But even if by "theft" you mean "copyright infringement" there is a strong argument for fair use, we humans consume large amounts of copyrighted content and are influenced by it in everything we do, I don't find it unethical or wrong in general, and I don't see how it should be any different for a machine, unless you are against job automation, but in this case you seem to be worried about "theft".
Edit: HN thinks that I'm posting too fast so my reply to the comment below here ahah: "Well, then we agree that it wouldn't be a problem if the model didn't output an image from the training dataset, which is extremely rare or essentially impossible with today's dedup steps used during dataset creation."