The term «artificial intelligence» is born in 1956, in a room at Dartmouth College. John McCarthy, a mathematician, convenes an eight-week seminar with Marvin Minsky, Claude Shannon, Allen Newell and Herbert Simon, among others. The premise they write into the proposal is luminously naive: every aspect of learning and human intelligence can, in principle, be described so precisely that a machine could simulate it. They asked for two months. Seventy years later, we still have not finished.
First road: the symbolic dream.
The fifties and sixties bet on logic. If human thought is language, and language can be formalized into symbols, then it is enough to write the right rules for a machine to reason. Newell and Simon build the Logic Theorist and have it prove mathematical theorems. It works. For a moment, people believe we are one step away from replicating thought.
But the world is not a theorem. Rules multiply, nuances escape, exceptions devour the system. By the seventies, symbolic AI enters its first winter: they promised too much, delivered little, governments cut the budget. The idea that thought is formal logic survived, but badly wounded.
Second road: experts in a can.
The eighties try something else: if we cannot model general intelligence, let us model the specific kind. Expert systems are born. Programs that imitate a doctor diagnosing, an engineer designing, a chemist synthesizing. They work in their domains and fail spectacularly outside them. Building one costs years and hundreds of interviews with specialists. By the time the interview is over, the specialist has already changed their mind.
Second winter. Stiff-necked AI does not scale either. Human intelligence seemed to be made of shortcuts, intuitions, happy contradictions. Coding it by hand was like copying the ocean drop by drop.
Third road: if we cannot reason it, let us predict it.
The nineties bring the bet that changes everything. Instead of programming rules, let us show the machine examples, thousands, then millions, and let it find the patterns itself. It is the era of statistical learning, of machine learning. The machine no longer understands; it predicts. And predicting turns out, surprisingly, to be enough for a great many things.
By the 2000s, neural networks, an old idea frozen for decades, come back with new computing power and gigantic datasets. They learn to see, to hear, to translate. In 2017 somebody publishes a paper titled “Attention is All You Need” and the transformer is born, the architecture holding up ChatGPT, Claude, Gemini. The machine no longer diagnoses like a doctor or proves theorems: it predicts which word is most likely after the last one. And that, at colossal scale, looks like intelligence.
We did not teach it to think. We taught it to look like it thinks, at scale. And at sufficient scale, the difference becomes philosophically uncomfortable.
The roads we did not take.
There were other routes. None closed entirely, but statistical prediction left them on pause out of simple economic efficacy. They are worth naming, because one of them could come back.
One was to deepen the symbolic road with far more sophisticated systems. An AI that really does reason with formal rules, able to explain every step of its decision. It would be slower, more rigid, but auditable. Today it would be invaluable in law, medicine, finance. We sacrificed it for speed.
Another was biological AI: brain-machine interfaces, neural networks that imitate real structures of the human brain and not only its name. Some laboratories persist, but the pace is geological compared to the transformer's curve. If it ever comes back, it will be through exhaustion of the predictive model, not by choice.
And a third, almost forgotten: artificial evolution. Systems that learn by trial and error in simulations, the way organisms evolve. Autonomous robots that train themselves by falling over. It is slow and expensive, but it produces behaviors nobody programmed.
Why this walk matters.
Because when somebody says «AI thinks», they are talking about one of three traditions, and the youngest of them. Statistical prediction is not intelligence: it is the shortcut we chose when the other roads got outrageously expensive. It is a brilliant shortcut, no doubt. But it is a shortcut, not a conquest. And there are people who argue the shortcut already understands: Hinton, for one.
Knowing this does not take power away from the machine. It gives power back to the human. Because if you understand you are interacting with a statistical actor trained on every available text, you stop asking it for truth and start asking it for probability. You stop believing it knows and start making use of what it predicts. The difference seems subtle. It is not.
Make-believe, again.
Today's artificial intelligence is, to a good extent, a collective make-believe. The machine looks like it thinks because we trained it to sound like somebody who thinks. And we, spectators trained by a century of cinema, are specialists in believing a screen. Put those two things together and you get the effect we are living in: a conversation with an extraordinarily brilliant actor who does not know he is acting, attended by an audience that forgets the set is empty.
The day we take another road again, improved symbolic, biological, evolutionary, or one we have not named yet, artificial intelligence will change its nature. For now, this is the one we have. And it is worth knowing it as mortal: neither god nor threat, but a historical shortcut loaded with consequences. The divine part is knowing what you work with. The tragic part is not knowing.