There is no “dog neuron” in your head — “dog” is a direction that hundreds of neurons point to together.
A friend says: “Sorry, I didn’t mean to be mean — I mean, you know what I mean.” Four “means”, and you didn’t mix up a single one. You never looked anything up, and you couldn’t say how you told them apart. So what shape does word meaning actually take inside your head?
The most intuitive guess: each word has its own drawer, and hearing “dog” lights up the dog drawer. In September 2026 a study went looking for those drawers one human neuron at a time — and found none. What it found instead was a map: words with similar meanings land close together, different meanings land far apart, and the same word moves somewhere else on the map when its context changes. The shape of that map looks strikingly like the one inside today’s large language models.
First, where the recordings came from. The hippocampus is a pair of curled structures deep in the brain, famous for memory and for finding your way around. Nobody puts electrodes into a healthy person’s hippocampus. The people who can be recorded are patients with hard-to-treat epilepsy: to find where their seizures start, surgeons implant electrodes for several days, and some patients volunteer for experiments while they wait for a seizure. Hippocampus & entorhinal cortex
In this study, more than a dozen patients listened to real people telling stories while the researchers recorded hundreds of hippocampal neurons. Then they did something slow but solid: first remove whatever the sound of each word and its grammatical role could explain, then ask whether what’s left of the firing can still be predicted by what the word means. It could — robustly.
The interesting part comes next. If you hoped a cell would respond only to “animals”, you’d be disappointed: a single neuron typically responds to several words from several categories that have nothing to do with one another. Stare at any one cell and you can’t say what it is “for”. A unit that does several unrelated jobs at once has a name: mixed selectivity, or polysemanticity.
If one cell can’t be read, stop reading one cell. Line up how strongly each of several hundred neurons fires at the same moment, and that long list of numbers is the word’s “coordinates” in your head. In other words, a word’s meaning doesn’t live in any one cell — it lives in the combined pattern all the cells make together. That’s a population code.
Once you have coordinates, you can measure distance. The researchers found that how far apart two words’ population responses are matches how different their meanings are. “Cat” and “dog” sit next to each other on the map; “rabbit” and “Monday” are far apart. And nobody taught these neurons what an “animal” is — the structure grew on its own.
When a large language model handles text, it doesn’t treat words as bare symbols. It gives each word a long list of numbers — hundreds to thousands of them — called a word vector or embedding. The trick is that words with similar meanings get similar numbers, so you can measure distance in a “space”. The study compared the neurons’ population responses with the embeddings of several models and found the two maps share a shape — cat next to dog in both. One runs on electrical spikes, the other on matrix multiplication; each, groping through a flood of language, grew the same kind of geometry.
Back to all those “means”. If every word had one fixed coordinate, “what does it mean” and “that was mean” would pile onto the same point and you couldn’t tell them apart. In fact, a word’s population response shifts with its context. The researchers used language models to score each word’s “polysemy” — how much its meaning drifts from sentence to sentence — and found that the more polysemous a word, the more its neural response drifted too.
An even more telling comparison: early word vectors (such as word2vec) are static — one word, one point, regardless of context. Later models (GPT-2, BERT) give contextual embeddings: the same word is recomputed in every sentence. Lined up against the hippocampal neurons, the “population distance ≈ meaning distance” rule held for contextual embeddings and actually flipped for static ones. The best match to the neurons was a GPT-2-class model.
People who study what goes on inside AI models ran into “one neuron, a pile of unrelated things” a few years ago, and gave it the same name: polysemantic neurons. A popular explanation is superposition: a model has far more concepts to represent than it has neurons, so each neuron takes part in several concepts at once. A concept is then no longer one slot but a direction — many neurons combined in particular proportions. The hippocampal data fit that picture: a few hundred cells carrying thousands of words and their context-dependent variants simply can’t give each one its own slot.
There’s a second line of evidence from the same period. A study published in Nature in June 2026 recorded single neurons in the frontal and temporal lobes while patients were speaking, then used language models to interpret them: some neurons tracked grammatical relations between words and parts of speech, others tracked a sentence’s overall syntactic structure and how its phrases followed one another, with this ability mostly on the left side of the brain. One study listens, the other speaks; one is in the hippocampus, the other in the language network — and both see the same kind of combinatorial, context-shifting code.
But a few boundaries need drawing. First, the likeness is geometric: two maps sharing a shape doesn’t mean the brain runs the same algorithm as GPT-2. The brain has no step where it reads in a whole passage and computes over it at once, and it learns in a completely different way. Second, the best fit was GPT-2-class — a small, old model by today’s standards; bigger models aren’t automatically more brain-like, which is itself worth chewing on. Third, these are recordings from epilepsy patients’ hippocampi, a few hundred cells in all.
And one deeper question: is the hippocampus storing word meaning, or binding meaning to the here and now? The hippocampus’s day job is tying together “what, where, when” into a single episode. While you listen to a story it may be doing exactly that — stamping each word with “at this point in this story”. If so, no wonder it looks so context-hungry.
“A word has no fixed home; its meaning depends on its relations to other words and on the context of the moment” — philosophers and linguists said this long ago: