You cannot hear your native language as noise. Two people are talking nearby in a language you understand, you are trying to concentrate, and their sentences keep letting themselves in one at a time. Switch them to a language you know nothing of, same volume, and within a few minutes you have filed it away as background sound.
That "can't switch it off" quality is one of the oldest sources of the intuition that people carry a device dedicated to language: a machine you cannot unplug, which runs to completion the moment input arrives. But it is a long walk from "can't switch it off" to "there is an organ." Here are the three things most often used to make that walk: a family, a gene, a patch of cortex.
Around 1990, clinicians in London noticed an unusual family. Across three generations, half the members had been hard to understand since childhood — not a small vocabulary, but sounds that came out distorted, and written sentences that were often missing parts. The other half had no difficulty at all. Half and half, in every generation: that ratio is conspicuous in genetics, the signature of a single dominant gene. By convention the researchers referred to them by two letters, KE.
The first person to test their language was Gopnik, using a very old trick: give someone a word they have never seen and watch whether they add the ending. A card shows a creature nobody has ever seen, and the experimenter says, "This is a wug. Now there are two…" A four-year-old English speaker fires back wugs — what the child learned was not the word wugs but the rule "add -s." Same for verbs: "Every day I brop. Yesterday I…" and out comes bropped. The affected members of this family fell behind badly on items like these, while doing far better on familiar words. Gopnik concluded that what was damaged was the machinery for grammatical marking itself — number, tense. The press picked it up and gave it a name: the grammar gene.
Vargha-Khadem and colleagues then ran a full battery on 13 affected and 8 unaffected members, and the picture came back much more complicated: what they struggled with went well beyond grammar. Affected members had clear difficulty controlling complex coordinated movements of the face and mouth — ask them to repeat "pa-ta-ka" in a run, which hands the job to lips, then tongue tip, then tongue back in turn, and the sequence falls apart. That difficulty is called orofacial dyspraxia. On nonverbal intelligence tests, the affected group also averaged lower.
The most interesting result is the test that separates the two groups most cleanly: nonword repetition. You hear a string of syllables with no meaning at all — perplisteronk, blonterstaping — once, and repeat it. On this task every affected member scored below every unaffected member, with no overlap. And difficulty on this task tracks how complex the articulatory movements are; it has nothing to do with grammatical marking. If what was broken were really "processing grammatical features," scores should not line up along how hard the mouth has to work.
One more thing: Gopnik's test is bolted to English. Mandarin verbs carry no obligatory tense ending — in 昨天我去 (zuótiān wǒ qù, "yesterday I go") and 明天我去 (míngtiān wǒ qù, "tomorrow I go") the verb 去 (qù, "go") is identical, and time is carried by words like 昨天 (zuótiān, "yesterday") and 明天 (míngtiān, "tomorrow"). The same damage in a Mandarin speaker simply cannot surface as "dropping the past tense."
In 2001, Lai and colleagues pinned the causal change to chromosome 7: a gene called FOXP2, one base altered, swapping the amino acid at position 553 from arginine to histidine. That position sits inside the stretch of the protein that grips DNA. Grip loosens, and everything downstream is affected.
Downstream is the operative word. FOXP2 codes for a transcription factor — it does no work itself; its work is deciding when other genes switch on and off, and its downstream targets number in the hundreds. Nor does it appear only where speech is handled: it is expressed in lung, heart, gut, and cerebellum. Break a master switch and a large territory shifts with it. From "damage it and speech breaks" you cannot infer "it encodes speech."
In 2002 Enard and colleagues gave the name a hard shove. Comparing FOXP2 across humans and other primates, they found two amino acid changes on the human lineage, plus traces of a selective sweep around the gene — when a stretch of genome is strongly selected, diversity around it gets flattened, and the estimated date of that flattening landed within the last 200,000 years, matching the arrival of modern humans. One gene, two changes, 200,000 years, humans start talking: a story too well-tailored to be true.
Then it came apart in layers. In 2007 Krause and colleagues recovered FOXP2 from Neanderthal bone: those two "uniquely human" changes are there in Neanderthals too. In 2018 Atkinson and colleagues recomputed the sweep on far larger and far more broadly sampled data, and the signal vanished — it had come from sampling. African populations and populations that left Africa have different histories, and pooling them manufactures a false sweep; analyzed separately, it is gone. A plainer point sits alongside: common variation in FOXP2 among ordinary people shows no relationship to individual differences in language development. It marks the site of a rare, disruptive, large-effect variant, not a dial that sets how good your language is.
So what layer does it govern? A species with no speech offers the hint. In 2007 Haesler and colleagues lowered FoxP2 in a region called Area X in zebra finches before the birds began learning to sing. The birds still vocalized, but copied their tutor's song imprecisely and sang more variably. Bird song learning and human speech learning share one thing: listen to a model, practice repeatedly, tune a long string of fine movements into place. This gene looks like it governs that layer.
Two clues shrank and collapsed; the third has only gotten stronger. What makes Fedorenko's method clean is the control condition: have someone read an ordinary sentence, then read a string of nonwords matched for length, format, and font but carrying no meaning (the wug kind of word, just not adding up to anything). In both cases the eyes move, letters are recognized, reading happens; the only difference is whether there is meaning and structure. Subtract one pattern of brain activity from the other and what remains is a network across the left frontal and temporal lobes. One step matters a great deal: this network sits in noticeably different places in different people, so it must be located person by person — averaging many brains together blurs it away.
Once located, the real test begins: knock on that same patch with other tasks. Mental arithmetic, listening to music, holding a sequence of locations in mind, reading program code, judging what someone else believes (the "he thinks there is chocolate in the box, but there isn't" kind of question) — none of these is easy, and all of them light up other regions, but this network barely moves.
The mirror half is just as hard. Fedorenko and Varley reviewed a set of patients with global aphasia in 2016: language essentially gone, comprehension and production both collapsed. These people can still add and subtract, solve logic problems, judge what another person is thinking, appreciate music, and find their way around a city. Put the two halves together and you have a double dissociation: healthy people doing nonlinguistic tasks leave this patch quiet, and people whose patch is destroyed can still do nonlinguistic tasks. Functionally, there really is something highly specialized here — that much is on firm evidence today.
The trouble is that "specialized" and "innate" are two different claims, and the phrase "language organ" ties them together. Malik-Moraleda and colleagues took the same localizer to 45 languages across 12 language families in 2022, from Basque to Japanese to Swahili, and the network's location, its leftward bias, and its indifference to nonlinguistic tasks came out the same everywhere; signers use that same left-hemisphere network too. Something that stable sounds like it could only be innate.
Except a counterexample sits right beside it. Reading also has a patch of cortex with a remarkably stable address, at the left occipitotemporal junction, and readers of every writing system grow it in the same place. Writing is only a few thousand years old — far too recent to have evolved an organ. That patch was repurposed from visual circuitry that was already there. So "everyone grows it in the same place" is evidence compatible with both "the genes specified the blueprint" and "given the brain's existing wiring constraints, it could only end up there." It cannot separate the two.
Split it into two sentences: one says there is a patch in the adult brain that responds only to language; the other says its location and division of labor are specified by genes. The first has hard evidence today; the second does not. How would you design a study that separates them? Is where this network ends up in people born blind or born deaf a useful clue?
Nonword repetition asks you to hold an unfamiliar string of sounds for a few seconds and deliver it to your mouth intact. It loads memory for sound, ordering, and articulation at once. If the most sensitive marker of a disorder lands on that kind of ability, then how much of what we call "language ability" is proprietary to language, and how much is general machinery that language keeps calling?
The reading region permits a comparison the language network does not: put literate and non-literate adults side by side. Language has no equivalent control group — there is no such thing as a typical adult who grew up without language. Without that comparison, is there another way to close in on the answer?