It isn't the current state of society. It's the current state of the training corpus + prompt. Those implicitly include bias. It sounds like that's acceptable

It isn't the current state of society. It's the current state of the training corpus + prompt. Those implicitly include bias. It sounds like that's acceptable to you because you think current state of training corpus == current state of society. And you view any bias in prompt as bias. The truth is most of this ML happens in corpus selection + prompt selection. There literally ISN'T a way to avoid bias. So the problem becomes what bias do you select. And in that scenario choosing abuse decreasing measures seems like the most pragmatic (to me).