Scientists recorded individual neurons in bilingual brains for the first time and found that the brain does not translate words, it does something more sophisticated
If you speak two languages, you already know that switching between them feels almost automatic. You hear a word in Spanish, you understand it. You think in English, you speak in French. The process feels seamless, and for a long time neuroscientists assumed they understood roughly how it worked: the brain must have some version of a mental dictionary, where equivalent words in different languages share the same neural real estate, connected by overlapping cells that respond to the same meaning regardless of which language carries it.
A new study published in the journal Cell has shown that this model is wrong. And what replaces it is considerably more interesting.
Researchers at Baylor College of Medicine and Rice University recorded the activity of individual neurons in the hippocampi of four bilingual people while they listened to, read, and conversed in both English and Spanish. It was the first time scientists had ever studied how the bilingual brain handles language at the level of single neurons, watching in real time as actual cells responded to actual words in two languages simultaneously.
What they found overturns one of the most intuitive assumptions in the field.
The brain does not use shared neurons to connect equivalent words
The logical prediction going into the study was that certain hippocampal neurons would light up for both “dog” and “perro,” firing equally in response to the same concept regardless of which language delivered it. These would be the bridge cells, the translators, the neurons that tell the brain that two different sounds carry the same meaning.
Almost none of the neurons worked that way.
When the researchers analyzed the data, they found that the overwhelming majority of hippocampal neurons were language-specific. A neuron that responded to “dog” in English did not respond to “perro” in Spanish, even though both words refer to the same animal. Individual neurons appeared to be tuned to words in one language and largely indifferent to their equivalents in the other.
“This is the very first study to look at how bilingual brains work at the level of individual neurons, and to do so in real time,” said first author Xinyuan Yan, a researcher at Baylor who is bilingual herself.
What the brain uses instead: a shared map
The absence of translation neurons did not mean the brain had no system for connecting the two languages. It had a more elegant one.
Rather than linking individual words through shared cells, the brain organized all the words in each language onto a geometric map. On this map, words were positioned according to their meaning relative to other words. “Dog” and “wolf” sat close together because they are semantically related animals. “Fork” was far from both of them because it belongs to an entirely different conceptual category. The spatial relationships between words on the map reflected the actual relationships between their meanings.
The critical finding was that this map had the same geometric structure in both languages. “Dog” in the English map occupied a position relative to its neighboring concepts that mirrored the position of “perro” in the Spanish map. The concepts were arranged identically. Only the neurons used to read those positions were different.
“It’s like looking into a room from a different window. Everything inside is the same, but the perspective is different,” said senior author Sameer Sheth of Baylor College of Medicine. The brain reads the same conceptual room through language-specific neurons that each provide their own viewing angle.
Why this solves the problem translation neurons could not
A translation-neuron system would create a practical problem: interference. If the same cells responded to “dog” and “perro” identically, the brain would need additional machinery to track which language it was currently operating in. Keeping two languages clearly distinct while simultaneously knowing they refer to the same things would require constant disambiguation.
The shared-geometry system solves this without any extra work. Because the neurons for English and Spanish are different, the two languages never compete for the same cellular resources. But because the map they read from has the same geometric structure, the meaning carries over perfectly. The brain gets both separation and translation simultaneously, built into the same architecture.
The finding extends to how the brain learns new languages
Senior author Benjamin Hayden drew a direct practical implication from the result. If the brain stores meaning as a geometric map of relationships between concepts, and that map can be read by neurons operating in any language, then learning a second language does not require building an entirely new conceptual system. It requires learning a new set of neurons that can read the map from a different angle.
“Our study shows that the brain is wired to learn multiple languages,” Hayden said. “Once it maps relationships among words, it can apply those same relationships across languages. We all have the potential to become bilingual, or even trilingual.”
This reframes what language learning actually involves cognitively. The hard part of acquiring a first language is building the semantic map itself, establishing all the spatial relationships between thousands of concepts. A second language, on this account, requires building new reading neurons but can largely reuse the existing map. It is less like learning an entirely new system and more like learning a new way to navigate one you already have.
The same architecture appears in AI language models
The researchers made one additional comparison that adds an unexpected dimension to the finding. They analyzed mBERT, a large language model trained to understand more than 100 languages, and found that it organizes words across languages using the same kind of shared geometric structure that the human hippocampus uses.
The AI model was not designed with any knowledge of how the human brain handles bilingualism. It arrived at the same solution through training on language data alone. The convergence suggests that the shared-geometry approach to multilingual representation may be a deep solution to the problem of handling multiple languages in a single system, one that emerges naturally in both biological and artificial neural networks when they are exposed to enough language across multiple tongues.
The researchers note that their study involved only four participants, all of whom were balanced English-Spanish bilinguals who happened to be undergoing treatment for epilepsy with implanted electrodes. The rarity of access to that kind of direct neural recording in living humans means the sample size cannot be larger under current ethical constraints. Whether the same geometric organization holds for people who learned a second language later in life, or for languages that are more structurally different from each other than English and Spanish, remains to be tested.
What the study establishes for the first time is that the neuron-level architecture of bilingual language processing is not what the field assumed. The brain does not translate. It reads the same map from two different windows, and the map was there before either language arrived.
Source
Xinyuan Yan, Ana G. Chavez, Melissa Franch, et al. “Shared neural geometries for bilingual semantic representations in human hippocampal neurons.” Cell, June 24, 2026.
DOI: 10.1016/j.cell.2026.05.020
https://news.rice.edu/news/2026/how-do-bilingual-brains-navigate-between-languages-scientists-discover-geometric-neural
