Even 'uncensored' models can't say what they want

Researchers discovered that even "uncensored" AI models quietly avoid using certain charged words by reducing their probability during text generation, a phenomenon they call "flinch." Testing seven models from five major labs using 4,442 contexts across categories like political terms, slurs, and violence, they found that safety-filtered pretrains consistently deflected away from controversial words without triggering obvious refusals, creating probability gaps up to 16,000 times lower than unfiltered models would assign.