A sentence in Japanese saves its verb for last. A sentence in Welsh puts the verb first, ahead of subject and object alike. Edit enough manuscripts that started life in one of those languages before arriving in English, and the difference stops reading as a stylistic tic and starts reading as two settings on the same underlying machine.
A test built to break its own assumptions
That machine is what a team led by Annemarie Verkerk of Saarland University and Russell D. Gray of the Max Planck Institute for Evolutionary Anthropology set out to test, in a study published in Nature Human Behaviour in late 2025. The team drew on Grambank, the largest existing database of grammatical features, covering more than 1,700 languages, and used it to check 191 rules that linguists have proposed over the past half-century as universals: patterns claimed to hold, with rare exceptions, across every language on earth, from how questions are formed to how possession is marked to which word classes exist at all.
Most of those 191 rules came out of an era when researchers picked languages from far-flung regions specifically to avoid comparing close relatives, then generalized from the sample. It is a sound instinct but an incomplete one. Languages in contact borrow from each other, and language families carry inherited traits down through centuries, so a sample that looks geographically diverse can still hide statistical ties that inflate how universal a pattern appears. Verkerk and Gray’s team used Bayesian spatio-phylogenetic analysis instead, modelling both descent and geography directly rather than trying to sidestep them by hand-picking a sample. It is a more demanding test, and by design a harder one for any given universal to pass.
What actually survived
About a third of the 191 proposed universals held up under that scrutiny. Two categories did most of the surviving: word order and what the researchers call hierarchical universals, the way grammatical relationships get layered and marked within a sentence, such as which participants in an event are singled out for special marking and in what order those markings stack.
The clearest example is one of Joseph Greenberg’s original 1960s universals: languages with normal subject-object-verb order are overwhelmingly likely to use postpositions, markers that follow the noun, rather than prepositions, which precede it.
Verkerk’s team did not just confirm that the correlation holds today. They traced how it arose independently, again and again, across unrelated language families and geographic regions with no contact with one another, evidence that something is nudging separate lineages of language toward the same solution rather than one lineage lending it to the next.
Order as evidence of assembly, not decoration
Word order looks, to most writers, like a fact to memorize about a language rather than something that means anything. The recurrence Verkerk and Gray describe argues otherwise. If unrelated languages, with no shared ancestor and no contact, keep landing on the same order for the same job, verb and object, marker and noun, that order is doing structural work rather than decorative work.
A related line of research makes the mechanism more concrete: Michael Hahn, Dan Jurafsky and Richard Futrell’s 2020 study in PNAS modelled word order across 51 languages as a trade-off between two competing demands, keeping sentences simple enough to produce and unambiguous enough to be understood, and found that real languages cluster near the efficient end of that trade-off far more than chance would predict.
Having these two findings in mind, the order words come in, looking like a rule imposed on thought after the fact. It starts looking like a visible trace of how the relationships between things, who did what to whom, what belongs to what, get organized before any particular word is chosen to carry them.
What the failed two-thirds says about confident rules
The more useful number, for anyone who edits for a living, might be the two-thirds that did not survive. Many of the rules linguists treated as settled for decades turn out to have looked universal partly because of which languages got studied first, and how those languages happened to be related to each other.
A pattern that seems to hold across twenty languages from twenty families can still be an accident of history rather than a fact about the human mind, if those families share more contact and inheritance than the sampling method accounted for.
That is a familiar problem to anyone who edits by inherited rule rather than by checking whether the rule earns its keep in the sentence in front of them. A rule that has been repeated in style guides for decades is not, by that fact alone, a rule that reflects how readers actually process a sentence. Confidence and sample size are not the same thing, in language study or in a style guide.
What holds up when you edit across word order
In manuscripts we work on at Global English Editing, the practical version of this shows up constantly in writing that started in another language’s word order and is being reshaped into English’s. A modifier that trails its noun in French sits awkwardly trailing it in English too.
A relative clause that a Mandarin sentence tucks before the noun it describes needs to move after it in English, or the sentence reads as though it is still translating itself in front of the reader. None of this is about correcting an error in the source language. The original sentence was not wrong; it was assembled on a different, equally coherent setting of the same machine. The edit is about recognizing which relationships in the sentence are load-bearing, who is the subject, what depends on what, which clause frames and which clause fills in, and rebuilding them in the order English readers expect that information to arrive: given material first, new material after, the governing clause before the one that depends on it. Editors who do this well are not applying a memorized rule about clause position. They are tracking the same hierarchical structure Verkerk’s data shows recurring across 1,700 unrelated languages, just rebuilt in one target language’s preferred sequence.
The Japanese sentence that saves its verb for last and the Welsh sentence that leads with it are not exceptions to some tidier rule sitting underneath them. They are two of the finite settings that the same underlying architecture allows, arrived at independently, by populations of speakers who never met each other and never read Greenberg’s papers. Editing between them is less like fixing a mistake than like reading two dialects of the same instinct for how thought puts itself in order before it reaches for words.
