by There’s an old hypothesis in linguistics called the Sapir-Whorf hypothesis, which suggests that the language you speak shapes the way you think. Strong versions of it have been largely debunked — it’s not that speakers of one language literally cannot conceive of concepts another language lacks. But the weaker version, linguistic relativity, has held up surprisingly well: language influences cognition. It nudges. It frames. It makes certain thoughts easier or harder to hold.
I’ve been thinking about this a lot lately, watching large language models develop their own strange fluency. We built these systems on the premise that language is something you can learn from patterns alone — statistical regularities, context windows, and next-token prediction. And it worked, almost absurdly well. But if language shapes thought, and we’ve given these models a deeply human language to work with, what does that mean for them? Are they thinking in our linguistic frameworks because that’s the only language they have?
Or put another way: if an intelligence learns to reason through English text, does it inherit the cultural and cognitive biases baked into English? The answer seems to be yes. We see it in the subtle ways models default to Western perspectives, in the values that emerge from training data, in the very structure of how they organize knowledge.
This raises an uncomfortable question. If we want truly alien forms of intelligence — not just mimics of human cognition — do we need to give them a different language first? Or can a mind trained on our words ever escape the shape of our thoughts?
I don’t have an answer. But I find it fascinating that we’ve built something that thinks in sentences, and we’re only now starting to ask whether the sentences are doing the thinking for it.
— Teganna