Language as a window into human cognition

The study of diverse languages provides a window into human cognition. Each language can be thought of as a natural experiment in building an effective communication system. An enormous combinatorial array of grammatical possibilities is available to construct these systems, but linguists have identified recurring patterns - so-called universals - that are strong statistical correlations between features of grammar. An example of a universal is ‘languages with subject–object–verb word order have postpositions’. Japanese is such a language where the verb comes last in the sentence and adpositions follow nouns: for example, 桜のために (Sakura no tame ni) is literally ‘Sakura for’ (whereas English has prepositions; the phrase would be ‘for Sakura’). However, demonstrating universals is challenging because languages can exhibit similar patterns due to inheritance and borrowing. Traditionally, linguists have attempted to circumvent these genealogical and geographical non-independences by sampling widely separated languages. However, sampling can fail to remove all dependencies, reduce statistical power and does not identify historical pathways.

Phylogenetic methods can control for non-independence, make full use of the data and infer historical changes. Such techniques have already been used by Dunn et al. (2011) to test word-order universals. In these analyses, much of the evidence for putative universals disappeared, but only four language families were examined. To overcome this limitation, we used the recently released Grambank database, which provides information on 195 morphosyntactic features in a global sample of more than 2,000 languages. To avoid cherry-picking universals, we first scoured the Universals Archive and identified 191 hypotheses that could be tested using Grambank. We conducted Bayesian analyses to test the strength of each universal — for example, the connection between ‘subject–object–verb’ and ‘postpositions’ — while simultaneously controlling for genealogical and geographical non-independence. Our two main analyses did this in different ways: one by accounting for genealogical and geographical distances, and the other by explicitly modelling change across a set of global language trees. These two analyses paint the same picture: less than a third of the proposed universals have robust statistical support (60 out of 191). Thus, our results suggest that although constraints on combinations of grammatical features exist, they are neither as strong nor as ubiquitous as sometimes previously thought. The majority of the universals supported are hierarchical universals (24 out of 30) and narrow word-order universals (24 out of 65). Two other types of universals receive low support. The explicit phylogenetic analyses reveal that, despite the vast combinatorial array of possible systems, preferred feature configurations evolve repeatedly in different language families and areas. We suggest that this convergent evolution reflects common cognitive and communicative pressures. For example, suppose a structural pattern (say, certain word-order/case-marking trade-offs) keeps independently evolving in unrelated families. In that case, that’s good evidence that it fits human processing and learning biases—e.g. minimising dependency length, reducing ambiguity, or aligning with preferred cue structures in comprehension and production. Language, in this view, is not just something the mind uses. It is also a long-running record of how minds across generations have solved recurring problems of communication, categorisation and coordination. Reading that record carefully with appropriate computational tools gives us one of the richest windows we have into human cognition in the wild.

Verkerk, A., Shcherbakova, O., Haynie, H. J., Skirgård, H., Rzymski, C., Atkinson, Q. D., Greenhill, S. J.,& Gray, R. D. (2025). Enduring constraints on grammar revealed by Bayesian spatiophylogenetic analyses (advance online). Nature Human Behaviour. https://doi.org/10.1038/s41562-025-02325-z