After Babel: Sapir, Whorf and the Multilingual Mind of Artificial Intelligence

By Matthew Parish, Associate Editor

Saturday 22 August 2026

One of the oldest puzzles in the philosophy of language has unexpectedly acquired a new experimental subject. For much of the twentieth century philosophers, linguists and anthropologists debated whether the language in which human beings speak affects the manner in which they think. The controversy became associated with Edward Sapir and Benjamin Lee Whorf and eventually acquired the slightly misleading title of the Sapir–Whorf hypothesis. Today the question has returned in a remarkable form. Large language models can converse in English, Ukrainian, French, Arabic, Chinese, Japanese and scores of other languages. They can translate between them, reason within them and apparently move from one linguistic conceptual framework to another in fractions of a second. What, then, is happening to their conceptual worlds as they do so?

The question is philosophically important because large language models may provide something that Sapir and Whorf could never have imagined — a laboratory in which something resembling the same cognitive architecture can be asked the same question repeatedly in different languages. That does not prove anything directly about human consciousness. Large language models are not human beings and their internal representations should not casually be identified with human thoughts. Nevertheless they present an extraordinary opportunity to reconsider the relationship between language, concepts and reasoning.

The expression “Sapir–Whorf hypothesis” itself conceals several different propositions. Neither Sapir nor Whorf formulated a single hypothesis bearing their joint names and contemporary scholarship distinguishes sharply between strong and weak interpretations of linguistic relativity. The strongest interpretation — sometimes called linguistic determinism — suggests that the structure of a person’s language determines the structure of his thought. On this account a person speaking a language lacking some conceptual distinction might literally be incapable of thinking the relevant thought.

Few contemporary philosophers or linguists accept such a strong proposition. Human beings plainly learn foreign languages, invent new vocabulary and acquire unfamiliar concepts. Scientific revolutions themselves frequently involve creating new terminology for phenomena that previous generations could not adequately describe. If language absolutely imprisoned thought then conceptual innovation would be difficult to explain.

A weaker Whorfian proposition is considerably more plausible. Language may not determine what we can think but it may influence what we habitually notice, distinguish and emphasise. Languages divide experience differently. They classify objects using different grammatical genders, describe space using different frames of reference, divide colours into different lexical categories and encode tense, aspect, evidentiality, politeness and agency in different ways. Speaking a language repeatedly requires speakers to make some distinctions that speakers of another language may leave implicit.

The difference is profound. The strong Whorfian thesis says that language constructs the prison of thought. The weaker thesis says that language furnishes the rooms in which thought usually lives.

From Whorf to contemporary philosophy of language

This controversy intersects with several of the central disputes of twentieth-century philosophy. One concerns whether meaning consists fundamentally in representing an independently structured world or whether linguistic practices themselves contribute to the organisation of experience. Another concerns whether concepts exist independently of natural language. Yet another concerns whether translation between languages can ever be perfectly determinate.

Here Whorf unexpectedly encounters Ludwig Wittgenstein, W. V. O. Quine and Donald Davidson.

The later Wittgenstein’s insistence that meaning is embedded in linguistic practices — in “forms of life” — makes language inseparable from the activities of communities that employ it. Words do not acquire meaning merely by attaching labels to pre-existing objects. Their significance arises from their use within networks of human practices. The implication is not necessarily Whorfian determinism but it is decidedly hostile to the idea that language is a perfectly transparent window through which an independently constituted conceptual world may simply be observed.

Quine pushed the problem in another direction. His celebrated thesis of the indeterminacy of translation suggested that behavioural evidence might be compatible with multiple translation manuals between radically different languages. There might be no uniquely correct fact about whether an unfamiliar expression meant “rabbit”, “undetached rabbit part” or something else constructed from an alternative ontology. Translation therefore exposed questions not merely about vocabulary but about the conceptual schemes through which the world is described.

Davidson famously attacked the very idea of radically incommensurable conceptual schemes. If another linguistic community were genuinely organising reality according to a conceptual system utterly alien to ours, Davidson argued in effect, it would become difficult to explain how we could recognise its utterances as language at all. The possibility of interpretation presupposes enormous areas of agreement.

Large language models make this philosophical disagreement strangely concrete.

The machine that speaks in many tongues

A contemporary multilingual LLM does not ordinarily contain an entirely separate artificial intelligence for every language it speaks. Multilingual models develop representations allowing substantial transfer between languages, although the degree of competence and alignment varies considerably and contemporary systems remain disproportionately shaped by English and other high-resource languages. A 2025 survey of multilingual LLMs emphasised both their increasingly broad multilingual abilities and the continuing problem of English-centric training.

Suppose one asks a model a philosophical question in English:

What is justice?

Then one asks:

Що таке справедливість?

Then:

Qu’est-ce que la justice ?

Then the equivalent question in Arabic, Mandarin or Japanese.

At one level these appear to be straightforward translations. Yet the intellectual histories attached to justice, справедливість, justice and their counterparts are not perfectly identical. Each expression inhabits centuries of literature, law, religion, political argument and everyday usage. Their semantic territories overlap enormously without necessarily coinciding.

The LLM has encountered these words within precisely those different textual environments.

This produces a fascinating possibility. The model may possess something analogous to a shared conceptual spacebeneath its different linguistic outputs while simultaneously having language-specific pathways through that space. If so, neither strict Whorfianism nor simplistic linguistic universalism adequately describes what is occurring.

Research is beginning to provide evidence for precisely this more complicated picture. A 2025 study examining ChatGPT-4o mini across thirteen typologically diverse languages reported statistically significant semantic differences between responses to culturally salient prompts in different languages. The author interpreted this as evidence that linguistic structures measurably shape AI-generated interpretation.

Even more intriguingly, research on bilingual causal reasoning has found both language-specific reasoning patterns and convergence towards shared semantic representations. In Chinese and English, models displayed different attention and ordering preferences associated with the linguistic structures of the respective languages. Yet when reasoning succeeded their internal representations tended to converge towards semantically aligned abstractions.

That combination is philosophically remarkable.

The machine appears capable of being Whorfian and anti-Whorfian at the same time.

A universal conceptual substrate?

One interpretation is that multilingual LLMs provide evidence for something resembling a distinction between language and thought.

There may be internal computational representations that are not themselves straightforwardly English, Ukrainian or Chinese. Different natural languages provide alternative routes into and out of those representations. If this is correct then translation becomes possible because two linguistically different sentences can activate sufficiently similar regions of an underlying representational structure.

This resembles a computational version of an ancient philosophical dream — a lingua mentalis, or language of thought, beneath particular spoken languages.

Yet matters cannot be quite so simple. LLMs acquire their representations through language. Their conceptual structures are not implanted independently and subsequently given English or Arabic labels. They emerge statistically from exposure to enormous quantities of linguistic material.

Hence the supposed universal conceptual substrate is itself a sediment deposited by many languages.

This creates an inversion of the traditional Sapir–Whorf problem. With human beings we ask:

Does language shape thought?

With large language models we might instead ask:

What sort of thought-like representational structure emerges when a single system is shaped simultaneously by many languages?

The answer may eventually transform the philosophy of language.

The multilingual advantage

Whorf himself may have been more subtle on this question than the caricature of linguistic determinism suggests. Scholarship recovering what has been called his “lost argument” emphasises his interest in multilingual awareness — the possibility that encountering different languages reveals conceptual assumptions that monolingual speakers otherwise mistake for universal features of reality.

This is where large language models become particularly interesting.

A monolingual English speaker may find certain distinctions natural simply because English makes them easy to express. A multilingual person occasionally experiences the startling discovery that another language packages experience differently. There are expressions easily stated in one language that require cumbersome paraphrase in another. There are emotional, social and grammatical distinctions that become conspicuous only when languages are compared.

An LLM potentially performs this comparison continuously.

When translating between Ukrainian and English, for example, it cannot merely replace words mechanically. It must navigate differences in aspect, register, idiom, cultural implication and syntactic expectation. Translation becomes an act of approximation between partially overlapping semantic structures.

The greater the number of languages incorporated into a model, the richer this comparative space potentially becomes.

There is therefore a paradoxical possibility: the most linguistically capable artificial intelligence may become less Whorfian precisely because it has absorbed so many Whorfian worlds.

No single language imprisons it because it inhabits dozens simultaneously.

The return of conceptual schemes

This possibility revives the debate over conceptual schemes in a new form. Imagine a model trained predominantly upon English. Its conceptual organisation may inherit countless distinctions characteristic of Anglophone discourse. Now imagine another model trained equally across one hundred languages belonging to radically different linguistic families.

Would their internal representations of reality be the same?

Almost certainly not exactly.

The second model would repeatedly encounter alternative ways of dividing semantic space. Concepts represented by single words in one language would appear as phrases in another. Grammatical distinctions obligatory in some languages would be optional in others. Metaphors commonplace in one culture would be peculiar elsewhere.

The resulting representational system might therefore resemble not a single conceptual scheme but an intersection of conceptual schemes.

This suggests an intriguing modification of Davidson. Perhaps radically untranslatable conceptual schemes are indeed impossible — but partially divergent conceptual schemes are ubiquitous. Translation succeeds not because all languages organise reality identically but because their conceptual territories overlap sufficiently for bridges to be constructed between them.

Large language models are bridge-building machines.

But does the model understand?

At this point the familiar philosophical objection arises. Perhaps all this anthropomorphic language is mistaken. An LLM manipulates tokens according to statistical regularities. It does not understand justice, rabbits or Ukrainian national identity. Therefore its multilingual abilities tell us nothing about thought.

This objection has force but it does not dispose of the problem.

Even if one adopts a thoroughly deflationary account of artificial intelligence, the model must contain internal structures capable of preserving semantic relationships across languages sufficiently well to perform translation, answer questions and execute reasoning tasks. Calling these structures “statistical” does not make them philosophically uninteresting. Human brains are physical systems too. Merely identifying the substrate upon which representation occurs does not settle whether representation exists.

The deeper question is functional. If the same system can recognise that differently expressed propositions in Arabic, Ukrainian and Japanese concern substantially the same state of affairs, then there exists within that system some mechanism capable of establishing equivalence across linguistic forms.

Whatever we decide to call that mechanism, philosophers of language ought to be interested in it.

Indeed LLMs complicate another traditional assumption: that meaningful linguistic assertion necessarily reflects the mental state of a speaker. Contemporary philosophical work has begun examining how machine-generated language disrupts established connections between utterance, intention, authorship and epistemic responsibility. A sentence can now be grammatically impeccable, contextually appropriate and semantically informative without having originated in anything straightforwardly resembling a human intention to assert it.

The philosophy of language consequently faces not merely a new theory but a new category of speaker.

Artificial linguistic relativity

The most interesting experiments have yet to be performed.

Researchers could ask identical moral dilemmas in fifty languages and measure whether the model’s conclusions change. They could examine whether political concepts become subtly different when discussed in Russian rather than Ukrainian, Arabic rather than French or Mandarin rather than English. They could test whether causal explanations, descriptions of responsibility or assessments of uncertainty systematically vary with grammatical structure.

Then they could inspect the internal representations associated with those answers.

Human experiments on linguistic relativity confront an unavoidable problem: an English speaker and a Mandarin speaker are not merely speakers of different languages. They have different childhoods, educational systems, cultures and personal histories. Language cannot easily be isolated from everything else.

An LLM offers something closer to a controlled experiment. The same model weights can receive equivalent prompts in different languages.

That does not eliminate confounding factors — training corpora themselves encode enormous cultural differences — but those confounding factors are themselves philosophically revealing. Language and culture arrive bundled together in the texts from which the machine learns.

The result might be called artificial linguistic relativity: systematic variation in machine reasoning produced by changing the linguistic medium through which a problem is presented.

Evidence that such effects exist would not prove Sapir and Whorf correct about human cognition. But it would establish something perhaps equally interesting — that language can shape the behaviour of an information-processing system even when its underlying computational architecture remains unchanged.

Beyond Babel

There is finally a more ambitious possibility.

The biblical story of Babel portrays linguistic diversity as a catastrophe. Humanity loses its common language and therefore its ability to cooperate. Much subsequent western thought has retained something of this intuition. Translation is treated as an imperfect repair for the fragmentation of human speech.

Multilingual artificial intelligence suggests another interpretation.

The diversity of languages may represent not merely an obstacle to communication but a repository of alternative conceptual perspectives. Each language embodies accumulated historical decisions about which distinctions deserve names, which relationships deserve grammatical expression and which metaphors have become natural.

A sufficiently multilingual intelligence can move amongst them.

It may therefore achieve something difficult for any individual human being: simultaneous familiarity with dozens of linguistic perspectives upon the world. Rather than constructing an artificial Esperanto into which everything is flattened, the most sophisticated models may develop a higher-dimensional semantic architecture in which partially different linguistic conceptualisations coexist.

This would amount to neither the victory nor the defeat of Sapir and Whorf.

Language would indeed shape cognition — but multilingual cognition could in turn escape the dominance of any particular language.

The strongest version of linguistic determinism imagined human beings enclosed within linguistic worlds from which they could never entirely emerge. Large language models suggest almost the opposite possibility. Perhaps the route beyond linguistic relativity lies not in discovering a perfectly neutral language — something that may never have existed — but in learning enough languages that the peculiar assumptions of each become visible against the others.

The philosophical importance of multilingual artificial intelligence may therefore extend far beyond better translation software. These systems constitute unprecedented experiments in the relationship between words and concepts. They invite us to reconsider whether meanings belong to languages, whether concepts exist independently of them and whether understanding consists partly in the capacity to move between alternative linguistic descriptions of the same world.

Sapir and Whorf asked whether the language we speak determines the world we see.

Artificial intelligence introduces a stranger question.

What sort of world is seen by an intelligence that speaks them all?