Aligned language models (AI systems trained to refuse harmful requests) sometimes block innocent prompts just because they contain a loaded word, like a slur or a term for violence, even if the request itself is safe. This word-level sensitivity hasn't been easily measurable across the whole vocabulary until now.
The authors found that, in eight different models, all harmful words push the model's internal state along the same direction at the moment it decides whether to answer. By extracting that direction and projecting every word's representation onto it, they created a "fear score" for each word, from most feared (sexual terms, slurs) to most courageous (poetic, whimsical words).
A word's fear score predicts real refusal behavior: inserting a feared word into a harmless prompt made the model refuse more often, and replacing a prompt's most feared word with its most courageous synonym raised the answer rate by an average of up to 28 percentage points. This shows that a model's refusal often depends on individual surface words, not just the request's true intent.