How Single Neurons Encode the Grammar and Meaning of Our Words

Language allows us to combine words into an endless variety of phrases and sentences, yet how individual brain cells support this unique human ability has remained a mystery. Now, by recording the activity of hundreds of single neurons while people spoke naturally, scientists have uncovered a cellular-level code for the building blocks of language—from the grammatical role of a word to its place in the hierarchical structure of a sentence. The findings reveal how the brain constructs meaningful speech and map the cortical landscape of language at an unprecedented microscopic scale.

Researchers report the results today in Nature. The work was led by senior author Ziv M. Williams, a neurosurgeon at Massachusetts General Hospital and Harvard Medical School, with first author Jing Cai. The team used tiny implanted electrode arrays to listen in on 579 well-isolated neurons across the frontal and temporal lobes of eight participants who were already undergoing epilepsy monitoring. As the participants spontaneously produced thousands of words and sentences, the scientists tracked each neuron's firing and used natural language processing models to label the linguistic features of every word.

They found that distinct groups of neurons responded selectively to specific linguistic properties. Some cells fired more when a word was a noun or a modal verb; others tracked the sentence’s hierarchical depth, its grammatical dependencies, or how words merged into phrases. About 9% of neurons were tuned to parts of speech, 16% to sentence constituents, and 10% to grammatical dependencies. These proportions far exceeded chance, and the neurons’ selectivity remained consistent across different sentences and contexts.

Importantly, the neurons didn’t just label words in isolation—they combined information. When the researchers built computational models that captured both the syntactic and semantic features of each word, those models could predict the neurons’ activity patterns. Models that also included the surrounding sentence context performed even better, with predictive power peaking about one second before a word was spoken. This suggests that single cells participate in a dynamic, combinatorial process that integrates grammar, meaning, and context long before we utter a sound.

The study also uncovered a nuanced geography of language in the cortex. While language-responsive neurons were found across the left and right frontal and temporal regions, the strength of their tuning was clearly left-lateralized: cells in the left hemisphere showed significantly stronger modulation, especially in prefrontal and posterior temporal areas. This mirrors well-known patterns from brain imaging but now at the resolution of individual neurons.

When the team compared the activity of single neurons with the local field potentials—the summed electrical chatter of surrounding brain tissue—they saw a striking difference. Many electrode sites had language-selective field potentials, but those signals often encoded different features than the neurons at the same spot. Moreover, neurons showed much sharper tuning, with some displaying an extreme degree of selectivity rarely seen in the population-level signals. This finding underscores that crucial linguistic detail may be hidden in the sparse spikes of single cells, not just in broader brain waves.

The recordings came from eight patients and captured natural speech production. Future work will need to examine whether similar cellular codes exist for listening, writing, or scripted communication, and how these patterns interact with other critical eloquent brain areas, such as the left inferior frontal gyrus.

Nevertheless, by mapping the microscopic building blocks of syntax and meaning, the research opens a new window on the neural basis of human language. It offers a prospective approach for studying the cortical organization of language across micro, meso, and macro scales, helping answer fundamental questions about how humans express complex thoughts.

Reference
Cai, J., Kfir, Y., Jamali, M. et al. Mapping the neuronal building blocks of human language with language models. Nature (2026). https://doi.org/10.1038/s41586-026-10691-5