The unexpected mirror: AI has become a laboratory for studying your own brain

THE MIRROR NOBODY ORDERED. For years the big challenge has been opening AI's «black box»: understanding why a model answers the way it does. That discipline is called interpretability. But something is happening in the opposite direction, and almost nobody tells it. Language models —LLMs, systems trained to predict the next word in a text— are being used as a «model organism» to open the other black box: the human brain. In biology, a model organism is a simple species, like the fruit fly or the mouse, studied to understand something more complex. Now the object of study is us. For the first time in history there is a non-human system that produces fluent language. That, more than a curiosity, is a laboratory. Our thesis: AI does not prove how the brain works, but it does offer, at last, a computable theory of language that can be tested against data. That is the real story.

THE EARTHQUAKE IN LINGUISTICS. The first tremor is felt in the theory of language. For half a century the dominant idea was Noam Chomsky's: we are born with a «universal grammar», an innate grammatical scaffolding. The classic argument is called «poverty of the stimulus»: the child hears too few examples, and too messy ones, to deduce the rules of a language alone; therefore part must come built in. LLMs complicate that story. Linguist Steven Piantadosi, of the University of California, Berkeley, argues in a 2023 paper that these models «refute» almost every strong claim about the innateness of language: a network with no grammar rules preloaded learns syntax from exposure to text alone. Chomsky has dismissed these systems as «a trick» and, in a 2023 New York Times essay, as a form of «high-tech plagiarism». But here comes the fact that tilts the debate toward evidence. Chomsky had claimed that LLMs learn «possible» human languages just as well as «impossible» ones, which would prove they tell us nothing. A Stanford team tested that idea in 2024 («Mission: Impossible Language Models»): they built impossible languages —with rules no human uses, such as counting positions or shuffling words— and found GPT-2 learns them worse and more slowly than English. That specific Chomsky claim, then, is empirically false. Our reading: beware of rushed conclusions. A machine extracting grammar from text does not prove the child does it the same way; it only proves a generic learner can do it given enough data. And the Chomskyans answer with a legitimate argument, about the amount of data, which we reach shortly. What matters is that a debate stuck in rhetoric is again being settled by experiments. That is healthy.

WHEN THE MODEL PREDICTS, SO DOES THE BRAIN. The second tremor comes from neuroscience, and it is more striking. Uri Hasson's lab at Princeton recorded the brain activity of nine people as they listened to a 30-minute podcast, with electrodes placed on the cortex (a technique called electrocorticography). In their 2022 study in Nature Neuroscience they found three principles the brain shares with GPT-style models. One: the brain, like the model, predicts the next word before hearing it. Two: it registers «surprise» when it predicts wrongly, a prediction-error signal. Three: it represents each word by its context, not in isolation (what AI calls a «contextual embedding», a numerical representation whose meaning depends on the surrounding words). A 2025 paper in Nature Communications went further: an LLM's layer-by-layer processing maps onto the time course with which the brain's language areas process a sentence. And Evelina Fedorenko's teams at MIT have shown that an LLM's internal activations predict the response of the human «language network» —a specific set of regions in the left hemisphere— better than any earlier model, measured with functional magnetic resonance imaging (fMRI, a technique that tracks blood flow as a proxy for neural activity). Why do they converge? There is a common theory: «predictive processing». The idea, formalized in 1999 by Rao and Ballard, holds that the brain is a prediction machine that endlessly anticipates what it will perceive and learns from the error. A transformer minimizes exactly that: the error in predicting the next unit of text. Same family of idea. Our reading: the convergence is real and repeats across several labs, but convergence is not identity. Two systems can land on the same solution —prediction— without being the same machine. Hasson himself stresses it: the principles look alike; how they are implemented inside each black box is radically different.

THE FINE PRINT. Here it pays to slow down. Some differences do not transfer, and they are huge. The first is the data diet. A child has heard fewer than 100 million words by age 13, according to the academic BabyLM consortium. A frontier LLM is trained on trillions —3 to 6 orders of magnitude more, that is, up to a million times more text—. They learn the same thing from incomparable amounts. That is precisely the Chomskyans' strongest counterargument: the human learns from the poverty of the stimulus; the machine, from abundance. The BabyLM challenge softens the blow —with good methods, models trained on just 100 million words rival trillion-word ones— but the natural gap remains vast. The second difference: the child has a body, goals, emotions and a developmental arc; they learn language while acting in the world. The LLM learns from static text, with no body and no purpose of its own. And there is a third caution, about the alignment itself. A 2025 study by Badr AlKhamissi and colleagues (EPFL and MIT) found that the resemblance to the brain appears very early in training —around 2 to 8 billion tokens (a token is the smallest unit of text, roughly a word or part of one)— and then plateaus; as the model grows more capable, it «outgrows» the human pattern. And it is worth attributing the skepticism: a 2025 preprint by Nima Hadidi, Ebrahim Feghhi and colleagues (UCLA) argues that much of the measured alignment is a «mirage» produced by fragile methods and confounding variables; with proper controls, they claim, it shrinks. We take no side. We flag it. The honest state of the question is that the signal exists, but its size is disputed.

THE UNCOMFORTABLE FINDING: LANGUAGE IS NOT THOUGHT. The deepest twist comes from Fedorenko's own lab. In a 2024 paper in Nature, her team argues that the language network is for communicating, not for thinking. The evidence is blunt: people with aphasia —the loss of language after brain damage— can still do algebra, play chess and reason about what others are thinking. Thought runs on other networks. This finding fits astonishingly well with a distinction that comes from AI: «formal» competence (grammar, fluency) versus «functional» competence (reasoning, world knowledge). LLMs nailed formal competence. The human language network also deals, above all, with formal competence. Hence the mirror's sharpest lesson: fluent language and thought are separable. That deflates both exaggerations at once. It deflates the hype —«it talks, therefore it thinks»— because a system can master words without being a mind. And it deflates the dismissal —«it's just autocomplete»— because mastering words turns out to be, in us too, a self-contained system of its own. Our reading: AI may have solved the part of language that was never the part about thinking. That does not make it less useful. It makes it easier to understand.

OUR READING AND THE HORIZON. Why does this matter outside academia? Because neuroscience, for the first time, has a model of language that works and can be tested, instead of a box of metaphors. That is a tool. It can sharpen what we know about aphasia, dyslexia and language disorders. It can power brain-computer interfaces that decode the speech someone wants to produce but no longer can —already demonstrated in the lab for people with paralysis—. In the short term, the discipline is clear: do not oversell, do not humanize the machine, measure the effect and also its statistical traps. In the long term, this fits squarely into the hopeful horizon we defend. Understanding the brain's code is a step toward repairing it: restoring speech, treating neurological disease, extending healthy life. The mirror is smudged and warped. But it is the first real mirror humanity has held up to the machinery of its own language. Learning to read it, carefully, is one of the great quiet stories of this decade.