Learn a foreign language and sooner or later you hit a word you cannot carry across. Portuguese saudade: longing for someone or some time that is gone, sweet with an ache in it. You can explain it, but you cannot swap one word for one word. And almost everyone has the same thought at that moment: do speakers of this language actually feel something I do not?
That thought has a name, the Sapir-Whorf hypothesis. The name is already misleading. Neither Edward Sapir nor Benjamin Lee Whorf ever wrote down anything they called a hypothesis; the label was attached by Harry Hoijer after a 1953 conference on Whorf's work, and although Hoijer said he meant to clarify it, he never actually stated it. A claim nobody ever formally uttered became the most widely circulated idea in twentieth-century linguistics.
Three steps: how Whorf's hardest piece of evidence was taken apart; where the strong version breaks as a matter of logic; and what exactly is left.
Whorf's day job was fire prevention inspector for the Hartford Fire Insurance Company. He reported an observation that has been quoted thousands of times since: in the room where full gasoline drums were stored, nobody smoked; in the room labelled "empty gasoline drums", workers smoked and tossed the butts away. But an empty drum holds vapour and is easier to ignite than a full one. His reading: empty says "nothing inside" and, riding along with it, "harmless" — and people acted on the second meaning.
The example has a picture in it, but what it shows is one ambiguous word steering one action, a long way from "language shapes a people's worldview". What pushed the idea that far was his claim about Hopi, a language of Arizona: Hopi, he wrote, contains no words, grammatical forms, constructions or expressions that refer directly to what we call "time" — and so Hopi speakers have no intuition of time as a smooth flowing continuum.
In 1983 Ekkehart Malotki published Hopi Time: 677 pages doing one job, laying out Hopi temporal expression item by item — verb marking that distinguishes future from non-future, words for "yesterday / tomorrow / that day", expressions for spans and frequencies, constructions that count time as units. The book opens with Whorf's sentence on the left-hand page and, facing it, a sentence of Hopi:
Pu' antsa pay qavongvaqw pay su'its talavay kuyvansat, pàasatham pu' pam piw maanat taatayna. ("Then indeed, the following day, quite early in the morning at the hour when people pray to the sun, around that time then, he woke up the girl again.") Four layers of temporal placement in one sentence.
The dispute belongs here too. Malotki's critics say he is hitting a straw man: read in the original, Whorf sounds more like he is saying Hopi does not objectify time into something countable. That defence has force, but Whorf's sentence was written in absolute terms, and his Hopi material came largely from one native speaker interviewed in New York rather than from long fieldwork.
The more general lesson: "language L has no word for X" is almost never evidence by itself. As the piece on words being assembled put it, a language that "lacks" something has usually packed it somewhere else — into a grammatical marker, a fixed collocation, a whole sentence. German Schadenfreude has no single-character equivalent in Chinese, but Chinese has 幸灾乐祸 (xìng zāi lè huò, "delight at misfortune"), four characters assembled from parts. That is not an unthinkable thought. That is a different way of spelling it.
The strong version has a name, linguistic determinism: a distinction your language does not give you is one you cannot think. To test it you need evidence that does not run through language. Otherwise the argument closes into a circle: they think differently, the evidence being that they speak differently; and why do they speak differently? Because they think differently.
Watch how a still-cited example ends up. Grammatical gender: German Brücke "bridge" is feminine (die Brücke), Spanish puente "bridge" is masculine (el puente); for "key" it runs the other way, masculine in German (der Schlüssel), feminine in Spanish (la llave). Boroditsky and colleagues reported in 2003 that when people pick adjectives for these objects, German speakers choose "beautiful" and "elegant" for bridges while Spanish speakers choose "strong" and "towering". It sounds like grammar has tinted the objects.
In 2014 Mickan, Schiefke and Stefanowitsch ran two versions of that experiment with German speakers — a word-association task and a primed lexical-decision task — and failed to replicate it both times. Their conclusion: the original was either an artefact of some undocumented aspect of the procedure, or a statistical fluke. A more recent multi-language test points the same way: whatever the influence is, it is weak and it varies by language.
The point is that picking adjectives is still a language task. To step out of the circle you have to measure something that does not require speaking — where the eyes go, what gets remembered, which two things get grouped together. That is how the strong version died: not by one fatal experiment, but by shrinking a notch every time someone rewrote it into a shape that could be measured.
The other blow came from Berlin and Kay's Basic Color Terms (1969): across nearly a hundred languages, basic colour vocabularies emerge in a strongly constrained order — black and white first, then red, then green and yellow, then blue. Carving the spectrum arbitrarily should not produce that order. The book later drew its own criticisms over sampling, but it accomplished one thing at the time: it turned "language carves up the world arbitrarily" from a default assumption into a claim that needed proof.
In 1996 Dan Slobin gave the surviving piece a name: thinking for speaking. Language does not change how you think in general. What it specifies is which pieces of information you have to go look up when you are about to say something.
Take a concrete difference. Someone runs into a room. The typical English way to say it is He ran into the room — manner (running) rides on the verb, path (in) hangs off the preposition. The typical Spanish way is Entró en la habitación corriendo, literally "he entered the room running" — path rides on the verb and manner drops to an optional tail. Berman and Slobin, in 1994, had speakers of several languages narrate the same wordless picture book, Frog, Where Are You?: English narrators nearly always said how the motion happened, narrators in the Spanish-type languages often left it out.
So far this is only a difference in what gets said. The decisive step was taken by Papafragou, Hulbert and Trueswell in 2008. Greek belongs to the same type as Spanish, with path on the verb. They showed Greek and English speakers identical motion clips while recording eye movements, under two instructions: you will have to describe this clip, or you will have to recognise this clip later.
The result splits cleanly in half. In the description task the two groups' eyes parted within the first second after motion onset, each favouring earlier and more looks at the component their language habitually encodes. In the memory task — where people are simply free to look — attention was not allocated differently by language.
The shape of that matters: the effect is real and visible at the millisecond scale, but it is online and task-bound. Language is not a pair of glasses you cannot take off. It is a checklist you pull out in order to speak — put the checklist away and you and everyone else are looking at the same world.
Evidence strength: the Papafragou study is eye-tracking with a between-language design, and later work on motion events broadly supports this task-dependent pattern; Slobin's material is a comparison of narrative corpora, not a randomised experiment. This whole weak-version area has been swept by the replication crisis too, and early reported effect sizes generally ran large.
Do not read the surviving piece as smaller than it is, or larger. It explains why translation always drops something, and why "give it a name" really does change what a team notices afterwards. What it cannot carry is the loudest sentence of all — that speakers of some language cannot think some thought.
That is a real risk: "language influences some cognition under some conditions" is nearly impossible to overturn. The way to give it content again is to bet on the boundary in advance — say which tasks it should show up in and which tasks it should vanish from. The Papafragou study counts as hard precisely because it staked a bet it could lose: if the two groups had also diverged in the memory task, task-boundedness would have been refuted on the spot.
This is the most brittle part of the story. If someone spends most of their waking hours talking inside their head, "task-bound" is not as limiting as it sounds. It is checkable: find the people who report having almost no inner monologue and see whether their attention in a free-viewing task looks like everyone else's. The trouble is that this can only be obtained by self-report, and self-report is itself language — which walks straight back into the circle from the section on not being able to prove thought from speech.
The most promising line in the weak version bets exactly here: obligatory is likelier to leave a trace than available. Turkish has a system of evidential marking — saying something commits you to indicating whether you witnessed it or heard about it, and leaving it unmarked is ungrammatical; English needs an added phrase like "apparently" to do the same work. If the expectation holds, the difference should show up in what must be said rather than in what can be said. Measure in the wrong place and naturally you find nothing.