How Bilingual Brains Store Languages

Introduction

Here is a question that puzzled scientists for over sixty years: when a person speaks two languages, does the brain keep them in separate compartments, or does it mix everything together? The question sounds simple. The answer turned out to be one of the most surprising findings in modern neuroscience.

More than half the world's population uses two or more languages every day [1]. A taxi driver in Brussels switches between French and Dutch mid-sentence. A graduate student in Göttingen reads a paper in English, takes notes in Turkish, and argues with her roommate in German. A six-year-old in Montreal answers his mother in Punjabi and his teacher in French, without pausing to think about it. Their brains are performing something that no computer can replicate well: storing, retrieving, and switching between entire language systems in real time, with almost no measurable delay.

For decades, the dominant assumption was straightforward. Two languages, two storage bins. Like filing cabinets in an office, each language would occupy its own neural territory. The brain would open one cabinet, close the other, and that was that. But starting in the late 1990s, brain scanners told a different story. Both languages live in the same neural neighborhoods. They activate simultaneously, whether the speaker wants them to or not. And the brain does not separate them by building walls. It separates them by building a traffic control system [2].

This article tells the story of how that discovery happened. It follows the researchers, the experiments, and the arguments. It moves from a linguistics professor in New York writing theory in 1953, through brain scanners in London and Toronto, all the way to a seven-tesla magnet in Paris in 2023 that revealed something nobody expected.

The Linguist Who Drew the First Map

The story begins in 1953, in a book most people outside of linguistics have never heard of.

Uriel Weinreich was a young linguist at Columbia University. Born in Vilnius, raised in several countries, fluent in Yiddish, English, German, French, and other languages, Weinreich understood bilingualism from the inside. His book, *Languages in Contact*, proposed something that seems obvious now but was radical then: the relationship between two languages inside a single mind is not one-size-fits-all [3].

Weinreich described three types. In the first, which he called "compound," both languages connect to the same meaning. A compound bilingual who knows French and English has one mental picture of "dog" that both *dog* and *chien* link to. In the second type, "coordinate," each language has its own separate meaning system. The word *chien* connects to a slightly different concept than *dog*, because the person learned them in different contexts. In the third, "subordinate," the second language can only reach meaning by first passing through the first language. The Spanish speaker learning English hears *table* and mentally translates it to *mesa* before understanding it.

This was 1953. There were no brain scanners. No fMRI machines. No way to check whether Weinreich's three types corresponded to anything real inside the skull. But his framework gave the field its vocabulary. For the next forty years, every study of bilingual memory started from this question: compound or coordinate? Shared meanings or separate ones?

The answer, as it turned out, was: it depends on when the person learned the language, how well they speak it, and which part of the language system you examine.

Two Dictionaries or One?

By the 1990s, the debate about bilingual storage had crystallized into competing models, each backed by experimental evidence. The question was no longer philosophical. It was testable.

In 1994, Judith Kroll and Erica Stewart at Penn State proposed the Revised Hierarchical Model [4]. The model made a claim: there are two separate word-form stores (one per language), but they both connect to a single shared conceptual store. The important detail was the strength of the connections. For a person who learned their second language after childhood, the link from L2 words to L1 words is strong (because that is how the words were learned, through translation). But the link from L2 words directly to their meanings is weak. Over time, with proficiency, the direct meaning link gets stronger and the translation link weakens.

This explains a common experience. A beginning Spanish learner hears *gato* and thinks "cat" before picturing the animal. A fluent bilingual hears *gato* and sees the animal directly. Same word. Same brain. Different wiring depending on proficiency.

Eight years later, Ton Dijkstra and Walter van Heuven at Radboud University in the Netherlands proposed something more radical: the BIA+ model [5]. Their claim was that the bilingual brain does not have two separate dictionaries at all. It has one integrated lexicon. When a bilingual person reads a word, both languages activate simultaneously. Reading the English word "coin" also activates the French word *coin* (which means "corner"). Reading "room" activates the Dutch word *room* (which means "cream"). The brain does not choose which language to activate first. It activates everything, in parallel, all the time.

If that sounds chaotic, consider what it means. Every time a bilingual person reads or hears a word, their brain is running two searches at once. Both languages are firing. Both are competing for selection. And this happens so fast, so automatically, that the person never notices.

What does this mean for someone studying a second language? It means the old advice of "think in your target language" is not quite right. The brain already activates both languages whether the learner wants it to or not. The real skill is not silencing one language. The skill is training the control system that manages the competition.

The Switch Inside Your Head

If both languages are always active, there must be a mechanism that prevents chaos. A bilingual speaker does not accidentally say *gato* when they mean to say "cat." Something in the brain is selecting the right language and suppressing the wrong one.

In 1998, David Green at University College London published the Inhibitory Control Model [6]. His idea was not that the brain turns off the unwanted language. Instead, it actively suppresses it. And the suppression is proportional to the activation. The dominant language, because it is more strongly activated, requires more inhibition when the speaker is using the weaker language. This creates a counterintuitive prediction: switching back into the dominant language should be harder than switching into the weaker one, because the dominant language has been more strongly suppressed.

Experiments confirmed this. The phenomenon is called asymmetric switch costs. It takes more time and effort to switch from L2 back to L1 than from L1 to L2. The brain is literally fighting its own suppression.

But where in the brain does this control happen? In 2006, Jenny Crinion and her colleagues at University College London published a study in *Science* that pinpointed a key player [2]. Using brain imaging on German-English and Japanese-English bilinguals during a word-meaning task, they found that the left caudate nucleus, a structure buried deep inside the brain within the basal ganglia, responded specifically when the language of words changed. It was not responding to meaning. It was responding to the language tag. The left caudate was acting as a language switch.

This was a surprise. The caudate is not part of the language cortex. It sits deep in the brain's motor-control circuitry. But it turned out to be perfectly positioned: it connects to the prefrontal cortex above and to the rest of the basal ganglia below, forming a control circuit that monitors which language is active and adjusts accordingly.

Later work by Jubin Abutalebi and colleagues at the University of Milan added another region to the control network: the anterior cingulate cortex, or ACC, a curved strip of cortex that monitors conflicts between competing responses [7]. They found that bilinguals used the ACC more efficiently than monolinguals during conflict tasks, and that bilingual ACC gray matter volume was larger. The brain was physically remodeling itself in response to a lifetime of managing two languages.

When You Learned Changes Where It Lives

In 1997, a study appeared in *Nature* that made headlines. Karl Kim and Joy Hirsch at Memorial Sloan Kettering Cancer Center in New York scanned the brains of twelve bilingual volunteers using fMRI while they silently generated sentences in each of their languages [8].

Six were early bilinguals who had learned both languages before age six. Six were late bilinguals who had learned their second language as teenagers. What Kim and Hirsch found was this: in Wernicke's area, the brain region associated with language understanding (a patch of cortex in the left temporal lobe), both groups showed overlapping activation for both languages. No separation. But in Broca's area, the region linked to language production (in the left frontal lobe), something different happened. Late bilinguals showed two distinct activation clusters, separated by several millimeters. Early bilinguals showed a single, overlapping cluster.

The interpretation: the age at which a language is learned affects where the brain stores its production machinery. Learn early, and both languages share the same frontal real estate. Learn late, and the brain carves out adjacent but separate patches.

Seven years later, Andrea Mechelli at University College London added a structural dimension [9]. Using voxel-based morphometry on the brains of English-Italian bilinguals, Mechelli found that gray matter density in the left inferior parietal cortex was higher in bilinguals than monolinguals. The density correlated positively with proficiency and negatively with age of acquisition. Early, highly proficient bilinguals had the densest parietal cortex. Late, low-proficiency learners had the least structural change. The brain was not merely reorganizing its functions. It was physically growing more tissue in the regions that handled bilingual processing. This study was published as a brief communication in *Nature*, just one page long, and it has been cited over 800 times.

What does this mean practically? For parents deciding whether to raise children bilingually, the neuroscience points in one direction: earlier is structurally different from later. A child who grows up hearing two languages will likely store them in more overlapping brain tissue, with less need for the effortful control systems that late learners rely on. But "later" does not mean "too late." Proficiency can partially compensate for late acquisition, and brain structure continues to change throughout life.

The Scanner Speaks Forty-Five Languages

For most of the twentieth century, language research was limited to English speakers. This created an enormous blind spot. English has rigid word order, no grammatical gender, limited verb morphology, and uses the Latin alphabet. Generalizing from English to all languages is like studying one species of bird and claiming to understand flight.

In 2022, Saima Malik-Moraleda and Ev Fedorenko at MIT published a study in *Nature Neuroscience* that tried to close this gap [10]. Their team scanned the brains of eighty-six bilingual speakers across forty-five languages from twelve language families, ranging from Mandarin to Swahili to Finnish to Basque. While in the scanner, participants listened to translations of passages from *Alice in Wonderland*.

The result was striking: every single language activated the same left-lateralized fronto-temporal network. The topography was so consistent that the correlation between activation maps for different languages approached r = 0.88 in the left hemisphere. Turkish and Norwegian. Mandarin and Tagalog. The same neural neighborhood.

Fedorenko's lab had previously shown that this language network is distinct from the domain-general "Multiple Demand" network that handles problem-solving, math, and executive control [11]. This distinction matters for bilingualism research because it means that when older studies found L1 and L2 "overlapping" in the frontal cortex, some of that overlap may have reflected shared task demands (both languages require attention and working memory) rather than shared language processing.

The implication is powerful: all human languages, no matter how different they sound or look on paper, feed into the same neural machinery. The substrate is universal. What differs between bilinguals is not where the languages go, but how strongly they connect, how quickly they activate, and how efficiently the control system manages the competition.

Chinese Characters and the Split in the Visual Brain

If all languages share the same network, does anything differ at all? In 2023, a study from Paris provided the most precise answer yet.

Minye Zhan and colleagues at the Paris Brain Institute (NeuroSpin) used a seven-tesla MRI scanner, which has roughly three times the resolution of a standard clinical scanner, to examine the Visual Word Form Area, or VWFA, a small patch of cortex in the left fusiform gyrus that specializes in recognizing written words [12]. They scanned twenty-one English-French bilinguals and ten English-Chinese bilinguals using unsmoothed 1.2-millimeter resolution fMRI, a technique that preserves fine-grained spatial detail usually blurred out in standard analyses.

In the English-French group, the VWFA responded identically to words in both languages. The brain saw English and French text as the same kind of visual input. No separation at the millimeter scale.

But in the English-Chinese group, something different appeared. The VWFA contained distinct sub-patches: some responded to Latin letters, some responded specifically to Chinese characters, and some responded to both. Unexpectedly, the Chinese-specific patches also responded to faces. The researchers speculated that Chinese character recognition, which involves holistic shape processing rather than the sequential letter decoding used for alphabetic scripts, may recruit face-processing circuitry.

This is the strongest evidence to date that *some* genuine spatial separation exists in bilingual storage. But the separation is highly localized, confined to visual input processing, and only appears when the two writing systems are fundamentally different (an alphabet versus a logographic script). Two alphabetic languages showed no separation at all.

What does this mean? It means the question "are two languages stored separately?" has no single answer. At the level of meaning, they share a store. At the level of grammar, they share a processing system. At the level of sound, recent intracranial recordings from UC San Francisco found language-specific sound patterns occupying adjacent but distinct cortical patches, just millimeters apart [13]. And at the level of visual recognition, it depends entirely on whether the two scripts look alike.

Switching Without Thinking

On any street in Hyderabad, in a coffee shop in Montreal, or at a dinner table in Luxembourg, bilingual speakers do something remarkable. They switch languages mid-sentence. Sometimes mid-word. They do it without planning, without pausing, and often without noticing. Linguists call it code-switching. And for decades, it was viewed as a sign of confusion or limited ability.

In 2021, Sarah Phillips and Liina Pylkkänen at New York University published a study in *eNeuro* that overturned this view entirely [14]. Using magnetoencephalography (MEG, a technique that measures the tiny magnetic fields produced by neural activity, with millisecond precision), they studied fluent Korean-English bilinguals. Participants saw two-word combinations that either came from one language or mixed languages, and judged whether they matched a picture.

The key finding: the left anterior temporal lobe, which is the brain's hub for combining words into meaningful expressions, treated mixed-language combinations identically to single-language ones. The compositional machinery did not detect that a language switch had occurred. It combined *Korean adjective + English noun* using the same neural process it used for *English adjective + English noun*.

Code-switching is not confusion. It is the brain doing exactly what it always does: combining meanings. The language labels are irrelevant to the meaning-building system. Phillips and Pylkkänen described this as evidence that the brain uses "a shared mechanism for combining words from a single language and for combining words from two different languages" [15].

This finding connects directly to the BIA+ model. If the mental lexicon is truly integrated, with no language gate at the entry point, then mixed-language expressions are not special. They are just another pattern of activation in a unified system. The brain does not care which dictionary the words come from. It cares about what they mean together.

The Shield Against Forgetting

In 2007, Ellen Bialystok, Fergus Craik, and Morris Freedman at York University in Toronto published a finding that caught the attention of the medical world [16]. They reviewed the records of 184 patients diagnosed with dementia at a memory clinic. Half were lifelong bilinguals. Half were monolinguals. The groups were matched on education, occupation, and cognitive test scores.

The bilinguals had shown their first symptoms of dementia approximately four years later than the monolinguals.

Four years. That is longer than any pharmaceutical treatment for Alzheimer's disease has ever achieved.

In 2013, Suvarna Alladi and colleagues in Hyderabad, India, replicated the result with an important twist [17]. Their sample was 648 dementia patients, 391 of them bilingual. Hyderabad was an ideal setting because bilingualism there is not linked to immigration (a potential confound in the Toronto study). The result: bilinguals developed dementia symptoms 4.5 years later. The effect held for Alzheimer's, frontotemporal dementia, and vascular dementia. It held even in illiterate patients, ruling out education as an explanation.

How does managing two languages protect the aging brain? The leading hypothesis is cognitive reserve. A lifetime of monitoring, selecting, and suppressing languages exercises the same control circuits, the ACC, the dorsolateral prefrontal cortex, and the basal ganglia, that are needed for general executive function [18]. This exercise builds reserve. The brain accumulates pathology at the same rate, but the symptoms emerge later because the control infrastructure can compensate longer.

Gigi Luk and colleagues at York University found structural evidence for this in 2011 [19]. Using diffusion tensor imaging (DTI, a technique that measures the structural integrity of white matter tracts), they compared fourteen lifelong bilinguals with fourteen monolinguals, all around seventy years old. Bilinguals showed higher fractional anisotropy, a marker of white matter integrity, in the corpus callosum and several major fiber tracts connecting frontal and posterior brain regions. Their wiring was, in a measurable sense, better preserved.

But there is an important caveat. Not all studies agree. The brain's maintenance systems are not fully understood, and a 2013 study by Brian Gold and colleagues found the opposite pattern: bilinguals with *lower* white matter integrity still performed as well as monolinguals, which they interpreted as evidence of compensatory reserve rather than preserved structure [20]. The DTI literature on bilingual aging is genuinely mixed.

The Debate That Will Not Die

No article about bilingualism and the brain would be honest without addressing the biggest controversy in the field: the bilingual advantage.

The claim, popularized in the 2000s, goes like this: because bilinguals constantly exercise their executive control systems, they should perform better on tasks requiring attention, inhibition, and task-switching, even when the task has nothing to do with language. For a decade, study after study seemed to confirm this. Bilingual children outperformed monolinguals on the Simon task. Bilingual adults were faster at switching between rules. The bilingual brain was, the narrative said, sharper across the board.

Then the replication crisis arrived.

In 2013, Kenneth Paap and Zachary Greenberg published a direct challenge [21]. Across multiple experiments with hundreds of participants, they found no consistent bilingual advantage in executive processing. In 2015, Angela de Bruin and colleagues went further and analyzed the publication record itself [22]. They found evidence of publication bias: studies that found a bilingual advantage were far more likely to be published than studies that found no difference. The file drawers were full of null results. A 2018 meta-analysis by Minna Lehtonen and colleagues, pooling data from over 150 studies, concluded that "there is currently no adequate evidence for a bilingual advantage in adults' executive functioning" [23].

Does this mean bilingualism does nothing to the brain? Not at all. The structural changes (Mechelli's gray matter, Luk's white matter, Abutalebi's ACC remodeling) are well documented. The dementia delay, while debated, has been replicated across continents. What is contested is whether these brain differences translate into measurable behavioral advantages on laboratory tasks in young, healthy adults. The answer, for now, appears to be: sometimes, in some populations, under some conditions, maybe. Not the resounding "yes" that early enthusiasm suggested.

The honest position is this: bilingualism changes the brain. Whether those changes produce a general cognitive advantage detectable in the lab remains an open question. The dementia-delay finding stands on stronger ground than the executive-function advantage, but even that finding has critics who point to possible confounds in retrospective clinic-based studies.

When Two Become Five (or Fifty-Five)

If two languages change the brain, what happens with five? Or twenty? Or fifty-five?

In 2021, Olessia Jouravlev, Idan Blank, and Ev Fedorenko at MIT studied seventeen polyglots, nine of whom were hyperpolyglots, meaning they spoke ten or more languages. One participant spoke fifty-five [24]. Using precision fMRI, the researchers measured language network activity while participants processed their native language.

The result was paradoxical. Polyglots showed *less* neural activity in their language network during native language processing, not more. Their language regions responded with smaller activations in both magnitude and extent compared to matched monolingual controls. The effect was specific to the language network. Domain-general executive regions showed no difference.

The interpretation: multilingual experience makes the language system more efficient. Instead of needing more brain power to handle more languages, the system becomes streamlined. It processes familiar language with less effort, freeing resources for the less familiar ones. In a follow-up study published in *Cerebral Cortex* in 2024, Fedorenko's team found that the polyglots' language network responded to unfamiliar but related languages (like a Portuguese speaker hearing Spanish) more strongly than to completely unfamiliar languages, suggesting a graded system of activation scaled to comprehensibility [25].

Structural evidence supports the dose-response relationship. Grogan and colleagues found that speakers of three or more languages had greater gray matter density in the posterior supramarginal gyrus than bilinguals, who in turn exceeded monolinguals [26]. More languages meant more structural adaptation, in a measurable, graded fashion.

The conclusion from the polyglot research is clear: the architecture does not change. A brain storing fifty-five languages uses the same network as a brain storing two. What changes is efficiency. And what drives the structural adaptations is not the number of languages per se, but the intensity and diversity of language use.

What the Electrical Signals Reveal

Brain scanners measure where activity happens. But electroencephalography (EEG) and magnetoencephalography (MEG) measure *when* it happens, with millisecond precision. And the timing differences between how bilinguals process their two languages tell a story that scanners cannot.

Two EEG components are central to this story. The first is the N400, a negative voltage wave peaking around 400 milliseconds after a word is presented, associated with semantic processing, meaning how easily a word fits its context. The second is the P600, a positive wave peaking around 600 milliseconds, associated with syntactic processing, meaning how the brain evaluates grammar.

In 1996, Christine Weber-Fox and Helen Neville at the University of Oregon tested Chinese-English bilinguals who had arrived in the United States at different ages [27]. They presented sentences with semantic anomalies ("The scientist criticized the math") and syntactic violations ("The scientist was criticized math") while recording brain electrical activity.

The results drew a sharp line between meaning and grammar. The N400 response to semantic anomalies was present in all groups, though slightly delayed in those who learned English after age eleven. Meaning processing transferred across ages. But the P600 response to syntactic violations told a different story. Participants who acquired English after age sixteen showed no P600 at all. Their brains were not processing grammar violations the way native speakers did. Syntax was the vulnerable domain. Meaning was the resilient one.

This distinction fits Michael Ullman's Declarative/Procedural model [28]. Ullman proposed that vocabulary and memorized word forms are stored in declarative memory (the same system that stores facts and events, centered on the hippocampus and temporal lobe), while grammar is stored in procedural memory (the same system that stores motor skills and habits, centered on the basal ganglia and frontal cortex). In a first language, grammar becomes automatic and procedural. In a late second language, grammar may remain declarative, relying on explicit rules rather than automatic processing, which explains why the P600 signature can be absent.

But there is a twist. Kara Morgan-Short at the University of Illinois trained adults on an artificial language called Brocanto2 [29]. One group learned through explicit instruction (rules and explanations). Another learned through immersion (context and practice only). At high proficiency, the immersion group showed the native-like P600 response. The explicit group did not. The implication: it is not just about age. The *type* of learning experience matters. Immersive, implicit learning can produce native-like brain responses even in adults.

The Whole Picture

The story of how bilingual brains store languages has moved through three phases. The first was speculation: Weinreich's compound-versus-coordinate taxonomy, built from logic and observation alone. The second was modeling: Kroll's hierarchies, Dijkstra's parallel activation, Green's inhibitory control, each a map drawn from behavioral experiments. The third, and current, phase is direct observation: brain scanners, electrode arrays, and millisecond-precision timing that show what actually happens inside the skull.

The picture that has emerged is not the one anyone expected. Languages are not stored separately. They are not stored in a single undifferentiated blob either. The brain uses a shared network, the same left-lateralized fronto-temporal system that all humans share for language, and it layers both languages into that network with graded connection strengths. A control system built from deep subcortical structures and prefrontal cortex manages the competition. Age of acquisition and proficiency determine how tightly the two languages overlap and how much control effort is required.

Genuine separation exists, but only at the edges: in the visual cortex when scripts are radically different, in the sound cortex where phoneme inventories diverge, and possibly in frontal production areas when acquisition is late. Everything else, meaning, grammar (in proficient speakers), and the compositional machinery that builds sentences, runs on shared hardware.

What does this mean for the millions of people who are bilingual, or who are learning to be? It means the brain was not designed for one language. It was designed for communication. And communication, for most of human history and for most humans alive today, has meant more than one language. The neural machinery is ready. It has always been ready.