Does language shape thought? Do the languages we speak affect how we live our lives? These are some of the oldest questions in the cognitive and social sciences, and most everybody reading these words probably has thought about them. These also speak to core questions in the philosophy of mind and language, on the role of language in intermediating between our brains and the world around us.
An emerging literature in the social sciences has given these questions renewed prominence by arguing that language systematically affects people’s values, beliefs, and behaviors. And in a forthcoming article in Language (preprint here [PDF]), I subject this prominent new literature to conceptual and empirical scrutiny. What follows is an expanded summary of my argument, accompanied by a discussion of why it matters, based loosely on this Twitter thread.
Linguistic Relativity and the Sapir-Whorf Hypothesis
You may have encountered the idea that language shapes reality in some form or another before. If you learned about it in a university setting, you probably heard it described as the Sapir-Whorf hypothesis; more generally, however, linguists, anthropologists, and those working in adjacent fields use the term “linguistic relativity” to describe a set of related theses about human language. Linguistic relativity means, loosely, that—at least to some degree—the language that you speak affects how you perceive and interact with the social world. For example, Arabic-speakers and Japanese-speakers experience the world differently because their languages differ. Do note that linguistic relativity is a precise claim: it is not just that Arabic speakers and Japanese speakers differ, or that communities of Arabic speakers and communities of Japanese speakers differ. Linguistic relativity holds that in some specific way, the differences between the Arabic language and the Japanese language are responsible for some of those differences between Arabic and Japanese speaking communities.
French versus English pronouns help to illustrate what I mean. English has one second-person singular pronoun: you.* French as two second-person singular pronouns: tu and vous. The two words, in French, differ in their familiarity, formality, or politeness. You’d address a stranger with vous, a friend with tu, a teacher with vous, a child with tu, and so forth.
The linguistic relativity thesis holds that this is not just a quirk of French that English speakers have to master. A strong proponent of linguistic relativity would predict that French people will be more attuned to rank, status, and hierarchy than English speakers because speaking French obliges one to attend to tu/vous distinctions in order to be communicatively competent. Of course lots of other things would explain speaker-community-level differences in attention to hierarchy; language isn’t everything. But counterfactually, English would be even more status-conscious if it also had a tu/vous distinction as well.
It is hard to think of a single research area that is so equally exciting for philosophers, experimental psychologists, ethnographers, literary theorists, and development economists. Think of the implications! Are certain language communities going to be better at certain cognitive tasks than others? Are there linguistic roots of hierarchical political systems? Is translation even possible? And it’s not just academics: growing up, I once heard a family friend announce to my parents that “the French are assholes” because of their attention to hierarchy, and used the tu/vous distinction as evidence.**
Although people have wondered about these sorts of questions for a long time, the history of linguistic relativity as a field of scientific inquiry dates roughly to the mid-19th century, concomitant with the emergence of anthropology as a scientific discipline. One common feature of mature scientific disciplines is that they began through various procedures and claims that we today would classify as pseudo-science.*** Alchemy to chemistry, astrology to astronomy, phrenology to physical anthropology, the list goes on.
And so too with the pseudo-scientific origins of what I consider to be a properly scientific field of inquiry on linguistic relativity. Linguistic relativity as a scientific claim developed in the context of the encounter between Europeans and the Americas and with Africa, in a time of settler colonialism and imperial race science. Although mine is not a systematic excavation of the intellectual history of linguistic relativity, it seems fairly plain that it emerged in conversation with views about “primitive peoples” versus “civilized peoples,” and capturing the notion that primitive peoples were primitive because they spoke primitive languages.
One important intellectual figure in the development of the (pseudo-) science of linguistic relativity was Wilhelm von Humboldt, a German linguist who made important discoveries about Basque and Malayo-Polynesian languages. He postulated that language is a human universal, and it is the “organ” that makes thought possible. Reality exists outside of our thoughts, but it is only through language that we are able to translate the world “out there” into thought. And because languages vary, he argued, it must follow that languages produce different thought-systems that will, naturally, vary across linguistic communities.
This is not an inherently racist position. The idea that thought is only possible through language is straight Wittgenstein, for example (see here [PDF] for more). But such a position is easily imported into a pseudo-scientific racist worldview as a way to justify a view of certain languages as inherently superior to others, better suited to scientific reasoning or logical exposition than others. (Guess which languages were considered the best for logic and science.) It also can be used to justify the act of creating modern or civilized people by raising them with a modern or civilized language. It is not hard to see how linguistic relativity might be used to argue that an Anishinaabe child can be made into a proper Canadian by replacing their Ojibwe thoughts with English thoughts.
This period in the intellectual history of linguistics and anthropology is endlessly fascinating: Boas, Sapir, Bloomfield, and others built linguistics and anthropology alongside one another. So much of making anthropology in the Americas was the task of learning and classifying the languages of the Americas. They worked both with and against Saussure and the European structuralists like Lévi-Strauss. The lineages are complex and it is sometimes hard to figure out who disagrees with whom about what, exactly. The arbitrariness of the sign unites basically everyone, but from there, just about anything goes.
In any case, by the 1950s Noam Chomsky took linguistics in a new direction by postulating a Universal Grammar inherent to all human languages, reformulating the problem as to explain the diversity of human language from the postulated common linguistic capacities that all humans possess. The limitless generativity**** of human language suggests that we must all possess some common ability to think through language.
In this now-standard view, the null hypothesis is that languages vary but that their phonological inventories, syntactic structures, semantic distinctions, and conversational pragmatics do not fundamentally affect human cognition.
With this history in mind, you can imagine how shocking it is for linguists to hear non-linguists arguing about the effects of language variation on thought. For example, on the TED radio hour on NPR, we have this long piece on the effect of having a grammatical future on savings rates across countries. The argument, in the words of author Keith Chen, is that
You speak English, a futured language, and what that means is that every time you discuss the future or any kind of a future event, grammatically, you’re forced to cleave that from the present and treat it as if it’s something viscerally different. Now suppose that that visceral difference makes you suddenly disassociate the future from the present every time you speak. If that’s true, and it makes the future feel like something more distant and more different from the present, that’s going to make it harder to save.
The research that underlies this claim appears in the American Economic Review, the flagship journal of the premier economics disciplinary association in the world.
Most Everybody Believes Some Version of Linguistic Relativity…
Before we dig into that kind of research, though, let’s step back to consider what we know about linguistic relativity. The Chomskian position does not rule out linguistic relativity completely. And indeed, at some basic level, some version of linguistic relativity seems almost certain to be true. For a simple example, take the word for the color blue. In English, we look at these two colors and call them both blue.
But in Russian, these are two different colors: goluboy/голубой (light blue) and siniy/синий (dark blue). You can’t go wrong calling either blue in English, but using siniy to describe light blue would be like calling the sky green. And it turns out that Russian speakers can do certain tasks related to classifying blue colors faster than English-speakers. This article in The Proceedings of the National Academy of Science of the United States of America finds that
….Russian speakers were faster to discriminate two colors when they fell into different linguistic categories in Russian (one siniy and the other goluboy) than when they were from the same linguistic category (both siniy or both goluboy).
This means that Russian-speakers go about their world a little bit differently than English-speakers: not just different words, but different competences. This finding tells us that our language affects our reality in some small yet precise way. Yet disambiguating among shades of blue isn’t really a socially meaningful activity. Most linguists don’t think that linguistic differences like this explain too much about our everyday lives, or gross patterns in human behavior or cognition across speech communities.
My favorite example of linguistic relativity that matters comes from speakers of Guugu Yimithirr, a language of Far North Queensland (and source of the word “kangaroo” in English). Guugu Yimithirr is probably unique among the world’s languages in that—at least in some analyses—it possesses no egocentric directions, like left/right or in front/behind. This means that to describe where something is, you have to reference cardinal directions like North/South, or “towards the mountains” or “towards the sea”. For example, at this very moment the computer that I’m typing on is to the West of me. If I turn my back to it, it is still to the West of me. In English, though, one would say that it was in front of me, and then it was behind me.
There is evidence that Guugu Yimithirr speakers are better at dead reckoning and solving maze puzzles than are speakers of related languages who do possess egocentric directions and who live otherwise similar lives. If this evidence is sound, then this is linguistic relativity in action. But I hasten to add that the analysis of Guugu Yimithirr as having only absolute frames of reference is contested (see here [pdf]). One worries about how we interpret linguistic and anthropological data about languages that seem “exotic” to Western audiences.
The big takeaway is that linguistic relativity is an interesting and relatively well-defined domain of research in linguistic anthropology and cognitive science, but its empirical claims are very narrow and very precise and there are disagreements even about those. There is a very good edited volume called Rethinking Linguistic Relativity that is over two decades old, but which summarizes the state of this literature for anyone who really wants to understand it. Dan Slobin’s chapter (pdf) on “thinking for speaking” is probably the most sophisticated reformulation of linguistic relativity into a psycholinguistic research program, but it does not make strong arguments about variation across human communities’ cognitive capacities.
…But Linguists (and Others) are Inherently Skeptical of Whorfian Socioeconomics
With that background, we now turn to the latest iteration of the linguistic relativity research program. In the past twenty years or so, a number of social scientists have embraced linguistic relativity to argue that language is a fundamental determinant of human behavior across the world. Examples? What if people who speak languages with grammatical gender in indefinite articles are more sexist than the rest of us? (Compare English the to German der/die/das.) What if people with obligatory politeness distinctions in pronouns (tu/vous) are less egalitarian than the rest of us? What if people who speak future oriented languages are more likely to care about the environment than the rest of us?
These are big claims. And to distinguish them from linguistic relativity as a research paradigm, I term this emerging literature “Whorfian socioeconomics.” You can think of it as linguistic relativity for the big leagues. Language affects no only how we complete certain cognitive tasks, but how we go about our lives—and when aggregated up to the level of the speaker community, how whole societies are oriented.
In my read, however, the evidentiary base for Whorfian socioeconomics is very thin. And in my forthcoming article, I review the evidence and the tools that people use to make such claims, showing it to be weak.
The first source of evidence for Whorfian socioeconomics is aggregate evidence correlating the presence or absence of a linguistic feature in a country or region’s dominant language with some outcome like gender discrimination, authoritarianism, or savings rates. This research is plausible, but to give any such correlations as causal interpretation, we need strong assumptions about ecological validity and unconfoundedness.
The second source of evidence is experimental: take bilingual speakers, randomly assign them to hear something or do something in two different languages, and see if their behaviors differ. Although experiments are indeed the gold standard for establishing causal relationships, such experiments are inherently limited by the multiple dimensions across which languages vary. In the language of causal inference, language is a bundled treatment, which means that if you compare two languages that differ on some characteristic D, you are also comparing languages that differ by characteristics A, B, C, … and so forth. Differences between languages might be a product of any of those characteristics.
For example, this article compares Russian and Estonian among bilingual speakers to see if speaking a futured versus futureless language affects time preferences. But to put it mildly, Russian and Estonian differ in a lot of ways. Russian has gendered pronouns, Estonian does not. Russian has six cases (in most analyses), Estonian has fourteen. Russian has no strict stress pattern, but stress in Estonian words is always on the first syllable. Russian allows for zero-copula for predicate nominals, Estonian does not. Russian is an Indo-European language, Estonian is a Uralic language. The list goes on.
In all, this means that a comparison of Russian and Estonian cannot isolate future marking from any other difference between these very different languages. A good experiment is one that can isolate the causal factor of interest by varying it while holding everything else equal in expectation, and that is very hard to do when comparing languages. If you want to argue that grammatical tone affects how we hear music, an experiment would need to compare a tonal language like Mandarin with a version of Mandarin that is structurally identical but which has no tones. Such comparisons are impossible because no pair of languages in the world is identical except for exactly one linguistic feature.
Separately, the claims of linguistic relativity become very odd when applied to fluent bilinguals. Remember the premise: language affects one’s social reality. So what social reality to fluent bilinguals inhabit? By assumption they operate with full competence in both realities, or they wouldn’t be fluent. So what is it that we learn from the observations that bilinguals differ based on what language they happen to be speaking? It must be that the language that one speaks in real time affects thought in real time. That is a Slobin-compliant position (see above)—but Slobin is not the theoretical foundation for Whorfian socioeconomics!
The only truly dispositive experiment would be the completely unethical experiment which randomly assigns babies to be raised in controlled conditions where they learn languages that are manipulated to vary by one dimension only. Not only is this unethical and practically impossible, it also raises questions about in what sense such a child would be a fully-formed social individual in our normal understanding of what makes us social beings. Tragic stories of so-called feral children show us the implications.
Returning to the point at hand, though: the third source of evidence used in Whorfian socioeconomics is a mix between experiments and aggregate data: crossnational surveys that record lots of features of people’s lives as well as the languages that they speak, for lots of people all over the world. If you can code all the world’s languages as to whether they possess a particular linguistic feature or not, then you can generalize across languages that vary in lots of ways to isolate the relationship that you care about, net of other differences across languages that also may vary.
The problem in this case is that statistical modeling decisions are paramount, and in my article, I show that the standard methodologies used when authors make such arguments are flawed. Specifically, they are very likely to uncover spurious statistical associations that look like evidence that linguistic features explain lots of variation across human populations. Some of you might enjoy learning just how easy it is for such an analysis to falsely discover completely outlandish linguistic effects:
I also review some simple statistical fixes that are nearly costless to implement but which address the problems inherent in such analyses. Employing these tools, I can replicate that Ted Radio Hour/American Economic Review article and show that its results do not hold when one accounts for uncertainty properly.
This Debate Matters
Clearly, linguistic relativity remains an important area of research, and Whorfian socioeconomics will almost certainly continue to be of interest to social scientists interested in cataloguing the social, political, and economic correlates of human language variation. But what I did not cover in my Language article are the normative implications of Whorfian socioeconomics and linguistic relativity.
Practically, of course, we should want to get the science right. Linguistic relativity might explain some subtle differences across speaker communities. But because these effects are probably very small, it should difficult to find them, and social scientists should accordingly start by engaging more with the voluminous post-Chomsky linguistics literature to see why linguists are so skeptical of these ideas.
But hold that aside. Normatively, I think this it is dangerous to argue that language affects our social world.
For reasons why, you can look back to the use of linguistic relativity to justify the practices of colonialism and imperial race science. Arguments that have been used to view certain peoples as primitive, and their languages as representing a barrier to their entrance into modernity, will continue to be used in this way—dressed up in new garb, applied to new problems, but with the same erroneous presumption that language constrains thoughts in socially meaningful ways.
The best articulation of this position is a fantastic (and underappreciated) book by John McWhorter entitled The Language Hoax. He argue that what he calls “Neo-Whorfianism”—which I see as the general theoretical idea that undergirds Whorfian socioeconomics—is not only plainly very wrong, but dangerous. Here’s a link to a long lecture by McWhorter that summarizes his argument, and here’s in illustration from his introduction:
Take the intransigent ultranationalist German historian Heinrich von Treitschke. Prussophile, xenophobic, and nakedly anti-Semitic, he was given in the late nineteenth century to insights such as “differences of language inevitably imply different outlooks on the world.” You can imagine the kinds of arguments and issues he couched that kind of statement in, and yet the statement itself could come straight out of Whorf, and would be celebrated as brain food by a great many today. “Surely,” after all, “the question is worth asking…”—yet somehow, we would rather von Treitschke hadn’t, and find ourselves yearning for thoughts about what we all have in common.
And that is the point. Language really is one of the only things that makes us all equally and equivalently human. Language is, for many working in the philosophy of mind, the quintessence of what it means to be human. The languages that we speak may differentiate the people with whom we can communicate, but we are united in that we all use language to do so. It is entirely certain that a speaker of Guugu Yimithirr could convey to me the meaning of a sentence like “my mother’s uncle who is seated on my left has just seen an airplane that reminds him of a crocodile,” which is a sentence that no one in human history has ever uttered, and yet we can both fully comprehend. And so can you, no matter who you are or where you live!**** That is miraculous.
There are no objectively good and bad languages, nor strong and weak languages. Nobody’s thoughts are crippled by the language that they speak, and nobody’s reality is fuller because their language makes it so. As humans, we should root against Whorfian socioeconomics because it promises that our languages imprison us in realities that divide us. The good news is that the science doesn’t support that.
* We used to have more: you versus thou (nominative) and thee (accusative).
** I don’t think that “the French are assholes,” but I remember this very clearly because this family friend was so adamant that pronouns were part of it.
*** Yes, the demarcation problem is real, and it is spectacular. So too is the social nature of the scientific enterprise. Acknowledging and moving on.
**** Attention Cornell PhD students who have noticed how I like to signal my approval of something by calling it “generative”: now you know where this turn of phrase comes from! It’s ultimately a Chomsky reference.
***** You may respond “but what if I’m from some community that has never encountered airplanes or crocodiles?” And I would respond “I’ve never seen a unicorn or a miracle either, but they can be explained to me.”