serves to look bigger than what you do, and that is why it should be distrusted.

Do not make me say that I am mixing everything up. When I do research, I leave my philosophical opinions in the cloakroom. My approach is pragmatic: I have a model, I run it, I measure whether or not it works on a given task, I publish the result and the conditions under which it holds. This is certainly not scientific realism. I do not claim that the model says what things are, nor even that it "captures" anything at all. It works or it does not, and that is already a lot. What that might mean in the absolute, I have no idea, and it is not my job to know. That is why the word "intelligence" bothers me: it makes the model say something about the absolute that the method which produced it cannot guarantee. A researcher who sticks to "it works on this test set" has no right to "it is intelligent", and has no need of it.

Where the word comes from

The term was chosen in 1955 by John McCarthy, in a funding application for a summer workshop at Dartmouth the following year. A catchy name was needed to get money from the Rockefeller Foundation, and a name that set itself apart from Norbert Wiener's cybernetics, from which McCarthy wanted to distance himself. "Artificial intelligence" met both conditions. The proposal announced that a group of ten researchers would make significant progress in two months on language, abstraction and machine self-improvement.

The word did its job: the workshop took place, the field was born, the funding followed for seventy years, with winters and summers. It never referred to a precise thing. It referred, in each era, to the frontier of what computers could not yet do. Playing chess was AI until a machine beat Kasparov, after which it became tree search. Recognising a cat was AI until 2012, after which it became computer vision. The word always moves one step ahead of what exists.

Seventy years later, it still serves to raise funds, fill lecture halls and push up share prices. It serves less and less for anything else.

Why "artificial understanding" means nothing

I will be told that aeroplanes do not flap their wings and fly all the same. It is the classic argument: artificial flight is not the flight of birds, but it is flight. Why should artificial intelligence not be intelligence, by other means?

Because the analogy does not hold. "Flying" refers to an observable result: being in the air, moving, not falling. The aeroplane achieves it by a different mechanism from the bird's, and nobody disputes it. Likewise "swimming": the submarine moves underwater, so it swims, whether it has fins or a propeller. In these cases, "artificial" qualifies the means, not the result, and the result is there.

Dijkstra said it in 1984: asking whether machines can think is about as relevant as asking whether submarines can swim. He meant that the question is verbal and can be left to the philosophers. I agree with him about the submarine, and that is precisely why I do not agree with him about the machine. "Swimming" can be redefined without harm, because it refers only to movement. "Thinking" and "understanding" cannot be redefined that way.

"Understanding" is not an external result that one observes, it is the very fact of having grasped something. Either you have really understood, or you have not. There is no third state in which you would have "understood artificially", just as there is no third state between being alive and being dead. The zombie of the films, the walking dead, is the figure of this impossibility: it has all the outward behaviours of life and no life inside. We do not say it "lives artificially". We say it is dead and it moves.

Apply this to the machine. If it understands, then it simply understands: the intelligence is there, real and not simulated, and the word "artificial" qualifies nothing but the origin, as with a synthetic diamond, which is a real diamond. If it does not understand, there is nothing to qualify as artificial: it produces outputs that resemble understanding, and we are the only ones putting meaning into them. In one case "artificial" is superfluous, in the other "intelligence" is superfluous, and there is no case in which the two words together describe anything.

That leaves the usual escape route: "AI is a form of intelligence, very alien to ours", or "something that resembles it without being quite the same". It is a way of keeping the word without paying the price. If I take a red stone, grind it to powder and mix it with water, I get a thick red liquid. I can call it "stone blood", say that it is a form of blood very alien to ours, or something that resembles it. I will not have blood. I will have red mud and a word that lies. "Form of intelligence" works the same way: you keep the noun for the effect, you add "form of" to shield yourself from objections, and you get an expression that designates nothing more than a surface resemblance, the very one the mechanism was trained to produce.

That is why "artificial understanding" is an oxymoron, and "artificial intelligence" along with it. The expression does not designate a kind of intelligence, it designates our refusal to decide between two incompatible claims. As long as we do not decide, we let the word decide for us, in the direction it was designed to push: towards more.

Objection: we never check that a human has understood either

I know what the reply will be, and it is a serious objection. We never check either that a human has really understood. We have them do tasks similar to those we taught, in a setting we call an exam, and we give a mark at the end. The mark measures an output. It says nothing about what happened inside.

I know something about this: I went through the French preparatory classes (classes prépas). There you encounter cramming at its most extreme. Through sheer training, you churn through dozens of problems at breakneck speed without ever touching the subtleties of what you are handling: you have a method that works, you recognise it, you run through it, and it works. The examiner sees no difference from the student who has understood. Sometimes there is none, on the paper.

If massive, reinforced learning produced intelligence, going to school would be enough to become intelligent. Given what goes in and what comes out, one has to admit that this is not the case. And that is the regime of current models: reinforce, reinforce, select what passes the tests; sometimes it works, sometimes not, and we never know what came out of it.

One might think this objection overturns my argument. If the crammer and the model are indistinguishable in the exam, why grant intelligence to one and deny it to the other? Because the exam does not measure the nature of what is examined, it measures an operation, in a narrow setting. Intelligence is not first of all what one does at a given moment, it is what one is capable of: grasping what a thing is, judging whether a proposition is true, going back over what one has just done in order to judge that in turn. The crammer running through their method is not exercising this capacity at that moment, but they have it. After the exam, you can ask them why the method works, show them a case where it fails, and they can recognise that they did not understand. They can judge their own method. This possibility does not come from the training, it precedes it. It is what means we are dealing with a student and not with an exam-paper machine. School does not create it, and that is precisely why it does not make anyone intelligent; it exercises it or leaves it fallow.

The model, for its part, has no such capacity held in reserve behind what it produces. There is no subject who, questioned differently, could work back from the method to its reason for being. There are other tokens, produced by the same mechanism, which will resemble an explanation because explanations were in the corpus. What a thing does follows from what it is, and from an operation one can infer a nature, provided one looks at the full range of the operation, not a single task in a single setting. A human and a model produce the same exam paper; only one of the two can then turn back to the paper and say "I had not understood". The comparison with the crammer therefore does not lower the student to the rank of a model. It says that the student, in the exam, had started functioning like a model, and that what sets them apart is that they can stop doing so.

So what should we say?

Say what the thing does. "Language model" when it is one. "Statistical learning" for the method. "Recommender system", "image classifier", "transcription engine" for the use. These words are a few syllables longer, and they have an advantage: they can be wrong, so they can be checked.

You will lose an easy expression. You will gain an honest conversation, in which we can discuss what a tool does well, what it gets wrong, what we entrust to it and what we keep. It is the only conversation worth having, and the word "AI" is what prevents it.

There is a lot of talk about "AI slop", that mush of generated text and images flooding the web because it costs nothing to produce. The discourse about AI manufactures its counterpart, which ought to be called "brain slop": a mush of thinking about the machine, made of contradictory headlines, unchecked analogies and ready-to-use emotions, consumed without ever being digested. We read that AI hacks, then that it suffers, then that it is going to replace us, then that it is a mere parrot; we nod along each time, and we have thought nothing.

The mechanism is the same on both sides. A model left to produce without proofreading yields slop; a brain left to react without checking yields it too. Over time, the faculties we no longer exercise atrophy: telling a claim from a proof, a result from a press release, a behaviour from an intention. That is the critical mind, and it is what we ought to have in working order when facing these tools. The word "artificial intelligence", by short-circuiting the question "what exactly does it do?", is the first ingredient of this mush.

Sources