New Science in Nature: How Human Brains and Language Models Converge

Ronni Holmvig Strøm · 2026-07-01

The resemblance between how a transformer handles language and how a brain handles language has lived mostly in analogy. Both build context. Both seem to carry structure and meaning in some distributed internal form. The comparison was suggestive, and it was hard to make precise, because the two systems are measured in such different units that lining them up felt more like metaphor than result. A paper published this year in Nature, "Mapping the neuronal building blocks of human language with language models" (Cai et al., 2026), makes the comparison measurable, and it does so at the finest grain available: a single firing neuron.

The recordings came from eight awake epilepsy patients who already had high-density microelectrode arrays implanted for clinical reasons. Across 14 sessions, the team from Massachusetts General Hospital and Harvard Medical School isolated 579 individual cortical neurons and tracked their firing while patients constructed sentences spontaneously, in their own words, without a script. The novel move was the second half. They took those firing patterns and compared them directly to the internal vector embeddings of modern large language models.

Setting the Two Systems Side by Side

Roughly half of the recorded neurons were language-selective, 289 of 579, firing in response to specific linguistic features rather than to sound: parts of speech, syntactic structure, dependency relationships between words, and semantic meaning. The selectivity held independent of acoustics and word frequency, which is the detail that makes the comparison legitimate. These cells tracked what a word meant and how it fit, independent of how it sounded or how common it was.

That isolation matters for the comparison, because an embedding is also acoustics-free. A language model never hears anything. If the neurons had been responding chiefly to sound, there would be nothing for a text-trained model to align with. The fact that they respond to linguistic structure rather than acoustics is what puts the two systems on the same axis in the first place.

When the Embedding Predicted the Cell

The central result is quantitative, and it is the reason the paper carries a model in its title. The embeddings of the language models could predict the firing rates of the neurons. Feed a model the same sentence the patient was constructing, read out its internal representation, and that representation tracked what the cells actually did.

The alignment was not uniform across models, and the pattern in the variation is the interesting part. The models that carried both syntactic and semantic information in their embeddings correlated most strongly with neural activity. Models that captured one without the other fit worse. The brain's cellular-level handling of language looked most like the models that had internalized the most complete picture of language, structure and meaning held together.

This is a link between two systems that share no design history. Transformers did not come from neuroscience. They came from attention mechanisms built to solve a machine translation problem, and the resemblance to cortex was never the goal. Two systems built for the same task by entirely different routes arrived at internal representations that line up well enough that one predicts the other.

What the Single Cell Reveals That the Model Already Carried

One of the older questions in neurolinguistics is whether the brain keeps grammar and meaning separate or processes them as one blended thing. At the scale of regions, the answer has been murky, because population signals average across cells with different jobs and the dissociation washes out.

The single-cell view is sharper. Individual neurons tracked specific grammatical features or specific semantic ones, and they did not collapse into each other at the microscale even though field potentials, the averaged signal, do blur them. Syntax and semantics are dissociable at the level of the firing cell.