New research shows models can be directly edited to hide selected voices, even when users specifically ask for them.