Saturday 07 November 2026About PsychregWrite for Us

Mental Health

Is Language Hierarchical? What a Viral Study Says About Chomsky and How Your Brain Stores Sentences

by Dennis Relojo-Howell · Updated

language analysis

Is language hierarchical? Since the 1950s, the standard answer in linguistics has been yes: words group into phrases, phrases group into larger phrases, and a sentence is a tree. A study published in January 2026 asked whether that is the only kind of structure the mind keeps, and its authors say it is not. A headline claiming that Chomsky was wrong would overstate what it found.

This is not a result from this week. The paper by Yngwie Nielsen of Aarhus University and Morten Christiansen of Cornell University appeared in Nature Human Behaviour on 21st January 2026, and this article draws on its published abstract, the authors’ own research briefing, and Cornell’s news release. The details below rely on those summaries rather than the full text of the paper.

Is language hierarchical, according to the standard view?

A hierarchical view says that words combine into units called constituents. In the example given in the Cornell Chronicle, “the” and “cake” combine into the noun phrase “the cake”, which combines with “ate” into the verb phrase “ate the cake”, and then with “she” to make the sentence. The picture is a branching tree, with phrases nested inside phrases.

The idea is most closely linked with Noam Chomsky, whose 1957 book “Syntactic Structures” became a cornerstone of modern linguistics. It has been challenged before. In 2012, Stefan Frank, Rens Bod, and Morten Christiansen reviewed evidence in a paper asking “How hierarchical is language use?” and argued that sequential structure may be fundamental to how we understand, produce, and learn language.

What did Nielsen and Christiansen test?

The team used structural priming, a method in which seeing or hearing a structure makes it slightly easier to process the same structure soon after. Psychologists generally assume that if a structure can be primed, it is part of what a person knows about language, and this logic has been used before to support constituents.

In four preregistered experiments with 497 participants in total, people saw a string of words and judged as fast as they could whether it could be used in an English sentence, such as “wondered if you”, or not, such as “it your for”. Some consecutive trials shared a sequence of word types that is not a constituent, like “knew that she” followed by “wondered if you”. The abstract’s own example is verb, preposition, determiner, shared by “added to a” and “defined by the”.

Priming sped up responses by about 30 to 40 milliseconds, according to the authors’ briefing. In a replication with 200 people, the effect for these non-constituent chunks was comparable to, or possibly larger than, the effect for ordinary constituents.

To see whether this holds outside the lab, the authors analysed two existing sets of data. In eye-tracking data, people (N = 68) read newswire sentences faster when they contained a non-constituent structure they had recently met. In transcripts of telephone conversations (N = 358), speakers were more likely to use such a structure after hearing their conversation partner use it.

Why “Chomsky was wrong” overstates it

The first problem is that the study shows something is missing, not that something is wrong. The abstract says the findings are “not explainable in terms of constituents as traditionally construed” and pose a challenge for accounts of linguistic representation. In Cornell’s release, Nielsen put it more narrowly: the main contribution is showing that traditional rules of grammar cannot capture all of the mental representations of language structure.

The second problem is that the experiments show that flat chunks exist and can be primed. They were not designed to show that nested phrase structure is absent, and the authors found that chunks and constituents were both primed. A structure being found does not mean the other one is gone.

The third problem is that the target is wider than Chomsky. The abstract says the result challenges generative approaches and constructionist approaches, which are two broad families of theory. Generative theories have also developed a great deal since 1957, so one experiment cannot speak for all of them.

There is a fair counter-point. The authors say their findings “raise the possibility” that language may not involve traditional constituent structure at all, and Christiansen told Cornell it might even be possible to account for how we use language in general with flatter structure. That is a hypothesis for future work, not something these experiments showed.

Outside experts have so far been broadly positive about the evidence. In the briefing, Edward Gibson of the Massachusetts Institute of Technology is quoted as calling it a fascinating paper and finding the reading-time and corpus analyses persuasive. A further limit is that the experiments were in English, and the authors note that languages differ widely in how they build structure, although they think the results are likely to generalise beyond English.

What do large language models have to do with it?

The authors say their results are consistent with demonstrations of structural priming in large language models such as ChatGPT, which rely on contextual representations rather than constituent structures rooted in grammar. That is a cautious link, not a claim that people and chatbots work in the same way.

Other researchers urge care here. Christine Cuskley, Rebecca Woods, and Molly Flaherty argue in a 2024 paper in Open Mind that any similarities between human language and the output of language models are functional, and that these models are tools, not theories of language. For a different angle on language in the brain, see this report on a small 2025 study of brain activity during real-world conversations.

So how does your brain store sentences?

The most careful reading is that people use both ready-made chunks and grammar. The authors point out that the most common three- or four-word sequences in language, such as “can I have a” or “it was in the”, are often not constituents, yet they seem to be part of what speakers know. Cornell compares this to snapping together pre-assembled LEGO pieces, rather than building every sentence from scratch.

The authors propose that these chunks may act as a kind of processing glue, used by speakers and listeners to connect meaningful sequences on the fly. That is a proposal for future research, and it does not mean grammar has gone: you can still understand a sentence you have never heard before. Language learning starts early, too, and a report on a 2025 study of infants found that babies as young as four months can abstract features of speech, which challenges older timelines for language acquisition.

What to watch for next

The next tests are replications by other teams, studies in languages with very different word-building, and work on what the chunks are for. Until then, the cautious reading is that language has more than one kind of structure, and that the balance between them is still being worked out.

FAQs

  • Is language hierarchical? In part. Sentences have nested phrase structure, as in “she ate the cake”, but a 2026 study found that people also store flat chunks of word types that this nesting cannot explain.
  • Did this study prove Chomsky was wrong? No. It challenges the assumption that mental representations of language are made only of grammatical constituents, but it did not test or show that nested structure is absent, and the authors say it is a challenge to constructionist theories as well.
  • What is a non-constituent? It is a run of words that does not form a complete grammatical unit, such as “can I have a” or “it was in the”. Such sequences are very common in speech and writing.
  • What is structural priming? It is the tendency to process a structure a little faster after you have just seen, heard, or said the same structure. Researchers use it to infer what people mentally represent about language.
  • Does this mean AI chatbots think like people? No. The authors say their results are consistent with priming effects in language models, which use context rather than grammar trees, but other researchers stress that language models are tools, not theories of human language.
  • Is the study new? It was published online on 21st January 2026 and appeared in the March 2026 issue of the journal, so it is not new this week.

Takeaway

Language still has nested phrase structure, but this study suggests it is not the only kind of structure: people also seem to store short, flat chunks of word types that grammar trees cannot explain. A “Chomsky was wrong” headline goes beyond the evidence, which shows chunks and constituents side by side, in English, in one set of experiments. The authors raise the possibility that language could work without constituents at all, and the next replications will show whether that holds.