Edward Gibson: Human Language, Psycholinguistics, Syntax, Grammar & LLMs | Lex Fridman Podcast #426

Across ~60 typologically diverse languages with parsed corpora, actual sentences consistently have much shorter word-to-word dependency distances than randomized but grammatically plausible alternatives, indicating a universal pressure to keep related words close for easier production and comprehension.

Nested structures like “The boy who the cat that the dog chased scratched cried” massively increase dependency distances and working-memory load; both humans and large language models struggle to complete or process such sentences, suggesting shared constraints on processing form.

Corpus and behavioral studies on contracts show unusually high rates of center-embedded clauses (e.g., definitions wedged between subject and verb), which severely hurt comprehension and recall for both laypeople and lawyers; low-frequency vocabulary matters somewhat, while passive voice has negligible effect.

fMRI work (Fedorenko et al.) finds a stable, left-lateralized “language network” activated by sentences (spoken or written) but not by math, music, programming, or other demanding cognitive tasks, while patients with severe aphasia can still reason, play chess, and do arithmetic—showing that high-level thinking doesn’t require language.

Groups like the Tsimane and Pirahã see the same colors and numerosities we do but have far fewer basic color terms and lack exact number words (even for ‘one’); experiments show they use approximate quantifiers (‘few/some/many’) and can match sets perceptually but can’t perform exact counting tasks, highlighting that lexical systems track what must be talked about, not what can be perceived.

Chomsky’s classic movement analyses (e.g., moving auxiliaries to form questions) create learnability problems and require innate “universal grammar” assumptions, whereas alternative accounts with word-specific argument-structure patterns (lexical copying) and dependency trees are simpler, more compatible with data irregularities, and more learnable from input.

They approximate construction-grammar-like patterns and even internally recover something akin to dependency structure, yet fail on small semantic twists (e.g., Monty Hall variants where probabilities are explicitly given), indicating that high-quality surface prediction doesn’t equate to human-like understanding or reasoning about underlying situations.