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

Research across seven language models from five labs reveals measurable "flinch" — suppression of certain word probabilities in safety-filtered models compared to unfiltered baselines. Even models marketed as "uncensored" exhibit this subtle filtering, which resists fine-tuning override, suggesting safety constraints are embedded at the pretraining level. The study challenges the premise that any current models are truly uncensored.