What this covers
Edward Gibson presents a theory of human language built on three interlocking claims: that language is fundamentally a system humans invented to communicate efficiently, that its form is neurologically and functionally separate from thought itself, and that dependency grammar describes how it actually works. The core argument rests on the observation that across languages, word order systematically minimizes the distance between words that depend on each other—a pattern Gibson has documented quantitatively across sixty language corpora. Brain imaging evidence, he contends, shows that the language network is anatomically distinct from networks for math, music, spatial reasoning, and code, and that people with language loss can still think, calculate, and strategize. This separates Gibson's view sharply from Chomsky's tradition, which posits innate universal grammar and treats language as a window into thought itself.
The conversation ranges widely across what that theory explains and what it unsettles. Gibson shows how center embedding—nested clauses that force readers to hold multiple long-distance dependencies in mind—becomes incomprehensible in every language, including in large language models, and how legal language's density of center-embedded structures explains why contracts are so hard to parse. He argues that learnability, not communication alone, shapes grammar's rigidity; that constructed languages like Klingon activate the language network precisely because they function like real languages; and that languages have fewer color words or number words than perception permits because cultures invent vocabulary for what they need to discuss. The material also touches on contested ground: whether the Piraha language truly lacks exact number concepts, whether the inner voice most people report is truly incidental to thinking, and whether claims about language's uniqueness hold up against what we actually know about animal communication.
Gibson argues that human language is an invented communication system optimized primarily to minimize dependency length between words, that form and meaning are separable (as shown by language being neurologically distinct from thought), and that grammar is best modeled by dependency grammar rather than Chomsky's movement-based phrase structure.
- Across and within languages, word order minimizes dependency distance, explaining harmonic word-order generalizations
- fMRI evidence shows the language network is distinct from networks for thinking, math, music, and code
- Center embedding (long nested dependencies) is universally hard, and even LLMs fail at it like humans
Dependency-length minimization organizes word order across languages to reduce processing difficulty.
- Across roughly 60 languages with parsed dependency corpora, the actual dependency lengths in real sentences are consistently shorter than in random reorderings of the same words, providing quantitative cross-linguistic evidence for dependency length minimization.
“the dependency lengths are always shorter in the real language compared to this kind of a control.”
- Gibson conjectures that dependency-length minimization likely evolved primarily for ease of production (the speaker) rather than comprehension, holding meaning constant—it's easier to say something when connected words are close together—though production and comprehension are hard to disentangle.
“I would guess it’s probably evolved for production. It’s about producing. It’s what’s easier for me to say that ends up being easier for you also.”
- Languages organize word order to minimize the distance between words that are grammatically dependent on each other, because shorter dependencies are easier to produce and comprehend; this explains why verb-initial languages tend to have prepositions and verb-final languages tend to have postpositions (harmonic word order).
“what we’re trying to do in a language is actually minimize those dependency links. It’s easier for me to say things if the words that are connecting for their meaning are close together.”
- There is a measurable cognitive cost associated with producing and comprehending dependencies between non-adjacent words, and the further apart the connected words are, the higher the cost; this can be measured by grammaticality ratings, reading times, and even the magnitude of activation in the brain's language network.
“there is a cost associated with producing and comprehending connections between words, which are just not beside each other. The further apart they are, the worse it is.”
- Nested center-embedded structures (where modifying clauses are stacked between subjects and verbs) become uninterpretable in every language, not just English, because they create multiple long-distance dependencies that overwhelm processing—even though each individual piece follows the language's rules.
“No matter what language you look at... it won’t be good. Guaranteed. 100% that will be uninterpretable in that language in the same way that was uninterpretable in English.”
- Languages have rigid, regular word-order rules (like English) even though pure communication efficiency would predict very free word order; this regularity probably exists because predictable, regular patterns are easier to learn, meaning learnability—not just communication—shapes grammar.
“If it were just about minimizing dependency lengths... then we might find grammars which didn’t have regularity in their rules. But languages always have regularity in their rules.”
- Legal language is hard to understand primarily because it is heavily center-embedded—about 70% of sentences in random contracts and laws contain a center-embedded clause (versus ~20% in other texts, including academic texts)—and experiments show that un-center-embedding the text, more than fixing passive voice or low-frequency words, dramatically improves comprehension and recall for both laypeople and lawyers.
“I think you get about 70%, something like 70% of sentences have a center embedded clause, which is insanely high. If you go to any other text, it’s down to 20%”
- Gibson's best current theory for why legalese is written in convoluted center-embedded style—despite even lawyers disliking it and being no better served by it—is that center embedding functions as a performative stylistic marker signaling 'this is legalese' and 'this is certainly true', analogous to how rhyming signals a magic spell.
“We think that that’s the kind of a style which tells you it’s legalese... it’s like a magic spell. We kind of call this the magic spell hypothesis.”
Exact number words and color vocabularies exist only where cultures need them, not universally.
- The Piraha language has no words for exact numbers, including no word for 'one', 'two', etc.; their quantifiers mean approximate concepts like 'few', 'some', and 'many' rather than exact counts, making it impossible to ask for an exact quantity of something.
“Naively I certainly thought that all humans would have words for exact counting, and the Piraha don’t. Okay, so they don’t have any words for even one. There’s not a word for one in their language.”
- The Piraha can perform one-to-one matching tasks perfectly regardless of set size (which requires no counting), but fail on tasks requiring memory of exact quantities above four or five (like matching a hidden set), demonstrating that having exact number words is the limiting factor enabling exact-quantity cognition.
“they would get them approximately right after four or five... it kind of shows that the words, you kind of need the words to be able to do these kinds of tasks.”
- The anthropological hypothesis for why some cultures invented exact counting systems is farming: tracking a set of livestock (e.g., goats) over time benefits from an abstraction over the set—a count—since you care about the number rather than each animal's identity, unlike tracking children whose individual identities matter.
“if you have a bunch of goats and you want to keep track of them... it’s easier to have a count system to do that. That’s an abstraction over a set.”
- Although humans with normal trichromatic vision can perceive millions of color distinctions, languages have far fewer color words (about 11 in English, sometimes only two in non-industrialized cultures like the Dani), because color vocabulary evolves based on what a culture needs to communicate.
“If you have normal color vision, normal trichromatic color vision, you can see millions of distinctions in colors. We don’t have millions of words.”
- Language is an invented system where humans create labels (words/morphemes) for the things they need to talk about in their world, which is why concepts irrelevant to a culture's life (like specific color shades or exact numbers) lack words in some languages.
“If you have normal color vision, normal trichromatic color vision, you can see millions of distinctions in colors. We don’t have millions of words.”
Language form is learnable from input without requiring innate universal grammar rules.
- Gibson argues that much of language, including its forms, can be learned from input rather than being innate; Chomsky posited innateness mainly because movement is hard to learn, but abandoning movement removes the need for substantial innate grammatical structure, which is consistent with LLMs learning language form well.
“I don’t really see a reason to postulate much innate structure. And that’s kind of why I think these large language models are learning so well is because I think you can learn the form, the forms of human language from the input.”
- Chomsky's grammar combines phrase structure with movement (e.g., auxiliary verbs moving to the front to form questions), but in 1971 he showed this combination creates learnability problems—it becomes impossible for a child to figure out the underlying structure—which he resolved by positing innate universal grammar; Gibson argues that abandoning movement in favor of lexical copying makes language learnable from input.
“Chomsky really famously, in 1971 showed that the movement story leads to learnability problems. It leads to problems for how language is learned.”
- The existence of specialized, modular brain areas does not prove they are innate; the visual word form area, for example, is highly specialized for reading but only develops in people who learn to read, showing modularization can be learned through experience.
“there’s something called the visual word form area... if you’re a reader. If you don’t read, you don’t have it... So the modularization is not evidence for innateness.”
Large language models predict grammaticality and model processing patterns better than Chomskyan theory.
- Large language models are arguably the best current theory of human language because they predict what is and isn't grammatical English better than any other theory, but they model form, not meaning—they are an opaque black box that is too large to be a good (simple) theory.
“the reason it’s arguably the best is that it does the best at predicting what’s English, for instance. It’s incredibly good, better than any other theory, but there’s not enough detail.”
- Large language models fail to complete deeply center-embedded (nested) sentences in exactly the same way humans do, despite not being explicitly trained to fail, suggesting they may be modeling the way human language form is processed and generated.
“If you ask a human to complete those, they can’t do it. Neither can a large language model. They’re just like humans in that.”
- Large language models are excellent at the form of language but do not understand meaning; they can be tricked by variations on familiar problems (like a modified Monty Hall problem where the answer is obvious) because they follow the form of solutions seen repeatedly in training rather than reasoning about the actual situation.
“I would argue they’re doing the form. They’re doing the form, they’re doing it really, really well. And are they doing the meaning? No, probably not.”
- Syntax (the forms of language) is a more tractable problem than meaning; studying form is comparatively easier while meaning remains a very hard problem, which is why approaches focusing on form—including LLMs—succeed where meaning-based approaches struggle.
“the syntax is easier, which is just the forms as opposed to the meaning. When you’re starting talking about the meaning, that’s a very hard problem”
Language is a specialized, modular brain system separate from thought and general cognition.
- fMRI can localize an individual's left-lateralized language network in about 15 minutes, and this network is remarkably stable over time—as stable as one's face—remaining essentially identical across nearly two decades of scanning.
“my language network is exactly the same a month ago as it was back in 2007. It’s amazingly stable.”
- Constructed languages like Klingon or those from Game of Thrones activate the brain's language network because they were designed to function like human languages and can express arbitrary thoughts, whereas computer programming languages do not activate it.
“that lights up the language area because they can speak pretty much arbitrary thoughts in a human language. It’s a constructed human language”
- The brain's language network is not engaged by non-linguistic thinking tasks: spatial memory, music perception, arithmetic, and computer programming all activate other brain regions, and global aphasics who have lost their language network can still play chess, do math, and drive—demonstrating language is a separate system from thought and is not necessary for thinking.
“none of those tasks tap the language network at all. At all. There’s no overlap.”
- The neurological evidence implies that we don't think in language; the inner voice that most people (roughly 70-80%) report is not the substrate of thought, since thought occurs separately from the language network.
“I guess the implication here is that we don’t think in language.”
Word-order patterns and verb-initial or verb-final structures show systematic cross-linguistic regularity.
- About 40% of the world's languages are subject-verb-object (like English), and a similar or slightly larger proportion (45-50%) are verb-final (subject-object-verb, like Japanese or Hindi); these two orders are by far the most common.
“About 40% of the languages of the world look like that, they’re subject-verb-object languages.”
- English word order changed dramatically over a thousand years: Old English was a verb-final language with case marking, like German, and through extensive contact with French it lost case marking and became a verb-medial (verb-initial-type) language.
“It used to look like German. It used to be a verb final language with case marking. And it shifted to a vermedial language, a lot of contact with French.”
- In every language, sentences can be represented as dependency trees where each word is connected to exactly one other word, with a root and branches; linguists broadly agree on this structure even if they disagree about the details.
“every word is connected to just one other word in that sentence. And so you end up with what’s called technically a tree, it’s a tree structure”
Language spread and death depend on economic incentives, not structural properties.
- In English, high-frequency words tend to have irregular forms (e.g., drink/drank), while the much more numerous low-frequency words tend to be regular (e.g., walk/walked).
“The frequent ones, the common words tend to be irregular. There’s many, many more low frequency words and those are regular ones.”
- The claim that human language is uniquely special (e.g., having compositionality) is not well-supported by evidence, because we cannot yet understand the communication systems of whales, crows, or monkeys; higher species with apparent communicative abilities may have far more signal—possibly something complex like human language—than we assume.
“this is a very odd reasoning to be making that human language is special because we’re the only ones who have human language... we can’t talk to them yet.”
- Gibson's biggest difference from Chomsky is methodological: Gibson gathers quantitative data through experiments and corpus analysis, while Chomsky relies on thought experiments and his own intuitions and has never been associated with any experiment or corpus work.
“our biggest difference is our methods... I gather data from people. I do experiments with people and I gather corpus data... He just doesn’t do that.”
- Languages die when there is little economic value in the native language relative to a dominant language; people learn the language that lets them make money and feed their families, which is why universal constructed languages like Esperanto don't take off and why English, Mandarin, and Spanish spread—driven by the size of their economies.
“there’s not a lot of value for the local people in their native language. So there’s much more value in knowing Spanish because they want to feed their families.”
- There is no evidence that any human language is intrinsically harder or easier for children to learn—by age three or four, children speak whatever their native language is—although second-language difficulty depends heavily on how close the second language is to one's first.
“There’s no evidence that there’s anything harder or easy about any language learned, by three or four, they speak that language.”
- Gibson found grammar fascinating as a kind of mathematical puzzle from a young age and approached language from an engineering and computer science perspective.
“I found language as a neat puzzle to work on from an engineering perspective actually”
Lexical copying better explains auxiliary-verb behavior than movement rules.
- The lexical copying account is favored over movement because auxiliary verbs don't all behave uniformly: some work only in interrogative form ('Aren't I invited?' but not 'I aren't invited'), which is naturally explained by learning word-specific usage but is hard for a general movement rule.
“it predicts that not all auxiliaries might move. It might depend on the word, and that turns out to be true.”
Translation difficulty arises from absent concepts, not surface form differences.
- Machine translation will be hardest where concepts exist in one language but not another—translating exact number concepts into Piraha is impossible because the words don't exist—so translation difficulty lies in the space of concepts (meaning) rather than the space of form, which can generally be mapped.
“good luck translating a lot of English into Piraha. It’s just impossible... there are no words for these concepts”