Sapir-Whorf, carefully

The claim most people have heard, that your language determines what you are able to think, is not a position any serious linguist holds, and is not one Benjamin Lee Whorf stated in that form. It survives because it is a good story, and because the single most repeated piece of evidence for it was never true.

The Eskimo snow vocabulary is a fabrication with a traceable history. Laura Martin followed the chain in 1986 and Geoffrey Pullum popularised her finding in 1991: a modest observation in Boas about a handful of roots was inflated by each retelling until it reached the dozens, then the hundreds. Boas had listed four. English does much the same thing with snow, and nobody concludes anything from it. Worse, the languages in question are polysynthetic; counting their "words" for anything is an ill-posed exercise, so the factoid was not merely wrong but unmeasurable.

What remains after that is smaller, better evidenced, and more interesting. Not language determines thought, but language reliably shifts performance on particular tasks, by measurable amounts. The question worth asking is where it does so, and by how much.

It helps to begin with what Whorf actually wrote, because the hypothesis that carries his name was assembled after his death. Whorf and Edward Sapir never co-authored anything, and Whorf never called his view a hypothesis; the label was applied by Harry Hoijer in 1954, and the tidy split into a "strong" deterministic version and a "weak" relativist one is a later reconstruction again. Whorf's own prose runs hot and cold. The most quoted line, from a 1940 magazine piece, is that "we dissect nature along lines laid down by our native languages"; elsewhere in the same essays he is far more guarded. He was a fire-insurance inspector with a gift for arresting sentences, and the arresting sentences are the ones that got remembered.

His famous case has fared badly in the particulars. Whorf claimed, with his own scare quotes, that Hopi contains no forms referring directly to what we call "time". Ekkehart Malotki's 1983 study Hopi Time answered with several hundred pages of Hopi temporal expressions, including a future marker. Whether Malotki refuted Whorf or a caricature of Whorf is still argued; what nobody now defends is the claim as popularly repeated, that the Hopi lack a concept of time.

The experimental literature that grew out of this wreckage is more careful, and its best results share a shape. Paul Kay and Willett Kempton asked, in 1984, whether the colour lexicon bends colour judgement. English splits green from blue; Tarahumara, in northern Mexico, covers both with one word. Asked to judge which of three colour chips stood apart, English speakers exaggerated the distance across their blue-green boundary; Tarahumara speakers did not. Then comes the telling part. When Kay and Kempton restructured the task so that silently naming the chips was no use, the English distortion vanished. The language effect was real, and it was also contingent; a naming habit the task could switch off, not a lens bolted to the eye.

The strongest body of evidence sits in space rather than colour. Guugu Yimithirr, in far north Queensland, runs its spatial talk on compass bearings; not "the cup to your left" but, in effect, "the cup to the south". Stephen Levinson's fieldwork, gathered in his 2003 book, found that speakers of such languages maintain a running sense of the bearings and tend to solve non-linguistic puzzles, like rearranging objects after being rotated, in absolute coordinates where English speakers use left and right. The correlation is robust and much replicated. The causal reading is contested; Peggy Li and Lila Gleitman showed in 2002 that English speakers can be nudged into absolute responding by changing the room, and Levinson's group answered in kind. The debate is genuinely open, which is a different thing from settled either way.

Number tells the tidiest story of the lot. The Pirahã language of the Amazon has no exact number words at all; its three quantity terms turn out to mean something like "few", "fewer" and "many". Peter Gordon reported in 2004 that Pirahã adults faltered on exact-quantity tasks, and read this as language limiting thought. Michael Frank and colleagues returned in 2008 with a sharper design and a sharper conclusion: the Pirahã matched quantities exactly when the objects sat in front of them, and faltered only when the task required holding an exact count in memory. Number words, they argued, are a cognitive technology; not a precondition for the concepts, but a tool for storing and comparing them, the way writing is a tool for storing sentences. You can think without the tool. The tool still matters.

Around these results lies a graveyard of famous findings that did not hold, and the graveyard is instructive. The much-repeated study in which German speakers called a key "hard" and "jagged" while Spanish speakers called it "golden" and "intricate" was never published in a peer-reviewed journal, and a 2014 attempt to replicate it found nothing. The claim that languages without a grammatical future make their speakers better savers did not survive statistical controls for the fact that related languages share cultures; one of the authors of the correction was the economist who proposed it. This field's pattern is consistent: the spectacular effects evaporate, the modest ones survive.

So what survives? Effects of language on cognition that are real, measurable, domain-specific and task-dependent; strongest where a language forces a choice, as Russian does with its blues and Guugu Yimithirr does with its bearings; and removable, in the laboratory, by occupying the verbal channel or redesigning the task. Winawer's group put the surviving thesis in one sentence: the difference is not that English speakers cannot distinguish light from dark blue, but that Russian speakers cannot avoid doing so. Language on this evidence is not a prison for thought. It is closer to a well-worn path across it; you are free to leave the path, and you mostly don't.