The Grammar Underneath

In 1957, a twenty-eight-year-old linguist at MIT published a thin book called Syntactic Structures and quietly detonated the social sciences. The claim inside it was almost arrogant in its simplicity. A child of three, hearing a few thousand fragmentary, error-strewn sentences, somehow arrives at the ability to produce sentences no one has ever spoken before — an infinite output from a finite, accidental input. That gap, Noam Chomsky argued, cannot be closed by experience alone. Something must already be there. The mind is not a blank page that the world writes on. It comes with a grammar.

This is the idea that puts Chomsky on the river of intelligence, and it is worth stating precisely, because it has been blurred by sixty years of argument. Chomsky did not merely study language. He proposed that the capacity for language is innate, species-specific, and structured — a universal grammar wired into the human animal the way echolocation is wired into a bat. Beneath the thousands of human tongues, he claimed, runs one deep design: a formal architecture of abstraction and control that the surface noise of any given language only decorates. The differences are surface; the architecture is shared. We do not learn to be linguistic creatures any more than we learn to grow arms. We unfold into it.

The reason this matters now — the reason a 1957 monograph is suddenly the most contested document in artificial intelligence — is that we have built a machine that appears to refute him. Large language models learn language the way Chomsky said no one could: by counting. They invert his whole picture of acquisition versus learning — ingesting a near-total corpus of human writing and extracting the statistical shape of it, and out the other end comes prose that is fluent, supple, often indistinguishable from ours. No innate grammar. No deep structure handed down by biology. Just an ocean of text and an appetite for patterns inside it. The empiricism Chomsky spent his life demolishing has, it seems, come back and learned to talk.

Colorless green ideas sleep furiously. Chomsky's 1957 sentence — perfectly grammatical, perfectly meaningless

Chomsky never conceded the point. To the end — he is still alive, in his nineties, his health frail since a 2023 stroke — he held that these systems are not doing what we do. A model that achieves fluency by absorbing the entire internet has, in his view, explained nothing about the human mind, because it would learn an impossible language as readily as a real one. What it produces is a kind of AI mirror — a statistical reflection of everything we have written, optimized for plausibility, not a window into thought. It has no theory of why human grammars take the narrow shapes they take. It is, he said, a triumph of engineering and a failure of science: a kind of high-tech plagiarism that mistakes the prediction of the next word for an account of thought. Performance is not competence. A thing can speak without knowing what speaking is.

Why the river runs through him

Here is the honest reckoning, the one the duty of care demands. Chomsky may be wrong about the machines and still be the indispensable thinker for understanding them. His permanent contribution was not the specific claim about innateness — that remains fiercely disputed — but the question he forced everyone to take seriously: what does it mean to know a language? Before him, behaviorism treated the mind as a stimulus-response box, and his 1959 evisceration of B.F. Skinner's account of language is the moment the cognitive revolution began. He made structure the central object of inquiry — and in doing so kept the asking human in the loop, refusing to let the explanation collapse into the behavior it was meant to explain. He insisted there is something inside that the outside cannot fully explain. Every argument we now have about whether a transformer "understands" anything is a Chomskyan argument, conducted on terrain he cleared.

And the tension is genuine, not rhetorical. The models work, and their working is not nothing. But fluency is cheap and comprehension is expensive, and Chomsky's life is a sixty-year warning against confusing the two. A system that has read everything and grasped nothing is exactly the creature he taught us to be suspicious of — the living shape of amplification without comprehension. When a chatbot produces a beautiful, confident, wholly invented citation, it is doing the most un-Chomskyan thing imaginable — generating well-formed surface with no structure of belief underneath. He would say: of course it does. It was never built to mean.

The questions of real intellectual interest are too hard for it to handle, and even questions of marginal interest, when the answer is not on the record, it cannot deal with. Chomsky on the new language models, 2023

There is a cost to canonizing him, and I will not hide it. Chomsky's certainty hardened into something that could not bend, and a field has to be able to update. The machines learned more from raw text than his theory permitted, and pretending otherwise is its own kind of refusal to look. But the deeper lesson survives the empirical fight. Amplification is not understanding. We have built an amplifier so vast it can reflect the whole of what we have written, and we keep mistaking the reflection for a mind. Chomsky's gift, in the end, is a discipline of attention: keep asking what is actually happening underneath the fluent surface, and never let the eloquence of the output stand in for an account of the thing producing it.

That is why he belongs on this river. Not as the prophet who got the future right — he may not have — but as the conscience who refused to let us off easy. In an age building intelligences that talk before we agree on what talking is, the seam between augmentation and automation runs straight through his question. The man who spent his life on the grammar underneath is not a relic. He is the question we still have not answered.