Neuroscience · Language & AI
When you hear the word "bank," something in your brain decides which kind you mean. For years, nobody could say which cells were doing the deciding.
Researchers listened in on single neurons while people heard stories, and used AI language models as a ruler for the answer.
Researchers at Baylor College of Medicine, Rice University, UC Berkeley, and Texas Children's Hospital recorded individual hippocampal neurons in 10 people, using electrodes implanted for epilepsy monitoring, while they listened to stories. Using encoding models and numerical word representations from , they found that word meaning is encoded across patterns of activity spread over groups of neurons, even after controlling for sound and grammar. Distances between neural population responses correlated with semantic distance, activity aligned most closely with GPT-2 embeddings, and response variation correlated with LLM-derived measures of polysemy, suggesting meaning coding is largely contextual. Published in Nature Neuroscience, the study is correlational, based on a small sample with epilepsy, and does not claim the brain works like an LLM.
Patterns across groups of hippocampal neurons carried information about what spoken words mean.
Reading Word Meaning in Individual Brain Cells
Researchers at Baylor College of Medicine, Rice University, the University of California, Berkeley, and Texas Children's Hospital have studied how groups of brain cells encode the meaning of words, drawing inspiration from how large language models process language.
Their findings, published in Nature Neuroscience, suggest that the meaning of words is represented through patterns of activity spread across multiple neurons in the hippocampus, a brain region involved in memory.
The data come from recordings of individual neurons in 10 people who listened to stories.
The headline is modest and specific. It is not a claim that the brain works like an AI system. It is a finding about where and how word meaning shows up in neural activity, and a method that borrows from AI to find it.
The rest of this article walks through what was recorded, how it was analyzed, and how much weight the result can bear.
What It Takes to Understand a Spoken Word
Understanding a spoken word involves more than recognizing sounds.
That context-dependence is what makes meaning hard to study. A neuron that responds to a word might be responding to its sound, its grammar, or its meaning, and these are tangled together.
Neuroscientists have long wanted to pull them apart and ask where meaning itself is represented.
Part of the difficulty is that meaning is not a single property. A word can sit near some words in meaning and far from others, and that web of relationships is what the researchers set out to look for in neural activity.
Why Large Language Models Became a Tool for Neuroscience
Large language models, the AI systems behind ChatGPT, Gemini, and similar platforms, have opened new possibilities for studying language.
They represent each word as a list of numbers, often called an embedding, that captures how the word relates to others in meaning and context.
That gives neuroscientists something to compare against. If brain activity recorded during listening lines up with these numerical representations, it may show how both the brain and the models interpret words in different contexts.
The study uses the models in exactly that way, as a measuring stick for neural patterns, which is why the source describes it as inspired by how LLMs process language.
Listening to Stories With Electrodes in the Brain
The team recorded the activity of individual neurons in the hippocampus of 10 people while they listened to stories.
The recordings were made with tiny wire electrodes implanted in the participants' brains as part of clinical monitoring for epilepsy.
That setting is why such data are rare. Recording from single human neurons is only possible when electrodes are already in place for medical reasons.
The researchers then used mathematical models to analyze both the language in the stories and the brain recordings. This let them test whether information about the meaning of words could predict the recorded neuronal activity.
The source notes the participants had epilepsy, which is worth keeping in mind when thinking about how far the results generalize.
Matching Neural Activity to Word Meaning
Using an encoding model approach, Melissa Franch and colleagues found robust encoding of semantics, even after controlling for phonemics and grammar.
In plain terms, the neurons carried information about meaning that could not be explained away by how the words sounded or how they were put together in sentences.
| Question | What the study reports |
|---|---|
| Is meaning encoded in hippocampal neurons? | Yes, robustly, even after controlling for sound and grammar |
| What do single neurons respond to? | Complex selectivities spanning multiple words in multiple semantic categories |
| How do population patterns relate to meaning? | Distance between neural responses correlated with semantic distance |
| Which model matched most closely? | GPT-2 embeddings |
| What about words with several meanings? | Response variation correlated with LLM-derived polysemy measures |
Individual neurons were not tied to one word. Each had complex selectivities that included multiple words and multiple categories, which fits the conclusion that meaning lives in groups of cells.
The researchers also compared the differences between population activity patterns for pairs of words with the differences between those words' meanings.
How Closely the Brain Tracked GPT-2 and Polysemy
Similar to the embedding vectors in LLMs, the distance between neural population responses correlated with semantic distance. Words that are closer in meaning tended to produce more similar population patterns.
There was a subtle exception. For very semantically similar words, even contextual embedding models showed an inverse relationship between semantic and neural distance.
Overall, neural population activity aligned most closely with GPT-2 embeddings. The source does not say which other models were compared, so no ranking is claimed here.
Finally, variation in response patterns correlated with LLM-derived measures of polysemy, the degree to which a word has several related meanings. The authors read this as showing that semantic encoding is largely contextualized.
What the Study Does and Does Not Show
The authors describe the work as a neurocomputational account of how hippocampal neurons encode word meaning during language listening.
The combined activity of groups of hippocampal neurons appeared to carry information about the relationships between word meanings and how those meanings vary with context.
Several limits are worth stating. The sample comprised 10 people with epilepsy; the recordings came from the hippocampus, and the findings were correlations between neural activity and model measures.
Looking ahead, the work could inspire more studies that use LLMs to investigate how the brain processes language. The source describes no clinical application.
Manohar Insight
What I like about this study is how plainly it uses AI. The models are not the object of the research; they are the ruler. They give neuroscientists a numerical way to measure how word meanings relate, and then the brain data are held up against it.
The result I find most interesting is that meaning is not stored in single cells. Individual neurons had complex preferences across many words and categories, and the information lives in the population pattern.
The detail about contextual meaning is the one to watch. If neural patterns track polysemy and context the way LLM measures do, that hints that the brain handles shifting meaning with something like context-dependent coding.
I would keep the caveats in front, though. Ten participants with epilepsy and a single brain region is a narrow base, and closer alignment with GPT-2 does not mean the brain is running something like GPT-2.
The source also notes a model comparison without listing the alternatives, so "closest to GPT-2" is best read as a finding about these particular comparisons.
The real story is not that brains resemble chatbots. It is that language models have become a useful measuring instrument for asking where in the brain the meaning of a word lives, and the hippocampus appears to hold a distributed answer.
Join the Signal
If language models can serve as a ruler for how the brain represents meaning, what other mysteries of the mind might AI help us measure? Share your take in the comments.
Discussion