Gibson, Edward "Ted"

psycholinguistics professor, MIT; head of the MIT Language Lab; author of Syntax: A Cognitive Approach

Edward Gibson: Human Language, Psycholinguistics, Syntax, Grammar & LLMs

Lex Fridman Podcast · Apr 17, 2024

• Edward Gibson
• psycholinguistics
• dependency grammar
• Noam Chomsky
• center embedding
• large language models and form vs meaning
• language and thought fMRI
• legalese
• Piraha and number words
• language death and WEIRD bias

Core ideas

Gibson, who came to linguistics from math and computer science rather than the humanities, frames his life's work around a sharp distinction between the form of language (its structure, word order, grammar) and its meaning — a distinction he considers wide and consequential, in contrast to Lex's more romantic intuition that form is simply "the shadow" of meaning. His central theoretical claim is that human sentences are best understood via dependency grammar: every word connects to exactly one other word, producing a tree rooted typically at the verb, and this framework makes visible a near-universal law — languages minimize the physical/linear distance between dependent words, because long-distance dependencies are costly to produce and to comprehend. He credits an ex-student, Richard Futrell, with demonstrating this empirically across roughly 60 languages with parsed corpora: real sentences have systematically shorter dependency lengths than randomly scrambled (even projectivity-constrained) versions of the same sentences, regardless of whether a language is verb-initial (like English, with prepositions) or verb-final (like Japanese or Hindi, with postpositions) — the two dominant word-order families (~40-50% each) that are internally "harmonic" specifically because that harmony keeps dependencies short. This dependency-length-minimization theory also explains center embedding (nesting a clause between a subject and its verb, e.g. "The boy, who the cat, which the dog chased, scratched, cried"): nesting multiple such clauses always lengthens several dependencies simultaneously and reliably breaks comprehension and production in any human language, which Gibson demonstrates via completion tasks (people asked to finish partial center-embedded sentences reliably drop required verbs) and via fMRI showing the language network's activation scales directly with dependency length.

He contrasts this dependency-grammar, corpus-and-experiment-driven approach directly with Noam Chomsky's phrase-structure grammar and especially with Chomsky's concept of syntactic "movement" (e.g., deriving "Will two dogs enter the room?" by moving the auxiliary "will" from a hypothesized declarative deep structure to the front in a surface structure). Gibson argues the alternative "lexical copying" account — where "will" simply has two related lexical entries, declarative and interrogative, with different argument structures and no movement at all — is both more learnable (Chomsky himself showed in 1971 that movement grammars create serious learnability problems, which is precisely why he concluded such structure must be innate/universal grammar) and better fits messy real usage, such as auxiliaries that only work in one form ("Aren't I invited?" has no declarative "I aren't"). He identifies methodology, not just theory, as the core difference between himself and Chomsky: Gibson insists on quantitative experiments and corpus data, whereas Chomsky's entire body of work rests on personal linguistic intuition with no experiments or corpus analysis ever attached to it.

A major thread concerns the relationship between language and thought, built on fMRI work by Gibson's wife and colleague Evelina Fedorenko. Her lab can localize an individual's "language network" (a stable, left-lateralized region, unchanged in Gibson's own brain across 15+ years of repeat scanning) using a simple contrast (real language vs. non-language/gibberish/backward speech), and critically, that network activates identically whether processing spoken or written language, in any language a person knows fluently, and even in constructed languages like Klingon — but it does not activate for chess, spatial reasoning, music (unless it contains words), or, strikingly, computer programming, even in expert programmers. Complementary evidence comes from "global aphasia" patients whose stroke has destroyed the language network but who can still do math, drive, and play chess perfectly normally — language and thought are dissociable, direct evidence against Chomsky's view (which Gibson states plainly) that language is the substrate of thought. Gibson personally lacks an "inner voice" entirely (which he initially assumed was a metaphor others used, until he learned most people, roughly 70-80%, report a literal auditory inner monologue), which he raises as a further open puzzle about how language relates to conscious thought.

On large language models, Gibson calls them "arguably the best current theories of human language" specifically because they are unmatched at predicting grammaticality judgments — but stresses this is a theory of form, not meaning: LLMs can be trivially tricked on meaning-dependent reasoning (e.g., failing a modified Monty Hall problem where the car's location is already stated with 100% certainty, because the model pattern-matches to the surface form of the classic version rather than tracking the actual state of knowledge) in ways humans reliably do not. Yet strikingly, LLMs fail at completing center-embedded sentences in essentially the same way and to the same degree as humans do, unprompted by any explicit training on "bad" sentences — which Gibson finds genuinely remarkable evidence that LLMs may be implicitly modeling something like the same memory-based processing constraints (dependency-length costs) that govern human sentence processing, even though their overall architecture has nothing designed to do that.

He describes a large applied research program on legalese conducted with former student and lawyer Eric Martinez: analyzing real contracts, they found legal text is extraordinarily center-embedded (~70% of sentences contain a center-embedded clause, versus ~20-30% even in dense academic writing) and uses unusually low-frequency words, but — surprisingly — passive voice, despite its bad reputation, makes zero measurable difference to comprehension or recall. Center embedding, however, devastates comprehension and recall for laypeople and for practicing lawyers (tested across 100 lawyers) equally, and lawyers themselves, when surveyed, strongly prefer simplified, non-center-embedded versions of the same content and don't want to write this way — undermining any simple "lawyers deliberately obscure meaning to create demand for themselves" explanation. Gibson's tentative alternative is a "magic spell hypothesis": center embedding may function as a performative marker signaling "this is authoritative legal language," much as rhyme signals "this is a magic spell," even though no one consciously intends or wants the resulting difficulty — though he and Lex explore, without resolving, whether this could still be an emergent, self-reinforcing systemic incentive rather than a stylistic accident.

On cross-linguistic and cross-cultural variation, Gibson discusses years of fieldwork with two Amazonian language isolates, the Tsimane (Bolivia) and the Pirahã (Brazil, studied via linguist/former missionary Dan Everett), arguing that vocabulary — color terms, number words — emerges from communicative need rather than perceptual capacity: industrialized English has ~11 basic color terms because manufactured objects are arbitrarily colored and need distinguishing, whereas cultures like the Pirahã have no exact number words at all (not even "one"), because their quantifier words ("hei," "hoi," etc.) turn out under careful elicitation to mean "few/some/many" on a relative, context-dependent scale rather than exact counts — confirmed by matching-task experiments showing Pirahã speakers perform perfectly on direct one-to-one object correspondence (no counting needed) but only approximately once a screen hides the objects and exact enumeration would be required. He connects this to the WEIRD (Western, Educated, Industrialized, Rich, Democratic) sampling bias problem in psychology and linguistics, and to language death more broadly (illustrated by the closely related Moseten language dying out in favor of Spanish) as fundamentally an economic phenomenon: languages persist because they carry functional/economic value to a community, which is also why global languages like English, Mandarin, and Spanish spread, even as language simultaneously serves as a marker of cultural and national identity (he cites the war in Ukraine as an example of language-as-identity cutting against pure communicative-efficiency logic). He closes optimistic that human language may not be as categorically unique as commonly assumed, criticizing arguments (including Chomsky's) that dismiss animal communication as mere "grunts and squeaks" as reasoning that would look equally dismissive from a whale's perspective, and expressing genuine openness to ongoing efforts (e.g., Earth Species Project) to find structure in whale, crow, or other animal communication systems.

Excerpts