What happened
Researchers found that multilingual AI models store informal language (slang, casual speech) using a shared abstract representation that works across different languages, rather than treating each language's slang as separate isolated knowledge. This means you can steer an AI model's formality level in one language and the effect transfers to languages it has never seen before — the model has internalized a portable, language-agnostic concept of "casual speech."
Why it matters
This is the first direct evidence that large language models build unified, causal representations of pragmatic meaning (how language tone and register work) rather than just memorizing surface patterns. If AI systems genuinely internalize abstract linguistic concepts across languages, that changes how we think about what these models actually understand versus what they're pattern-matching.