I tend to agree - but you assume that it would not be possible to create a model that can train on copyrighted work and only output text which would be considered fair use.

That seems very possible to me, and undermines the "training is copyright violation" argument. It's not the training, it's the output.