The grammar tree might be a myth—or at least, not the whole story.

For decades, the reigning idea in linguistics—the one tied to Noam Chomsky—was that human language lives in the brain as hierarchy. Nested grammatical trees. Deep, abstract rules we supposedly compute before we speak.

A new study says: not so fast. https://www.nature.com/articles/s41562-025-02387-z.

Researchers ran four experiments with nearly 500 English speakers, plus two analyses of real-world language—eye-tracked reading and everyday conversation. What they found: our brains also store simple, flat sequences of word types, like “verb + preposition + determiner,” that cut straight across traditional grammatical boundaries. People could be primed on these linear chunks even when no full grammatical phrase was present—and the effect showed up in natural reading and conversation too.

To be clear: this doesn’t kill hierarchical grammar outright. The paper says these flat structures can’t be explained by traditional constituents—which is a real challenge to major theories, generative and constructionist alike. A challenge, not a funeral.

But it does poke a hole in a familiar story.

Critics have long dismissed AI like ChatGPT as “just a next-token predictor”—not truly intelligent, because it doesn’t grasp deep grammatical hierarchy the way humans supposedly do.

This study doesn’t prove humans are next-token predictors. But it does suggest we lean on simpler, sequential patterns far more than the textbook picture allowed.

Which means the line between “real” human language and machine prediction may be blurrier—and stranger—than either side claimed.

Notes:

This is a single paper, in one language, published in 2026. Cross-linguistic replication is the obvious next step, and until someone runs those experiments, the answer is: plausible, authors expect it, but unproven. English evidence, with the authors predicting broader applicability.

#ArtificialIntelligence #language #NoamChomsky

Language is often thought to be represented through hierarchically structured units. Nielsen and Christiansen find that non-hierarchical structures are present across reaction-time tasks, eye-tracked reading and natural conversation.