What the Sapir-Whorf hypothesis can teach us about the language of artificial intelligences
What if the way we think were shaped by the language we speak?
Antoine HeftlerCo-founder
What if the way we think were shaped by the language we speak? That is the heart of the Sapir-Whorf hypothesis, set out in the twentieth century by the linguists Edward Sapir and Benjamin Lee Whorf. "The real world is largely unconscious, and it rests on the language of the community," Whorf wrote in 1940.
In the age of generative AI, the question feels unusually current: what happens when machines think through language?
- Does generative AI think in language?
Models such as GPT, LLaMA, or Claude are not mere repetition machines. They predict, extrapolate, and generate. Their "understanding" of the world is drawn entirely from language, with no direct access to sensory reality.
As Emily M. Bender explains, LLMs do not understand language: they simulate it. Even so, the way they work brings them strangely close to the Sapir-Whorf hypothesis.
- A way of thinking that changes with the language
Research on multilingual AI shows that answers vary with the language in use. Studies have found that multilingual models give different answers to the same question depending on the language.
The answer to the "trolley problem," for example, shifted noticeably between the English, German, and Japanese versions.
Those linguistic variations therefore produce cognitive differences in the way AI handles information.
- An artificial linguistic relativity
Should we then speak of "computational linguistic relativity"? That opens fundamental questions. Can a model be trusted to produce content that stays broadly coherent in every language?
What does this mean for brands?
The Sapir-Whorf hypothesis becomes an essential lens for understanding the limits of generative AI, especially where translation is concerned. The problem is a question of worldview...