06 · The Two Blues of Russian

In 2007, at MIT, twenty-one Russian speakers judged pairs of blue that straddled the line between goluboy and siniy faster than pairs sitting on the same side of it: 1,164 milliseconds against 1,288.

A difference of 124 milliseconds. That is the true size of the effect this book will accept, and it is far smaller than what has been claimed of it. So this chapter prints what replicated and what did not on the same pages. Print them apart and the reader gets the number without its boundary.

In 1969 the colour question had already been answered once, and answered the other way. Brent Berlin and Paul Kay published Basic Color Terms: Their Universality and Evolution, surveyed 98 languages, and proposed that every language draws its basic colour terms from a universal set of eleven categories, admitting them to the lexicon in a broadly predictable order. The proposal was aimed straight at the strongest reading of Whorf: the belief that each language carves the spectrum as it pleases.

What has survived to today is not the eleven-stage sequence exactly. The World Color Survey that followed — 110 languages, and the dataset cited by both Winawer et al. 2007 and Kay & Regier 2007 — pushed the fixed order into a statistical tendency. The surviving conclusion is weaker, but it is not empty: some force presses wholly unrelated languages to name colours in the same direction.

And the two sides do not really contradict each other. In their own 2007 reanalysis, Kay and Regier hold two things at once — colour boundaries are not set arbitrarily, and colour boundaries correlate with perceptual differences between speakers of different languages. The field's principal researchers call the "universalist versus relativist" frame a false dichotomy.

Russian has no neutral word for blue. A speaker must choose between goluboy for light blue and siniy for dark blue, every time, not only when precision is wanted. English carries no such obligation. Winawer, Witthoft, Frank, Wu, Wade and Boroditsky took that difference into the laboratory and published it in PNAS 104(19):7780–7785 in 2007.

The participants were 26 native Russian speakers and 24 native English speakers, recruited in Boston and tested at MIT; after exclusion for missing more than 25% of trials, 21 remained on each side (3 English, 5 Russian excluded). The task was to look at colour squares and say, as fast as possible, which of the two options matched the target. It ran in three forms — no interference, a digit string held in mind, and a spatial grid to remember.

The result is much narrower than the retelling. Russian speakers were indeed faster when a pair straddled the goluboy/siniy line, but only for pairs close enough to be hard to tell apart, and only when there was no digit string to hold. The category × interference interaction came out at F(2,40) = 5.3, p = 0.009. English speakers showed no category advantage under any condition: F(1,20) = 0.150, p = 0.703.

The sharpest number is how much of that advantage the digit string ate. On the Russian side, 173 milliseconds; on the English side, 14 [t(40) = 2.142, p = 0.038]; and in a parallel comparison, 189 against 15 milliseconds [t(40) = 2.17, p = 0.036]. Within the Russian group alone, the advantage on near colours was significant with no interference (1,164 against 1,288 milliseconds; t(20) = 2.59, p = 0.0176) and under spatial interference (1,162 against 1,270 milliseconds; t(20) = 2.18, p = 0.041). With a digit string to hold, it was not significant.

The authors themselves concluded something narrower than their readers passed on. The effect is small, it depends on the task — only hard discriminations, only under speeded conditions that block thinking in words — and it is generated on the spot. Language is recruited during the judgment and can be switched off by occupying the verbal channel. This is not a visual system permanently retuned.

Supporting work exists, but it does not stand as far off as it should. Winawer et al. cite three studies using the same triad paradigm — Witthoft et al. 2003 (conference proceedings), Drivonikou et al. 2007 (PNAS 104:1097–1102), and Gilbert, Regier, Kay and Ivry 2006 (PNAS 103:489–494), whose own title reports the effect in the right visual field and not the left. The two laboratory lines, Kay/Regier/Ivry and Boroditsky/Winawer, overlap. The corroboration is real, but it is not corroboration from outside.

Here the other half has to be printed. Several other studies in this same line did not replicate — and the ones that did not replicate are the ones retold most often.

Grammatical gender does not change how people describe objects. That is the verdict after two replication attempts. The original claim comes from Boroditsky, Schmidt and Phillips 2003, "Sex, Syntax, and Semantics," in Gentner and Goldin-Meadow's Language in Mind (MIT Press). In German, key is masculine, and German speakers described keys in English as hard, heavy, jagged. In Spanish the same word is feminine, and Spanish speakers described them as beautiful, elegant, tiny. Its companion study was Phillips and Boroditsky 2003, on similarity judgments between pictures of people and pictures of gendered objects. At the time they were replicated, neither had passed peer review in a journal: one sat in a book, the other in conference proceedings.

The first time anyone tried to replicate it, nothing came out. Mickan, Schiefke and Stefanowitsch 2014, "Key is a llave is a Schlüssel: A failure to replicate an experiment from Boroditsky et al. 2003," in the Yearbook of the German Cognitive Linguistics Association, ran two separate attempts — a word association task and a primed lexical decision task — with German and Spanish speakers. Neither recovered the original effect. The authors were left with two options: the original was an artefact of undocumented procedural detail, or it was a statistical fluke.

The second attempt was preregistered and better powered, and still nothing came out. Elpers, Jensen and Holmes 2022, "Does grammatical gender affect object concepts? Registered replication of Phillips and Boroditsky (2003)," in the Journal of Memory and Language 127:104357, collected a total N = 375. Once the appropriate sources of error variance were modelled, Spanish and German speakers' similarity judgments showed no gender-congruency effect at all. The original pattern surfaced only in a separate control condition, where English speakers were explicitly trained on artificial categories built to mimic gender. That is an effect of training, not of a native grammar.

Mandarin's vertical time fell the same way. Boroditsky 2001, "Does Language Shape Thought? Mandarin and English Speakers' Conceptions of Time," in Cognitive Psychology 43(1):1–22, ran three experiments — the first with 26 English-speaking and 20 Mandarin-speaking students, the second with 25 bilingual students, the third with 70 native English speakers — and reported that Mandarin-English bilinguals think about time vertically more than English monolinguals do. The reason offered was that Mandarin has additional vertical spatial metaphors for time.

Chen 2007, in Cognition, titled the paper "Do Chinese and English speakers think about time differently? Failure of replicating Boroditsky (2001)," and found no reaction-time difference of the predicted kind. Worse for the original was the corpus analysis in the same paper: Chinese speakers use horizontal time metaphors more often than vertical ones, which pulls the floor out from under the explanatory mechanism. Several further replication attempts are named in this literature — January & Kako 2007, Tse & Altarriba 2008, Chen & O'Seaghdha 2013 — but this book's dossier did not retrieve and check the texts itself, so they are printed as the citing literature reports them, not as independently confirmed.

And colour-memory effects in general, weighed across the whole set, are not reliable. Wright, Davies and Franklin 2015, "Whorfian effects on colour memory are not reliable," in the Quarterly Journal of Experimental Psychology 68:745–758 (DOI 10.1080/17470218.2014.966123), gathered cross-linguistic and developmental data and used Bayesian model comparison to weigh "Whorfian effect present" against "Whorfian effect absent" directly. The evidence consistently favoured absent. This is the paper that makes the present chapter most uncomfortable, because it does not knock down the opposing camp's work. It knocks down the line siniy and goluboy stand on.

The tidiest explanation for statistics scattered like this comes from Regier and Xu 2017, in WIREs Cognitive Science 8:e1440. What decides whether an effect appears is not language but uncertainty in the non-linguistic representation. Tasks answerable with confidence — easy, simultaneous, no memory load — show almost nothing; uncertain tasks — delayed, difficult — give larger effects. The two of them offer this as an explanation of a mixed replication record, not as a new experimental finding, and this book prints it that way.

So the size of the real effect can be stated only this far: small, task-dependent, and visible only under time pressure, or when something else has taken the verbal channel. Outside those conditions it disappears; and weighed across the whole set the way Wright, Davies and Franklin weighed it, it does not stand. Anyone claiming that speakers of different languages live in different worlds will find no number in this chapter to borrow. A hundred and twenty-four milliseconds is not another world.

But a translator works in exactly those conditions: one sentence at a time, under time, at the boundary where two options are nearly the same — precisely where this effect shows up, on every count. When a Russian sentence obliges a choice between goluboy and siniy and the English sentence can take only blue, what goes missing is not the shade. What goes missing is the necessity of choosing.

There is no aggregate number for this field to print. This book's dossier searched for a meta-analysis quantifying an average effect size for linguistic relativity as a whole and did not find one. No such number appears on this page, and none should appear anywhere without a source.

What could not cross: Russian obliges its speaker to place every shade of blue on one side or the other of the goluboy/siniy line before the sentence can be finished. English does not oblige it, and cannot receive it — "light blue" restores the shade but not the necessity of choosing, and what is not necessary a translator may always drop.

Overturned if: a laboratory unconnected to both the Boroditsky/Winawer and the Kay/Regier lines preregisters and repeats the Winawer et al. 2007 triad task at higher power. Recover the category × interference interaction and this page grows firmer. Fail to recover it and this page joins grammatical gender and vertical time on the same pile. Either result overturns this chapter.

Sources: Winawer, Witthoft, Frank, Wu, Wade & Boroditsky 2007, PNAS 104(19):7780–7785 — http://lera.ucsd.edu/papers/pnas-2007.pdf · Berlin & Kay 1969, Basic Color Terms — https://press.uchicago.edu/ucp/books/book/distributed/B/bo3643648.html · Kay & Regier 2007, Cognition 102:289–298 — https://lclab.berkeley.edu/papers/berinmo-cognition-in-press.pdf · Mickan, Schiefke & Stefanowitsch 2014 — https://www.degruyterbrill.com/document/doi/10.1515/gcla-2014-0004/html?lang=en · Elpers, Jensen & Holmes 2022, Journal of Memory and Language 127:104357 — https://www.reed.edu/psychology/thought-lab/assets/publications/Elpers,%20Jensen,%20_%20Holmes%202022%20-%20JML.pdf · Phillips & Boroditsky, "Effects of Grammatical Gender on Object Description" — https://escholarship.org/uc/item/3j37f380 · Boroditsky 2001, Cognitive Psychology 43(1):1–22 — http://lera.ucsd.edu/papers/mandarin.pdf · Chen 2007, Cognition — https://pubmed.ncbi.nlm.nih.gov/17070793/ · Wright, Davies & Franklin 2015, QJEP 68:745–758 — https://www.tandfonline.com/doi/abs/10.1080/17470218.2014.966123 · Regier & Xu 2017, WIREs Cognitive Science 8:e1440 — https://terry-regier.github.io/lclab/papers/whorf-wires-preprint-v2.pdf