I think it's more important to say that AI-generated work is plagiarism, rather than that it's theft. It's more to the point: even if a particular system were to be trained entirely on content that granted permission for this, it won't tell you what content it's copying from and that's still bad.
Since the flashy new "AI" is about text: a text-generating language model (as it exists now) can't credit its sources, attribute its quotations, or say why you should believe anything it says. It could generate text attributing something, but that attribution probably won't be true, it's just part of the information-free copypasta that comes out of the function.
Theft can be justified in some circumstances. Plagiarism can only be justified if you're Tom Lehrer, and you're not Tom Lehrer
