They have no preferences, LLMs spit out what's in the training data or what the filters tell it to do.