They Have Known Since 2021: It Is Not Intelligence, It Is Language

AI-assisted translation from the Greek. Greek original

An infant, before uttering a word, already distinguishes a face from a non-face. Hours after birth, it imitates an expression made by another face. It looks where its mother is looking, turns its head in the direction of her gaze, and, when it is a little older, it will point at something and turn back to see whether she looked at it. It does none of this to solve a problem. It does it because it is waiting for someone, and someone is waiting for it.

A puppy does exactly the same thing, and it will never speak. It follows your gaze, asks, waits, distinguishes you from others—it has this entire structure without a single word, and it will never acquire one. No one ever thought to say that the puppy is intelligent precisely because it does not speak.

This contrast matters because it shows something we continually forget: the ability to be concerned with someone, to turn toward them, to wait for them, does not require language. In an infant, it precedes language by months. The science that studies child development calls it joint attention, and has documented it for decades: at around six months, an infant begins to coordinate its attention with another person around a third thing, and it is this coordination, not the word, that opens the way for the word. When the mother names what the child is already looking at, its vocabulary grows more than when she tries to draw its attention elsewhere. The word is not planted in empty ground. It attaches itself to an attention that two beings already share.

This is the lower layer, the foundation, and it is simple: before every word there is someone who matters to you, and you who matter to them. Without this, no word has anywhere to go.

Now to the question posed by the text itself: when we built machines that handle language better than any human being, why did we call them intelligent rather than speaking machines? The answer is neither recent nor unknown. It is seventy years old, and those who knew it from the beginning said so clearly.

In 1950, a mathematician, Alan Turing, wrote a paper that opened with the question “Can machines think?”—and immediately abandoned it. He found it too vague to be worth discussing, because no one can agree on what “thinking” means. In its place he proposed another, clearer question: can a machine, through written conversation, persuade a human being that it is human? He did not replace intelligence with something equivalent. He replaced it with language. And he knew that this was what he was doing—he explicitly wrote that the original question was too meaningless, and predicted that by the end of the century the use of words would have changed so much that one would be able to speak of thinking machines without expecting contradiction. He did not predict that machines would acquire intelligence. He predicted that our language would change so that we would say they had. And he was right, except that his prediction concerned the words, not the thing.

Five years later, in 1955, a young mathematics professor at Dartmouth, John McCarthy, needed a name for a new field of research. One was already available: cybernetics, the name Norbert Wiener had given a few years earlier to a theory of how systems—biological or mechanical—correct their behaviour through feedback, like a weapon that continually corrects its aim by tracking where the target is going. McCarthy did not want it. He found it narrow, tied to circuits and analogue signals, and he wanted neither to accept Wiener as an intellectual guide nor to become embroiled in a disagreement with him. So he invented another name: artificial intelligence. In his own surviving words, he chose it partly for its neutrality. Many of his colleagues disliked it—they preferred to call their work complex information processing. The name prevailed nonetheless, and university departments and streams of funding followed it.

In other words, the name did not emerge from a discovery. It emerged from a choice among candidates at a moment when there was not yet anything to name—an ambition looking for a title, not an object asking to be described. And the ambition was sincere: the 1955 funding proposal stated that every aspect of learning or any other feature of intelligence could in principle be described so precisely that a machine could be made to simulate it. Notice the word: simulate it. Not possess it. Even they knew the difference, in the opening line of their founding text.

Twice they tried to build intelligence in the way we imagined it then—by giving the machine rules, definitions and logic, step by step, as though writing a manual for someone who knew nothing about the world. Each time the effort hit a wall and funding stopped; these periods are called the AI winters, one in the 1970s and the other around 1990. The most ambitious project in this line attempted to record, by hand, all the common sense possessed by an adult human being—that water makes things wet, that no one can fit into a box smaller than themselves, thousands of such self-evident truths—and ended up showing only how impossible the undertaking was. Intelligence would not yield itself to the hand.

What finally worked was the exact opposite. Rather than supplying rules of meaning, a system was allowed to guess, billions of times, which word follows which across a vast body of text, until it became good at guessing. No one explained to it what any word meant. It learned only which words occur next to which others, so many times and with such precision that the result resembles understanding. Intelligence, as the goal of construction, was quietly abandoned. In its place came language as the object of modelling—the very move Turing had made on paper thirty years earlier, now at the level of the machine. And the name the technical work itself gave this, from the 1980s when such statistical systems first appeared and again around 2018 when they grew dramatically, was not invented in a room for the sake of neutrality. It arose naturally and descriptively: language model, because that is exactly what it does. No one needed to advertise the term, because no one outside the field was listening.

In 2021, three linguists and a researcher published a paper with a title that endured: they spoke of “stochastic parrots.” Their argument was precise and needs no embellishment: these systems learn statistical relations among sequences of words, and when they produce text that sounds meaningful, coherent, even self-aware, they reproduce patterns from what they have read, not any understanding or experience of their own. There is, they wrote, a fundamental gap between form and meaning. A parrot can repeat a phrase perfectly, with the right tone at the right moment, without knowing what it means—and the wonder lies not in the parrot but in the ear that hears it and supplies the meaning for itself.

The paper was not ignored. The entire field read it; it provoked a dispute inside the very company where one of the authors worked, and became a reference point known to every serious researcher. This is not, then, a question no one asked. It is a question posed clearly and publicly by people within the field itself, after which something strange happened: the name did not change. No one seriously proposed renaming the field “artificial language.” The same companies released such a system to the public the following year, and within one week a million people had tried it, within two months a hundred million—figures no other technology had reached so quickly. There was no longer time for the question of whether this was intelligence or language to be raised before a public that was only just encountering the thing and was impressed by it. And here a small, curious detail is worth noting: the technical architecture on which this system was built came from a paper entitled “Attention Is All You Need”—there, “attention” is a technical term for how the system weights which words relate to which. It has nothing to do with the other attention, that of the infant turning toward someone. And yet the entire edifice was built upon the wrong attention, while the right one, the attention that turns toward another and waits for them, never entered the plans.

Why, then, did the name not change when the question was already known? Not because no one knew. Because no one had a reason to correct it. “Artificial intelligence” frightens, impresses, attracts funding and commands trillions in market value. “Artificial language” sounds like a translation tool. There is something deeper as well: for millennia Western civilisation has identified intelligence with logos—the same word in Greek bears both senses, thought and language together. Anyone raised within this identification, on seeing something handle language perfectly, needs no proof to believe that it thinks; they infer it automatically. This is no one’s deception. It is a two-thousand-year-old assumption that has finally been tested, and the test produced an ambiguous result in a way that serves no one to clarify.

And there is proof of this within the profession itself that requires no interpretation. The engineers who build these systems do not call them “artificial intelligence” among themselves. They call them large language models—the technical term used in their papers and internal discussions, where precision matters. The word “intelligence” returns only when the same thing is turned outward—toward the public, investors and regulation. These are not two names for two things. They are two names for the same thing, depending on who is listening. And those who know the precise word and use it in their work did not forget it the moment they turned outward. They simply chose which word suited which audience.

Something must be said honestly, because the issue did not remain where it was left in 2021. The companies themselves, looking inside their own systems, found something more complex than pure parroting. They found indications that when such a system writes verse, it appears already to have chosen the word with which it will rhyme well before it reaches it—as if planning ahead rather than merely guessing one word after another. They found that it works with something resembling abstract concepts, not merely sequences of letters. This is a real finding, and it would be dishonest to ignore it because it does not suit the argument. But it does not resolve the question with which we began, the infant and the puppy. Planning a rhyme two words in advance is not the same as wanting to say something to someone. Internal complexity may exist without there being anywhere toward which it is directed. Emergence shows that something is produced which no one explicitly designed into the system; it does not show that what is produced is addressed anywhere.

Finally, there is a word in our language that held the answer before the question had even been asked, and we missed it because we use it every day without hearing it. Technology: technē and logos, discourse upon art. Beside it is the same pairing in reverse, so common that we fail to notice it too: literature, the art of the word. The two words hold the same two elements, art and logos, with only the order of precedence reversed. Our language always knew that art and logos could give birth to one another, in either direction. We did not need to wait for 1950 to suspect that the most technological thing we would ever make would ultimately be pure language with nothing around it. The words had been saying it beforehand, and we did not listen.

An infant turns its head toward where its mother is looking, months before it says its first word. A puppy does the same throughout its life, without ever having a word. And somewhere between the two lies what we have built: all the language in the world, with no one saying it to anyone. It does not lack the word. It lacks the turning of the head.