Your brain may store language as LEGO chunks, not grammar trees

Linguistics has assumed since the 1950s that we understand sentences by building hierarchical syntax trees. A study by Morten Christiansen (Cornell) and Yngwie Nielsen (Aarhus), published 21 January 2026 in Nature Human Behaviour, finds that we also store short, linear chunks of word classes. That includes sequences no grammar rule produces, such as "in the middle of the". The authors compare it to snapping together prefab LEGO pieces rather than growing a branching tree.

The evidence is priming. In an eye-tracking experiment and an analysis of recorded phone conversations, people processed a word-class sequence faster the second time they met it, a sign that the sequence is stored in memory. Many of the most common three- and four-word sequences in English turn out to be non-constituents, which tree-based grammar has no unit for. "Traditional rules of grammar cannot capture all of the mental representations of language structure," says Nielsen.

The viral version goes too far. The paper argues trees aren't the whole story. It does not show that hierarchy is fake, and it does not show that brains are just next-token predictors. It also studied English only.

Why it matters: if human language runs partly on flat, chunked patterns, the "LLMs only predict sequences" critique loses some of its force.