Human Language: Psycholinguistics, Syntax, Grammar & LLMs

High-Level Overview

Lex Fridman talks with Edward "Ted" Gibson, an MIT professor of psycholinguistics who heads the MIT Language Lab and whose book Syntax: A Cognitive Approach is due from MIT Press. Gibson explains his data-driven view of why human languages look the way they do: sentences form dependency trees, and every language studied so far keeps related words close together to reduce the memory cost of producing and understanding them. From there the conversation covers his disagreements with Noam Chomsky over movement and innate grammar, brain-imaging evidence that language is separate from thought, what large language models do and don't understand, why legal contracts are so hard to read, Claude Shannon's noisy channel, and fieldwork with Amazonian groups whose languages lack words for exact numbers or most colors.

The tone is curious, technical and good-humored. Gibson, trained in math and computer science, approaches language like an engineer and is refreshingly willing to say "we don't know," while Lex keeps pushing on the relationship between form and meaning and is repeatedly surprised, especially when Gibson reveals that he has no inner voice and that thinking doesn't happen in the brain's language network. The conversation flows from linguistic theory to neuroscience, AI, law, anthropology and the future of human and animal communication.

The Discussion Explores:

• Word-order universals: why verb-first languages use prepositions and verb-final languages use postpositions, and the idea of minimizing dependency length.
• Dependency grammar versus Chomsky: trees of word-to-word connections, phrase structure, the "movement" debate and lexical copying.
• Experiments versus intuitions: Gibson's quantitative methods compared to Chomsky's thought experiments.
• Center embedding and cognitive cost: why nested sentences are hard in every language, and how that shows up in the brain.
• Language is not thought: fMRI evidence of a dedicated language network and patients who think without language.
• Large language models: excellent at form, shaky on meaning, and possibly the best current theory of language.
• Legalese: contracts packed with center-embedded clauses, and the "magic spell" hypothesis.
• Language, culture and communication: color and number words among the Tsimane' and Pirahã, the noisy channel, language death and talking to whales.

Deep Dive into Main Ideas

From Math Geek to Language Puzzles

Gibson was drawn to language as a child confused by grammar lessons, seeing sentences as puzzles. A math and computer science undergraduate at Queen's University, he went to Cambridge for a one-year master's in computational linguistics without having taken a single language class, and ended up spending his life on it. An early AI course's natural language section struck him as "a bunch of hacks" rather than a real theory. He deliberately chose to study form (how words combine) rather than meaning, because form seemed more tractable. Lex, who sees form as a shadow of deeper meaning, pushes back on how separable the two are, a thread that runs through the whole conversation.

What he finds most beautiful about language are generalizations across languages. About 40% of languages are subject-verb-object like English, and those tend to have prepositions ("talk about a book"); roughly 45% to 50% are verb-final like Japanese or Hindi, and those tend to have postpositions, putting the marker after the noun and the verb at the end. Of about 1,000 languages with good word-order data, out of some 7,000 spoken today, about 95% follow this harmonic pattern, first observed by Stanford typologist Joseph Greenberg. Gibson's explanation is that both consistent patterns keep connected words closer together than mixed patterns would.

Dependency Grammar: Every Sentence Is a Tree

Gibson distinguishes three layers of language: sounds, words (form paired with meaning) and syntax, the combinations of words that produce compositional meaning. In any sentence, each word connects to exactly one other word, forming a tree, which he says virtually all linguists accept; they disagree only about the details. In "Two dogs entered a room," the root is the verb "entered," which connects to "dogs" (which connects to "two") and to "room" (which connects to "a"). This dependency approach goes back to French linguist Lucien Tesnière and even to Pāṇini in ancient India. A word's part of speech is defined by how it's used rather than by its meaning.

He explains morphemes, the smallest units of meaning, and morphology, how endings or prefixes attach to roots. English has only one or two morphemes per word, while Finnish can pile up ten, yielding millions of forms of a single word. Russian's case endings let words move around freely, which may help explain its rich poetic tradition, while English relies on rigid word order to show who did what to whom. High-frequency English verbs tend to be irregular ("drink, drank), while rarer ones are regular (walk, walked"). How such systems evolve is largely unknown, he says, because we have mostly snapshots of current languages, and contact between groups constantly reshapes them; English itself shifted from a German-like verb-final language with case marking to its current form, partly through contact with French. Even Queen Elizabeth's vowels shifted measurably over her reign.

Chomsky, Movement and Lexical Copying

Gibson describes Chomsky's phrase structure grammar, with rules like a sentence expanding into a noun phrase plus a verb phrase and his formal language hierarchy, from regular languages (too simple for human language, lacking center-embedded recursion) to context-free and context-sensitive ones. Phrase structure and dependency grammars can generate the same sentences; Gibson prefers dependency grammar because it makes the distance between connected words transparent.

The real disagreement is "movement." Chomsky saw that questions like "Will two dogs enter the room? put the auxiliary verb in front, and theorized that words move from an underlying deep structure" to a "surface structure. In 1971, Gibson says, Chomsky showed that grammars with phrase structure plus movement would be impossible for children to learn from data alone, and concluded that grammar must be innate: universal grammar. The alternative Gibson favors, developed by linguists like the late Ivan Sag, is lexical copying: auxiliaries simply have two stored forms, declarative and interrogative, with no movement. This makes language learnable bottom-up and explains quirks, such as Aren't I invited?" existing without "I aren't invited. Gibson thinks Chomsky wanted language to be simple and never cared much about words, even though a word like take" has dozens of senses. In the brain, Gibson notes, words and combinations of words activate the same language network.

The deeper divide is method. Gibson has been at MIT since 1993 and describes himself as a data-driven psychologist who runs experiments and analyzes corpora, whereas Chomsky, he says, has never been associated with an experiment or corpus study and relies on intuitions. Gibson says he evaluates theories quantitatively and Chomsky doesn't.

Center Embedding and the Cost of Distance

The core finding of Gibson's work is that the farther apart two connected words are, the harder a sentence is to produce and understand. "The boy cried" is easy. "The boy who the cat scratched cried" is still fine. But "the boy who the cat which the dog chased scratched ran away" becomes nearly impossible, because nesting clauses inside clauses, center embedding, stretches several dependencies at once. He says this is uninterpretable in every language, not just English. In one experiment, when people are asked to complete "The book which the author who..." they leave out one of the three needed verbs about 60% of the time.

His former student Richard Futrell tested this across about 60 languages with parsed text collections, comparing real sentences to randomly scrambled versions (including scrambles that avoid crossed dependencies). In every language, the real dependencies were shorter, whether the language was verb-initial, verb-medial or verb-final. Gibson believes this reflects ease of production first: it's easier for the speaker to keep connected words close, which happens to help listeners too. Lex argues language is optimized for the listener; Gibson clarifies he's holding meaning constant and asking only how its form is chosen. Cognitive cost can be measured through acceptability ratings, reading times and, in one study, stronger language-network activation for longer dependencies. Gibson suspects cost grows in a way related to memory retrieval and interference, and wishes more researchers would pursue an exact formula.

Language Is Not Thought

Gibson's wife, MIT neuroscientist Ev Fedorenko, can locate a person's language network with fMRI in about 15 minutes by contrasting understandable language with things like backward speech or nonwords. It is a left-lateralized network that responds the same to spoken and written language, to comprehension and production, to every language a person speaks fluently, and even to constructed languages like Klingon. Gibson's own network has looked the same in scans from 2007 to today: "my language network is like my face." Crucially, tasks such as spatial memory, music, math and reading computer code do not activate it at all; they use other areas, including a "multiple demands" network for hard tasks.

A second line of evidence comes from Rosemary Varley's patients at University College London with global aphasia, whose strokes wiped out their language networks. They cannot speak, understand, read or write, yet they can play chess, drive, do arithmetic and perform as well as age-matched controls. Together, Gibson says, these findings suggest language isn't necessary for thinking: "language is a communication system for the meaning. It's not the meaning." You can think in language, but you don't need it. Along the way, Gibson reveals that he has no inner voice and assumed for years that the idea was a metaphor, while surveys suggest most people report one. Lex says he vocalizes when reading but doesn't hear a voice and refuses to believe most people do. They agree this may challenge Chomsky's view that language is the foundation of thought.

What Large Language Models Get Right and Wrong

Gibson has written that LLMs are arguably the best current theories of human language, because they predict what is and isn't good English better than any other theory, though they are too large and opaque to be satisfying theories. Chris Manning and colleagues at Stanford found evidence that models like BERT build something resembling dependency structure internally. LLMs are closest to construction-based theories, which pair forms with meanings in a usage-based way, and Gibson thinks dependency grammar might be the right formalization of those.

On understanding, Gibson says: "They're doing the form, they're doing it really, really well." Meaning is another matter. He cites a modified Monty Hall problem: if you tell a model you know the car is behind door one and the host opens door three, it still tells you to switch, because it follows the familiar form of the puzzle. Humans rarely make that error. Strikingly, LLMs struggle with center-embedded sentences just as humans do, without being trained to, which Gibson finds fascinating and suggests they may be modeling how humans process form. Lex argues that anecdotes aren't systematic evidence and that form alone gets surprisingly close to apparent understanding, partly because truth may be overrepresented online. Gibson notes that Fedorenko's group, with first author Kyle Mahowald, has argued that LLMs lack the thinking and reasoning that the brain handles separately from language.

Why Legalese Is So Hard to Read

Legal language is the one exception Gibson has found to short dependencies. Working with Eric Martinez, then a Harvard law student in his class, he analyzed contracts and found that about 70% of sentences contain a center-embedded clause, compared with about 20% to 30% in other texts, including academic writing. Legal text also uses more low-frequency words and more passive voice. In experiments that rewrote passages to remove each feature, passive voice made no difference, rare words hurt a little, and center embedding hurt a lot, slowing reading and badly reducing comprehension and recall. He reads an example in which the definition of "total payments" is dropped between a subject and its verb, and suggests simply putting definitions elsewhere.

Surprisingly, the same pattern held for about 100 lawyers: they understood legalese better than laypeople but showed the same penalty, preferred plain versions, and said they'd rather hire someone who writes plainly. So why does it persist? Gibson's leading idea is the "magic spell hypothesis": center embedding may be a performative style that signals legal text the way rhyme signals a magic spell. Lex suspects incentives, since incomprehensible contracts make people hire lawyers; Gibson says they found no evidence for that but remains suspicious too. He is hopeful, because unembedding is easy, and notes that presidents from Nixon to Obama issued plain-language directives without knowing what actually to fix.

The Noisy Channel and How Languages Evolve

Gibson describes Claude Shannon's work on communication, which Gibson says began with a master's thesis and included examples very much like language, and notes that Shannon was interested in linguistics but the field moved in Chomsky's direction. Communication happens over a noisy channel: background noise, speaker errors and listener limitations. Listeners often infer what a speaker meant despite errors, and some aspects of word order may be optimized for robustness to noise, though Gibson admits these are "just so stories" compared with his solid dependency results.

Asked whether languages are optimized for learning, Gibson says every language is a noisy solution to the problem of communication, and that no language is harder for children: every baby learns their native language by about three or four. But regular word-order rules probably exist to make learning easier, since pure dependency minimization would allow much freer word order. On innateness, Gibson thinks much of language is learned and that LLMs support this. A dedicated brain module doesn't prove innateness: the visual word form area develops only in people who learn to read, and people who lose left-hemisphere regions early in life can develop language on the right. Fedorenko studies such "interesting brains," many of whom contacted her after a New York Times article.

Colors, Numbers and Languages at the Edge

Following Joseph Henrich's warning about generalizing from "WEIRD populations, Gibson works with isolated Amazonian groups whose languages have no known relatives. I think language is invented," he says: people create words for what they need to communicate. English speakers share about 11 basic color words, while the Dani of Papua New Guinea have two, roughly black and white, and the Tsimane' of Bolivia have three that everyone knows (black, white and red) plus a few others. They see colors perfectly well and enjoy bright objects, but in a world without identical dyed items, they rarely need to name colors to distinguish things.

With linguist Dan Everett and then-student Mike Frank, Gibson studied the Pirahã of Brazil, whose language has no word even for one. Their three quantity words, once thought to mean one, two and many, actually mean roughly "few, some and many," and depend on context: counting down from ten, the word for few appeared around four. Pirahã speakers matched one-to-one sets perfectly, but when a set was hidden they could only approximate beyond about four or five, suggesting exact number words are needed for exact tasks. Anthropologists guess counting systems arose with farming, to track interchangeable animals; people don't need numbers to track their children, Gibson jokes, because they care which children, not how many.

On why universal languages like Esperanto haven't caught on, Gibson says languages survive by being useful, and economics drives which ones spread. Mosetén, a close relative of Tsimane', is dying because Spanish is what feeds families. Lex adds that language is also identity, citing Ukrainian during the war, and Gibson agrees, hoping languages survive. On translation, he says form can probably be mapped but concepts can't always: you can't translate much English into Pirahã. He suspects Hemingway's prose has short average dependencies.

Talking to Whales and Advice for a Curious Life

Lex mentions Aza Raskin's Earth Species Project, which seeks common structure between whale, crow and human communication. Gibson finds it quite reasonable and criticizes the claim that human language is uniquely special: dismissing whale or crow sounds as grunts is "bad reasoning," since we don't speak whale, and a whale might say the same of us. Lex adds Neri Oxman's work on plant communication. If he met aliens, Gibson says he'd bring Dan Everett, whose gift for fieldwork comes from being deeply social; he describes Everett's method of throwing down a stick, then two sticks, and asking for the words.

His advice for young people: "when you see something interesting just go and do it." When he couldn't return to the Pirahã because of problems with Brazilian authorities, he searched for another group and found the Tsimane'. Just try things other people haven't done, he says. Lex closes with Ludwig Wittgenstein: "the limits of my language mean the limits of my world."

Conclusion

Ted Gibson offers a strikingly simple, testable account of human language: sentences are trees of word-to-word dependencies, and speakers in every language keep those dependencies short because long, nested connections are costly for memory. That single idea explains word-order universals, why center-embedded sentences are unreadable, and why legal contracts are so painful, and it gives legislators a clear fix. His broader picture treats language as an invented, learned tool for communication rather than an innate foundation of thought, backed by brain scans showing a language network separate from reasoning and by aphasia patients who think without words. That view shapes his take on AI, too: large language models have mastered form, even sharing our struggles with nested sentences, but meaning is another matter. From the Pirahã's missing numbers to the hope of talking with whales, the conversation shows how much language reveals, and hides, about the minds that use it.

This is a summary of "Human Language: Psycholinguistics, Syntax, Grammar & LLMs" from Lex Fridman Podcast, giving you the GIST so you can decide if you want the REST. Listen to the full episode.