If you have ever learned Japanese or Korean, you have probably groaned: why is everything backwards? "Go to Tokyo" becomes "Tokyo to go". The odd thing is how tidy the reversal is: when the verb moves after its object, "to" moves after its noun as well. Languages almost never flip just half.
Tidiness like this convinced linguists that one shared system lies underneath every language. But "universal" can be read three quite different ways: something every language must have; something most languages tend to have; and something every human brain can handle but a given language need not use. Much of the heat in this debate comes from the two sides meaning different ones.
In 1963 Greenberg drew a set of universals from thirty languages, and the sturdiest kind took the form "if…, then usually…": if a language puts its object after the verb, it usually has prepositions (little words like in, to, from placed before the noun); if the object comes before the verb, it usually has postpositions (placed after the noun). English go to Tokyo has both the verb and to up front; Japanese 東京に行く (Tōkyō ni iku, literally "Tokyo to go") has both ni and the verb at the end.
A pattern like this is called an implicational universal. It does not say every language does something; it says a certain combination almost never turns up. Later typologists pushed the sample into the thousands and the diagonal held firm. For a more absolute claim — that the subject nearly always comes before the object — the topic on whether large models have made linguistics obsolete tells how an Amazonian language with the object first broke it.
In 2011 Dunn, Greenhill, Levinson and Gray asked the question differently in Nature. Counting languages has a loophole: four hundred Austronesian languages all descend from one ancestor, so their shared verb-first order may be a single inheritance counted four hundred times. They borrowed biology's method of drawing evolutionary trees and examined 589 languages in four families — Austronesian, Bantu, Indo-European and Uto-Aztecan — asking: when a language changes its verb–object order, does its adposition order change with it? The answer was that most dependencies between word-order features show up only in some families, and take a different shape in each.
The reply from the other side is strong too: four families are few, each family contains only so many word-order changes, and the statistical power is too thin to overturn a worldwide correlation. Today's rough consensus: the diagonal is a real tendency, but whether it is a law or a well-worn path is still open.
Pirahã is spoken along the Maici River in the Brazilian Amazon by only a few hundred people; the other members of its Mura family are no longer spoken. Daniel Everett did fieldwork there for some thirty years. In 2005 he published the claim that the language does not put clauses inside clauses. Where English says "John's brother's house", Pirahã speakers would split it up: "Brother's house. John has a brother. It is the same one." His explanation was cultural: the community talks about what has been seen first-hand or reported by someone who was there, leaving no use for structures nested layer within layer.
This was a challenge to a very strong claim. In 2002 Hauser, Chomsky and Fitch proposed that the part of the human language capacity that is truly unique might be recursion alone — a structure that can hold another structure of the same kind, so that "I know you think he said…" can keep going down. If a fully working human language did without it, the hypothesis would at least need rewriting.
In 2009 Nevins, Pesetsky and Rodrigues wrote a long rebuttal with two core points. First, Everett's own 1986 grammar of Pirahã analysed some sentences as embedded clauses; in 2005 he changed the analysis, but not the material. Second, the "gaps" occur in other languages too: German's possessive placed before the noun also resists stacking a second layer, yet nobody says German culture limits German grammar.
In 2016 Futrell and colleagues went sentence by sentence through 1,149 sentences of natural speech that Everett had helped collect years before. They found no unambiguous embedding, but conceded that the corpus is small and was only shallowly analysed, so embedding cannot be ruled out. The evidence needs a clear label: one language, with almost all first-hand material from a very small number of researchers, hard for outsiders to check. And one more layer should not be forgotten: the much-cited "exhibit" is a living community whose language exists for living in, not for this dispute.
The Pirahã dispute finally jams at a fork. If "recursion is universal" means every language must use it, one language that truly does not refutes it. If it means every human brain has the capacity, Pirahã cannot refute it — Pirahã children raised elsewhere learn the embedded clauses of Portuguese like anyone else. Everett's side asks back: is a universal that no language could ever refute still an empirical claim about language? Both sides have a point, and the argument continues today.
Meanwhile another kind of universal has been quietly gathering evidence. In 2015 Futrell, Mahowald and Gibson took large annotated corpora of 37 languages and measured how far apart every pair of grammatically linked words sat — a verb and its object, say. In every language the average distance was shorter than a baseline with the word order randomly shuffled, and the gap widened as sentences grew longer. This is called dependency length minimization.
You do it yourself. In English both throw out the old rug and throw the old rug out are fine; but once the object gets long, speakers nearly always choose the first, keeping out beside throw instead of leaving it stranded behind a long string of modifiers. Look back at the figure: in English go sits next to to, in Japanese ni sits next to iku; the crossed combinations would wedge a whole noun phrase between verb and adposition. The diagonal may be a product of exactly this pressure.
On this account, universals live not in a grammar written into the genes but in speakers' limited working memory. It explains the tendencies, though not why some languages take the rare road anyway; and the universal-grammar side can reply that such pressure and innate structure may well coexist. Evidence strength: corpus statistics across 37 languages, pointing the same way as the finding, covered in the topic on which machines can handle which languages, that real texts rarely nest clauses more than three deep — but these are correlations, not direct proof of a mechanism.
The grid already holds 14 such languages. One counterexample can topple an absolute universal but not a tendency. What is informative is how it got that way, and whether it has just been through heavy contact.
One approach is to look at whether speakers of that language show layer-within-layer structure outside syntax — in storytelling, planning, counting. If it is missing there too, the capacity claim has to shrink.
Not necessarily: they learn from human writing, where the traces of the pressure are already present. To tell them apart, you would check whether a model trained on an artificial language that deliberately pulls related words apart shrinks the distances again on its own when generating.
The discomfort comes from expectations your first language has already set, not from the target language itself. Japanese children learning English find go to Tokyo the backwards one.