Edward Gibson on Human Language, Psycholinguistics, and LLMs

MIT psycholinguist Edward (Ted) Gibson argues that the shape of every human language, from English word order to the impossibility of certain legal sentences, follows from one simple pressure: keep the words that depend on each other close together. Along the way he makes the case, backed by brain-imaging evidence, that language and thought are separate systems — a finding with direct implications for what large language models are and are not doing.

Key ideas

1. Dependency length minimisation explains word order. Every sentence in every language decomposes into a tree of word-to-word dependencies. Languages keep those dependencies short because long-distance connections are harder to produce and to understand — and this single pressure predicts why verb-initial languages take prepositions, verb-final languages take postpositions, and roughly 95% of the world’s documented languages fit one pattern or the other.
2. Center embedding is the one universal way to break a sentence. Nesting one dependency inside another — ‘the boy who the cat, which the dog chased, scratched cried’ — lengthens every dependency at once and produces sentences no speaker of any language can reliably produce, understand, or even complete. Legal contracts embed clauses at roughly 70% of sentences, against 20–30% in ordinary technical prose, which is why lawyers and laypeople alike find them equally hard to parse and equally prefer the plain-language rewrite.
3. Language is not thought. fMRI work from Gibson’s MIT colleague (and wife) Evelina Fedorenko localises a dedicated ‘language network’, active for comprehending and producing sentences in any language a person knows, that stays silent during chess, arithmetic, music, and computer programming. Patients with global aphasia, whose language network has been destroyed by stroke, can still reason, plan, and do maths — evidence that thinking does not require language at all.
4. LLMs have mastered form, not meaning. Gibson calls large language models arguably the best current theory of a language’s form — better than any hand-built grammar at predicting what is grammatical — while arguing they still fail at meaning, citing their tendency to pattern-match on surface structure (as in a rephrased Monty Hall problem) even after being told the underlying facts. Tellingly, LLMs struggle to complete deeply center-embedded sentences in exactly the way humans do, unprompted — a sign that they have absorbed something real about how the human sentence-processing system works.
5. Words for number and colour are inventions of necessity, not universal cognition. The Piraha of the Brazilian Amazon have no words for exact counting at all — not even ‘one’ — and cannot reliably distinguish six identical objects from eight once the objects are hidden from view. Gibson’s account: cultures invent the vocabulary distinctions they need to communicate, not the ones their perceptual systems are capable of making.

Content

Dependency grammar and the shape of the world’s languages

Gibson frames language structure through dependency grammar, in which every word in a sentence connects to exactly one other word, producing a tree rooted at the verb. This tradition runs from the ancient Indian grammarian Panini through the French linguist Lucien Tesnière (1959) to Gibson’s own work; he prefers it to Noam Chomsky’s rival phrase-structure grammar because it makes the distance between connected words explicit, where phrase structure buries that distance inside nested brackets. The two formalisms are mathematically close to equivalent, generating the same sentences, but dependency grammar puts the quantity Gibson cares about — dependency length — in plain sight.

That framing explains a cross-linguistic pattern first catalogued by the Stanford typologist Joseph Greenberg: languages that put the verb before its object (English: ‘the dog chased the cat’) consistently take prepositions (‘to John’), while languages that put the verb after its object (Japanese, Hindi) consistently take postpositions (‘John to’). About 95% of the roughly 1,000 languages with adequate documentation — out of some 7,000 languages worldwide — fit one of these two ‘harmonic’ patterns. Gibson’s explanation, formalised in work by his former student Richard Futrell, is dependency length minimisation: scrambling a real sentence’s word order at random, even while preserving whether dependencies cross, produces reliably longer dependencies than the word order languages actually use. The effect holds across all roughly 60 languages with parsed corpora available, regardless of whether the language is verb-initial, verb-final, or (rarely) verb-first-of-all (VSO). See Dependency Length Minimisation.

Center embedding, and why legalese is the one exception

The clearest demonstration of the cost of long dependencies is center embedding: nesting one subject-verb dependency inside another, as in ‘the boy who the cat, which the dog chased, scratched, cried.’ Every language permits this construction grammatically, and every language makes it nearly impossible to produce, understand, or complete once nested more than once. Gibson’s crowdsourced completion experiments found that people given a partial doubly nested sentence (‘The book, which the author who…’) supply the wrong number of verbs roughly 60% of the time, even in writing with unlimited time.

Legalese is the sole systematic exception Gibson has found to the rule that natural languages keep dependencies short — and it does so by embracing center embedding rather than avoiding it. Working with the lawyer-turned-linguist Eric Martinez, Gibson found that roughly 70% of sentences in ordinary contracts contain a center-embedded clause, against 20–30% in other technical prose. A typical culprit is the habit of defining a term inline, mid-sentence, between a subject and its verb — ‘any payment or benefit by the company (…), such payments and benefits, including the payments and benefits under section 3(A) hereof, being hereinafter referred to as total payments, would be subject to the excise tax…’ Passive voice and rare vocabulary also appear more often in legal text, but neither affects comprehension once measured directly; only center embedding does, degrading recall and understanding equally in laypeople and in a sample of 100 practising lawyers Gibson tested. Lawyers were no better at parsing their own profession’s center-embedded sentences, and — asked directly — preferred the plain-language rewrites just as strongly as laypeople did. Gibson’s best guess for why the convention persists despite nobody wanting it is a ‘magic spell hypothesis’: center embedding functions as a stylistic signal that a text is binding law, the way rhyme signals an incantation, rather than a device anyone consciously chooses for its meaning.

The Chomsky disagreement: movement versus lexical copying

Gibson and Chomsky, MIT colleagues for over three decades, differ first on method — Gibson runs experiments and analyses corpora; Chomsky, by Gibson’s account, has never been party to either, relying instead on his own intuitions as thought experiments — and second on a specific claim about English auxiliary verbs. Chomsky’s phrase-structure grammar (from the 1950s) explains ‘will’ moving to the front of a question (‘Two dogs will enter the room’ → ‘Will two dogs enter the room?’) as literal syntactic movement from a single underlying ‘deep structure’ to a ‘surface structure’. Gibson prefers the alternative defended by the linguist Ivan Sag and others: lexical copying, in which a word like ‘will’ simply has two listed forms — declarative and interrogative — with no movement between them. The lexical account handles a wrinkle the movement account struggles with: some auxiliaries only work in one form. ‘Aren’t I invited?’ is a valid question with no corresponding declarative (‘I aren’t invited’ does not exist); ‘ought’ works declaratively but sounds archaic as a question (‘Ought I?’). Chomsky’s 1971 argument that movement is unlearnable from ordinary child input (and must therefore be innate, part of a universal grammar) is, on Gibson’s reading, an artefact of having assumed movement in the first place; drop movement, and the learning problem gets much easier, undercutting the case for innate grammatical structure.

Language and thought are different systems

The episode’s most striking empirical claim comes from the work of Gibson’s MIT colleague (and wife) Evelina Fedorenko, who uses fMRI to localise, in each individual’s brain within about fifteen minutes of scanning, a left-lateralised ‘language network’ that activates for comprehending or producing sentences — spoken or written, in any language the person knows, including constructed languages like Klingon that are fluent enough to carry arbitrary thoughts. That same network stays essentially silent during chess, spatial memory tasks, music perception, arithmetic, and — notably — reading or writing computer code, even for expert programmers. A separate, bilateral ‘multiple demands’ network handles effortful cognition of any kind instead. The clinical confirmation comes from patients with global aphasia: a stroke that destroys the language network (without also destroying the right hemisphere) leaves comprehension, production, reading, and writing devastated, while chess, driving, and arithmetic remain intact. Gibson takes this as evidence that language is a communication system for meaning, not the substrate of thought itself — directly opposed to Chomsky’s view that language underpins thought. See Language-Thought Dissociation.

What LLMs are and are not doing

Gibson calls large language models arguably the best current theory of a language’s form — better than any competing grammar at predicting what is and is not acceptable English — while immediately qualifying the compliment: a theory that is just a giant set of learned parameters is too large to count as an explanatory theory in the way a compact grammar would. Analysis by Christopher Manning’s group at Stanford suggests some LLMs internally build representations that resemble dependency structures, which would explain why their grammatical output is so reliable. Gibson locates LLMs closest to construction grammar, a usage-based theory in which the basic unit is a form-meaning pairing (a word, or a combination of words) rather than an abstract, Chomsky-style set of generative rules.

Where LLMs falter, on Gibson’s account, is meaning. He cites the rephrased Monty Hall problem: told explicitly that a prize is behind door one with 100% certainty, then asked whether to switch after a different door is opened, an LLM will often still recommend switching — pattern-matching the surface form of the classic (and different) probability puzzle rather than tracking the stated fact. Humans essentially never make this error. Yet in one specific respect LLMs mirror humans closely without being trained to: they struggle to complete deeply center-embedded sentences in the same way people do, which Gibson finds genuinely interesting, since nobody trained them on labelled bad sentences — the failure emerges from the same kind of form-only learning that (Gibson argues) also drives human sentence production.

Piraha, number, and the invention of vocabulary

Working with the linguist-missionary Dan Everett, Gibson studied number words among the Piraha, an Amazonian isolate-language group with no history of contact-driven change. The Piraha have no words for exact counting whatsoever — not even ‘one’. What had been read as a ‘one-two-many’ counting system turned out, on controlled testing, to be three context-dependent approximate quantifiers (‘few’, ‘some’, ‘many’), not counting words at all: shown an ascending sequence of one to ten identical objects, speakers’ word choice depended on where in the sequence they started, not on the exact quantity. Piraha speakers match sets of objects perfectly when the objects stay visible — a task requiring no counting — but fail systematically at exact matching once the objects are hidden from view and counting would be required, typically defaulting to approximate estimates once past four or five items.

Gibson draws the same lesson from English colour vocabulary: the Piraha and the Bolivian Tsimane, both farmer-forager groups Gibson has worked with, use far fewer colour terms than industrialised English (which settles on around eleven), and some non-industrial groups — the Dani of Papua New Guinea — use only two. This is not a perceptual limit; Gibson’s account is functional. Vocabulary tracks what a community actually needs to distinguish for communication (which manufactured, arbitrarily-dyed object do you want?), not the full range of what human vision or exact-counting reasoning is capable of representing. The same logic explains why English and Mandarin thrive globally and minority languages such as Moseten (a close relative of Tsimane, in heavy contact with Spanish) are dying: language usefulness tracks economic and communicative need, though Gibson agrees that language can simultaneously serve as a marker of group identity — Ukrainian in the war with Russia is his example — independent of its purely communicative value.

Noisy channels, translation, and animal communication

Gibson connects word order to Claude Shannon’s information theory — itself first sketched in a 1948 MIT master’s thesis considering language before Shannon moved fully into communication engineering. Word order, on this view, may be partly shaped by robustness to a noisy channel: background noise, the speaker’s own errors, and the listener’s imperfect attention, three distinct sources Gibson keeps analytically separate. On translation, Gibson expects some concepts to resist translation altogether in either direction (English number concepts into Piraha; presumably some Piraha concepts into English) even as most vocabulary maps across languages reasonably well — with harder, open questions remaining about whether the ‘music’ of prose style (Hemingway’s short sentences against a more elaborate stylist) is itself something a measurable, and eventually translatable, quantity.

Asked about interspecies communication — prompted by Lex Fridman’s mention of the Earth Species Project’s work on whale and crow vocalisation — Gibson pushes back on the common assumption that human language is categorically unique. He argues that claims like Chomsky’s, that only human language has full compositionality, rest on a form of question-begging: nobody making the claim actually speaks whale, crow, or monkey, so the judgment that other species only produce ‘grunts and squeaks’ is an assessment from outside a communication system nobody inside the field has learned to read. He does not predict every species has language-like communication, but treats the possibility as open and worth taking seriously for species with large brains and complex social structures.

See also

• Edward Gibson — speaker
• Lex Fridman — host
• Dependency Length Minimisation — the cognitive-cost theory of word order and center embedding developed across this episode
• Language-Thought Dissociation — Fedorenko’s fMRI evidence that language and reasoning are separate brain systems
• Noam Chomsky — Gibson’s MIT colleague and the episode’s recurring point of contrast
• Language and Mind — the wiki’s broader synthesis on language’s relation to thought
• Large Language Models — the wiki’s hub concept on LLM capability and limits