TOPIC 7 · PHASE A COGNITION

The Language Brain

Speaking and thinking are two different machines

2026-07-13 · BigCat

Someone who can't produce a single word or follow a sentence can still do arithmetic, play chess, get a joke — language breaks, and thought carries on unharmed.

It's tempting to feel that language just is thought — that the running commentary in your head is your thinking. But the brain doesn't divide the labor that way. Some stroke patients lose language completely: sentences fall apart, they can't say what day it is — yet hand them a math problem and they solve it; set up a chessboard and they play. The language machine is wrecked, but the thinking machine keeps turning. This issue opens up "the language brain," and we'll hit two things that cut against intuition: the two famous language areas from the textbooks tell a story that's mostly wrong, and language and thought turn out to live in different rooms. That second fact is exactly the key to making sense of today's AI language models.

// 01

Not two dots and a wire

Open any old textbook and language is drawn as two dots: Broca's area near the forehead handles "speaking," Wernicke's area above and behind the ear handles "understanding," and a bundle of nerve fibers wires them together. That picture is a hundred and fifty years old — back when doctors could only wait for a patient to die, dissect the brain, see which part had broken, and reason backward from what ability had been lost.

OLD · two dots NEW · one network Broca speak Wernicke grasp
Language isn't two bricks — it's a whole tightly-linked network across the left brain

The trouble is, modern brain imaging dissolves that clean picture on contact. The truth: language runs on a whole tightly-interconnected sheet stretching from the left frontal lobe to the temporal lobe — far more than two dots. And even those two "classic" areas don't split the job so neatly — knock out Broca's area alone and you often won't get lasting speech loss; to really destroy language you have to take out a much broader chunk of the network. The old map isn't all wrong; it just flattened a network into two words. (For which regions and connections make up that net, language network)

// 02

The language room — thought doesn't live there

This network has a striking personality: it lights up for language and ignores nearly everything else. Neuroscientist Evelina Fedorenko did something very clean — first she pinpointed each person's own language network using sentences, then she poked it with all sorts of non-language tasks: arithmetic, listening to music, reading computer code, figuring out what someone else is thinking. The network barely stirred. It only lights up when you comprehend or build meaningful language.

"She read the letter" 7 × 8 = ? ♪ a melody left brain · language net lights for language ON
Only meaningful language switches it on; arithmetic and music flow right past

Put that finding together with the patient from the opening and the conclusion hardens: language is a dedicated communication tool, not thought itself. People with severe aphasia can still do arithmetic, reason, grasp cause and effect — thinking runs on other circuits. And the reverse holds: infants who can't yet speak, and some congenitally deaf people who grew up without exposure to language, still have rich thought. That chattering "inner monologue" in your head is more like thought's announcer than thought itself — it translates the already-formed idea into words, but the thinking got done before you opened your mouth.

AI cross-read

Hold on to this "language ≠ thought" split — it's the key to reading AI language models in the final section. The brain files "wording things well" and "reasoning" in two separate systems. Today's AI, by contrast, is trained by cramming mountains of text into one and the same network — so the two get welded together. That difference is worth chewing on more than the resemblance is.

// 03

Understanding a sentence means guessing all the way through

You think listening to someone is "reception" — sound comes in, meaning comes out. In fact your brain is jumping the gun the whole time: before the sentence is done, it's already guessing the next word. This isn't mysticism — EEG shows it directly. Read you "he was thirsty, so he poured a glass of —" and your brain has "water / tea / coffee" primed; if what actually shows up is "socks," your EEG throws out an unusually large wave that scientists call the N400 — and its size is essentially a ruler for "how unexpected that word was."

"…a glass of water" "…a glass of socks" expected · tiny wave surprise · a big N400 wave
The more unexpected the word, the bigger the brainwave — comprehension is itself a prediction

See the pattern? This is the same thing as perception as inference from the visual pathway issue, just wearing a different organ: the brain doesn't receive passively — it guesses first, then corrects with real input. Language is just the fanciest application of that prediction machine. It's why you can fill in the half-sentence you didn't catch in a noisy restaurant, and why you can instantly get a friend's meaning before they finish — all thanks to that word-guessing engine racing ahead.

AI cross-read

Here comes the most literal brain↔AI face-match on the whole site. A large language model (the ChatGPT kind) does exactly this, literally: give it the first half of a sentence and it predicts the most likely next word, one after another — its entire training objective is "guess the next word right." Even wilder: scientists took the internal activity of these word-guessing models and lined it up against the language-network activity of a human hearing the same sentences, and it matched astonishingly well; the better a model is at guessing words, the more it looks like a human brain. Two independently-grown paths — carbon evolution and silicon training — hit on the same trick for understanding language: predict the next word.

// 04

Speaking well isn't thinking well

Precisely because the word-guessing trick lines up so tightly, a sharper question surfaces: does an AI that speaks flawlessly therefore "think"? Stitch the two ends of this issue together and the answer is clear — no.

Fedorenko's group offers a clean cut: language ability comes in two layers. One is formal competence — is the grammar right, do the words fit, does the sentence flow — purely "does it sound like language." The other is functional competence — using language to reason, do math, tie into the real world, track who knows what. The brain assigns these two layers to different systems: the language network handles form, while reasoning, arithmetic, and social cognition have their own territories. That's exactly why aphasia patients keep "the words gone, the mind intact."

Today's language models happen to have drilled form to near-perfection while function is another matter — they'll write airtight sentences yet stumble on a simple multi-step inference or a spot that needs real-world common sense. They're so fluent, so fluent that we can't help projecting "it really understands" onto them. The one line to take from this issue: fluency is form, thinking is function; sounding good has never guaranteed thinking clearly. True of AI — and just as true of people.

AI cross-read

This draws a cleaner line under the "does AI really understand" debate: stop taking "how smoothly it talks" as evidence of "how smart it is" — that only measures formal competence. To really test it, you have to probe functional competence separately: multi-step reasoning, using real-world knowledge, matching words to reality. Tellingly, the fix also mirrors the brain: rather than expecting one network to do everything, do what the brain does and divide the labor — let the model call external tools (a calculator for math, retrieval for facts), splitting "can speak" from "can compute / can look up" into a relay. That's one of the main directions in AI engineering right now.

🌀 Crossing Over · Interdisciplinary Echoes

"Can language reach the real thought and being underneath?" — a few ancient traditions argued this out long ago:

// Going Deeper

If thought doesn't need language, what is that voice in my head that never shuts up?
That's inner speech — most likely the brain borrowing the language system to add "subtitles" to thought that's already taken shape, so you can grip it, order it, chew it over. The evidence: some people report having almost no inner monologue and still think, decide, and live perfectly well. So the inner voice is more of a commonly-used add-on than the engine. It genuinely helps — putting a thought into words can pin something vague into a clear string, which is why "writing it down made it click."
Language models predict brain activity so well — doesn't that mean they work "on the same principle" as the brain?
Careful with that jump. "Can predict" shows the two share structure at some computational level — both approximate the optimal solution to "predict the next word from context," so their internal representations end up looking alike. But that's not the same as implementing it the same way: the brain runs on tens of watts, learns as it goes, and is bolted to a body and world; the model runs on megawatts, freezes once trained, and lives only in text. What lines up is the task logic of "predict the next word"; what doesn't is the flesh and the energy underneath. Similar in function, divergent in implementation — a line the computational-neuroscience finale will keep returning to.
If "can speak" ≠ "can think," how badly do we err by judging intelligence from fluency?
Worse than you'd guess. Humans carry a deep bias: speaks smoothly = thinks clearly. So the eloquent get overrated and the inarticulate-but-deep get buried; that same bias is now poured wholesale onto AI — it's fluent enough that we instinctively fill in understanding for it. The fix is to test the two separately: never mind how pretty the talk is, only whether the conclusion survives questioning and matches the facts. That's the right way to grade AI — and a fairly hard rule for reading people, too.
Why is language almost always on the left side of the brain? What does the right side do in language?
In about nine in ten right-handers, and most left-handers, the core language network sits mainly in the left hemisphere — why it's so stubbornly there is still unsettled (one idea: the left brain is better at processing fast signals that unfold in time, and language is exactly that kind of signal). The right side isn't idle: it handles more of the tone, prosody, subtext, humor, and irony layer. People with right-hemisphere language damage can follow the literal meaning but often miss what's between the lines — every word registers, they just can't hear that you're joking. Which shows "understanding a sentence" was never only understanding the literal words.

// Further Reading