You spend a whole day in meetings in a second language, then call home in the evening, and the first sentence in your native language is the one that snags. The word is slow to come. You file this under "tired." It can in fact be measured on its own, and it runs the opposite way from what you would guess.
People who speak two languages are usually pictured as having two drawers in the head, opening whichever one they need. Thirty years of experiments have taken that picture apart. Three things here: how both languages stay online at once, why switching between them turns out to be tightly rule-governed, and what happened to the claim that bilingualism makes you smarter.
A few objects sit on a table: a marker, a postage stamp, some others. You are wearing a device that records where your eyes go, and you hear an instruction in English: pick up the marker. English monolinguals look straight at the marker. Russian–English bilinguals also pick up the right object, but during that first fragment of sound, mar-, their eyes flick briefly toward the stamp — because the Russian word for stamp is марка (marka), whose opening sounds collide with marker.
Not one Russian word appears anywhere. The task is in English, the setting is in English. Yet in the few hundred milliseconds while the sound is still arriving, words from the other language light up on their own. Marian and Spivey ran this in both directions, and it replicates with other language pairs.
A more deeply buried version exists. Thierry and Wu recruited a dozen-odd people whose first language was Chinese and who learned English after the age of twelve, and asked them to judge whether pairs of English words were related in meaning. Train and ham obviously are not. What the participants did not know was that half the pairs had a trap in them: translated into Chinese, the two words share a character — 火车 (huǒchē, "train") and 火腿 (huǒtuǐ, "ham") both carry 火 (huǒ, "fire"). The behavioural data showed nothing at all; button-press speed and accuracy were unaffected. The brain response was not: a wave that arrives about 400 milliseconds after a word (called the N400, and larger the worse a word fits its context) differed reliably on the pairs with the hidden character overlap. Asked afterwards, nobody knew Chinese had been involved.
So "which language" is not a choice made at the door. The words light up together and the selection happens much later. That structure sets up the bill for the rest of this piece: both the cost and the benefit have to come out of "both are open, and one must be picked."
The title of a 1980 paper is itself the example: "Sometimes I'll start a sentence in Spanish y termino en español" — the second half means "and I finish it in Spanish." Poplack recorded a great deal of everyday conversation in a Puerto Rican community in New York in order to test what was then taken for common sense: that mixing like this means neither language was properly learned. Code-switching is the name for it.
The data pointed the other way. Switch points are not sprinkled at random. First, they almost never fall inside a word: you hear switches at a clean joint, as in y termino en español, but not a hybrid gluing an English stem to a Spanish ending. Second, they avoid the places where the two grammars disagree. Spanish puts many adjectives after the noun — la casa blanca, word for word "the house white" — while English puts them in front, the white house. Word order does not line up there, so switches steer around that spot and land on joints where both languages agree.
Myers-Scotton later gave a more general account: within a clause, one language supplies the frame — the word order and the little function words — while the other fills content into the slots.
One finding is blunter still. Poplack found that switching inside a sentence is done by the most fluent speakers; less fluent bilinguals more often finish a sentence before switching, or insert a single noun. Switching mid-clause means keeping both grammars in hand at once, so only those who can do that do it.
The strength of the evidence needs stating. Those two proposed universal constraints later collected counterexamples in bulk — change the language pair, change the corpus, and violations account for a large share of the sentences. The settled statement today is narrower: the exact form of the constraints is still open, but that switch points are jointly constrained by both grammars, rather than random, holds up. As for "mixing is sloppy" — that is a social judgement, and it is not in this data.
Meuter and Allport had people do something very simple: a digit appears on screen, and the background colour tells you which language to name it in. Sometimes several trials run in the same language, sometimes a switch is required. Switch trials are slower than non-switch trials, which is unsurprising. The direction is what surprises: switching from the second language back into the native one is slower than switching the other way. The better-known language is the more expensive destination.
Green's account: while you are speaking the second language, the native one is not sitting quietly. It is stronger and competes harder, so you have to hold it down for the second language to get out at all. The harder you press, the slower it climbs back once you let go. The cost is not unfamiliarity with the route; it is how much suppression you were applying a moment ago.
That account makes a testable prediction: if the cost comes from the amount of suppression, then people whose two languages are evenly matched need not press either one hard, and the asymmetry should shrink. Roughly so — in balanced bilinguals it often shrinks, vanishes, or even reverses, and a 2021 analysis pooling many studies concluded that it is more reliable in unbalanced bilinguals.
Because that instruction — the colour telling you to switch now — is itself a variable. Blanco-Elorrieta and Pylkkänen had the same bilinguals do two things: switch on an external cue, and switch freely, facing a person who understands only one of the languages, so the choice makes itself. Recorded with an instrument that tracks brain activity millisecond by millisecond, cued switching drove control-related regions early, 100 to 160 milliseconds after the word. In free switching, that rise did not appear.
So "bilinguals pay a switching cost all day long" needs discounting: some of what the lab measures is the cost of having someone else decide when you switch.
The famous chain of reasoning runs like this. Every sentence you speak requires choosing a language and holding the other one down (the mechanism from "The cost of switching shows up on the road home"), which amounts to drilling a general-purpose set of abilities all day: blocking distraction, switching between tasks, monitoring your own errors — collectively, executive control. Drill it enough and the benefit should show up on tasks with no language in them. Over twenty years a body of experiments did report such benefits, mostly in children and older adults.
Then the claim ran into trouble. In 2015 de Bruin and colleagues did something clever: instead of reading published papers, they went through conference abstracts from 1999 to 2012 — abstracts are submitted when the result comes in, before the publication filter. Of the abstracts supporting a bilingual advantage, 68% later became papers; of those challenging it, only 29%, with no difference between the groups in sample size or method. In 2018 Lehtonen and colleagues pooled 152 studies of adults and 891 effect sizes: the advantage was small to begin with, and correcting for publication bias — the tendency for unflattering results to go unpublished — took it to essentially zero. In 2020 Nichols and colleagues ran 12 tasks on more than eleven thousand people: bilinguals were better on 1, monolinguals on 4, and once age, education and the rest were matched, both sides disappeared.
The strongest version of the other side deserves stating too: splitting people into "bilingual" and "monolingual" piles is a crude cut. How many hours a day each language is used, how often they are switched, in what social setting — these vary enormously, and the variation may well be washing a real effect out. The argument is not over.
One thread runs further out and is messier. In 2007, across 184 case files at a memory clinic, bilingual patients showed symptoms of dementia about four years later than monolingual ones. The obvious worry is confounding: those bilinguals were largely immigrants, with different education and different routes to a clinic. A 2013 study of 648 cases in Hyderabad removed a large piece of that worry — there bilingualism is the local norm, both groups came from the same environment, and the 4.5-year gap showed up even in patients who could not read. But prospective follow-up studies find no difference in incidence or in the rate of decline. The restrained statement today: possibly a delay in symptoms surfacing, not a lower risk of the disease, on evidence that is mostly retrospective chart review.
What replicates solidly is the other side of the ledger. Bilinguals are a little slower to name pictures, have more moments where the word will not come, and can name somewhat fewer things in each language. Gollan's account calls it weaker links: each word's share of use is divided across two languages, and low-frequency words take the hit. That is a frequency ledger, not an ability one — the vocabulary of both languages combined is in the same range as a monolingual's.
What is a person who switches freely actually practising — holding down the other language, or judging which one the listener understands? What experiment separates those two? And if what remains after the separation is small, was the bilingual-advantage argument aimed at the wrong object all along?
Someone who uses each language four hours a day and switches ten times an hour lands in the same pile as someone who speaks a second language only in meetings. If you measured continuously instead — hours per language per day, switches per day, whether each switch was chosen or imposed — which variable would you expect to attach to which outcome?
Are two grammars online at once as well? Could a sentence heard entirely in one language be tugged during parsing by the word order of the other? To see that, what should an experiment watch — where the eyes land, how long the sentence takes to read, or the direction in which comprehension goes wrong?