AIs benefit from negative examples too. As long as the inputs are accurately labelled then it enhances what it knows.

For example, if I go to ChatGPT I can ask it to produce a wild parody of a far-right rant about something and it'll be able to do so. I could feed it an existing far-right rant and ask it to analyze what it's actually talking about and it'll have a better ability to figure out what's going on if it knows that style from its training. I could tell it "write an article about this subject that specifically avoids the tropes and pitfalls of far-right media."

We've seen frequently the pitfalls of trying to censor a model's training data to exclude something that's part of the common world around it, the model either "fills in the gaps" itself or just ends up insane. When the nude human form was excluded from training data for early image AIs it wound up making weird Cronenbergs because it had no idea what was really going on underneath the clothing, for example.