🍼 In 2004, a Japanese science book asked a deceptively simple question: how does a baby learn to talk? The answer, given by one of Japan's leading brain scientists, was that we are born already wired for grammar. Eighteen years later, a chatbot learned fluent language with no such wiring at all — just by reading the internet. That collision reopened a debate that has run for nearly seventy years, and it is still wide open today.

A question with a confident answer

The book was What Are Researchers Seeing at the Frontiers of Science?, a 2004 collection of dialogues by the novelist Hideaki Sena, drawn from interviews he ran in the magazine Nikkei Science between 2002 and 2004. One chapter — "How Do Babies Learn Language?" — featured Kuniyoshi L. Sakai, a neuroscientist at the University of Tokyo who studies language using brain imaging.

The publisher summed up his chapter's premise in a single line: the brain comes equipped, from birth, with a mechanism for building grammar.

That idea is not Sakai's invention. It traces back to the linguist Noam Chomsky, who argued in the late 1950s that children acquire language far too quickly, and from far too little and messy input, to be learning it from scratch by imitation. Something must already be in place — a kind of built-in scaffolding Chomsky called universal grammar. Linguists know the core argument as the "poverty of the stimulus": the speech a toddler hears simply doesn't contain enough information to explain how completely and how fast they master their native tongue. Sakai's own research uses fMRI to look for where that machinery might live in the brain, pointing to specific regions in the left frontal lobe as a candidate "grammar center."

In 2004, this was the frontier. A baby's mind was treated as a small machine running pre-installed software, and the job of science was to find the code.

Then a machine learned to talk without any of that

On November 30, 2022, OpenAI released ChatGPT. It reached a million users in five days and around a hundred million within two months — the fastest adoption of any consumer app at the time.

What makes ChatGPT relevant to a debate about babies is how it works. A large language model is, at bottom, a system that reads staggering amounts of text and learns to predict the next word. It is given no grammar rules, no built-in scaffolding, no theory of language. It just absorbs patterns from data — and out the other end comes prose fluent enough that most people can't tell it from a human's.

For anyone who had spent decades arguing that fluent language requires innate structure, this was an uncomfortable result. Here was a system with none of the supposed prerequisites, producing exactly the thing those prerequisites were meant to explain. The UC Berkeley cognitive scientist Steven Piantadosi made the empiricist case in the title of a 2023 paper: "Modern Language Models Refute Chomsky's Approach to Language." His argument was that LLMs undercut nearly every strong claim for language being innate, by succeeding at what generative theory said couldn't work without built-in structure.

The debate didn't end — it sharpened

Chomsky did not concede. In March 2023, he co-wrote an essay in The New York Times titled "The False Promise of ChatGPT," and the title says most of it. The human mind, he argued, is not a "lumbering statistical engine" gorging on data; it is an efficient system that works from tiny amounts of information and builds explanations — accounts of why something is or isn't the case, not just guesses about what word comes next. A child can grasp a rule from a handful of examples. A language model needs a sizeable fraction of everything ever written.

That gap is the crux. A toddler becomes fluent on a few years of part-time, error-filled, wildly incomplete input. ChatGPT needed text on a scale no human could read in a thousand lifetimes. If both end up speaking, but one got there on a teaspoon of data and the other on an ocean, are they really doing the same thing?

Sakai, for his part, has become one of Japan's more outspoken skeptics. In interviews and recent books on the brain and AI, he has argued that systems like ChatGPT don't actually understand meaning — they only appear to converse, assembling plausible word sequences without grasping the intent behind them. He has gone further than most, warning that rushing generative AI into classrooms before we understand how the mind works risks eroding the very thinking skills it claims to assist. Whether you find that persuasive or alarmist, it is a consistent position: the man who explained innate grammar in 2004 sees the machine of 2022 as something fundamentally unlike a human speaker.

The empiricists have a sharp reply: if a model with no innate grammar can do this, maybe grammar was never as "innate" as the theory needed it to be — maybe powerful general learning, fed enough data, is most of the story. Linguists are still publishing papers fighting it out, which is the honest signal here. Nobody has won.

What a chatbot can't tell you about your own first words

It's tempting to read ChatGPT as the final word — proof that language is just pattern-matching after all. But it shows only that one route to fluent language runs through brute statistical exposure on an ocean of text. Whether that's the route a baby takes is a separate question, and the baby got there on a teaspoon. So the interesting question is no longer "can a machine talk?" but "why can a child do it on so little?" — and every fluent adult once solved that one without lessons, a dataset, or any memory of doing it.

In Japan, some of the researchers who first asked how babies learn to talk — Sakai among them — are now also among the loudest voices urging caution about handing thinking over to AI. How is that debate playing out where you are? And when you think back, do you have any sense of how you learned your own first words?

References