You definitely need permission to make copies of people's art on your computers to use whilst learning though (copyright holders tend to be OK with or completely unaware of that, but it's a right they reserve)

More to the point, this feels like the "have your cake and eat it" of legal arguments. If we're going to anthropomorphise AI to the extent we argue storing data in its data corpus without authorisation is just like allowing a human to look at your legally owned stuff, and reinterpreting that data in a [sometimes not particularly] lossy fashion as a set of weights and biases is just like a human learning, we should probably follow the "AI is just like a person" argument through to its logical conclusion and hold it legally responsible for not respecting copyright law if it ever produces outputs with trademarked images or uncanny resemblance to copyrighted material that aren't deemed "fair use", just like a person would be.

(tbf, it's not too hard to imagine a world in which ChatGPT says "I'm sorry, but I am not able to write stories involving Mickey Mouse, because Mickey Mouse is the intellectual property of the Disney Corporation. Respecting intellectual property rights is very important", and I'm not sure that's an improvement...)