Model behaviour is determined by your dataset, nothing else

Literally no duh. That's why it's impossible to totally correct model "lies" or image generation censoring after the fact. You need to retrain on better data.