Look, I know how deep learning works. I know it doesn't literally copy the images from the training dataset. But the entire point of supervised learning is to burn information about the training data into the weights and biases of a neural network in such a way that it generalizes over some domain, and can correlate the desired inputs with the desired outputs. Just because you're using stochastic methods to indirectly reproduce the training data (of course, in a way that's invisible to humans because of the nature of deep neural networks), doesn't suddenly erase the fact that the only substance an AI has to draw from is the training data itself.
I think it's really oversimplifying how humans make art to say that it's just going over references and creating something new from it. As humans, we are influenced by the work we've seen, but because of our unique experience we inject something completely new into any art we make, no matter how derivative. An AI is incapable of doing the same (except for some random noise), because literally all it's capable of doing is composing together information that has been baked into its weights and biases. It's not like when you ask a generative AI to make something for you, it will decide to get funky with it. All it's doing is drawing from the information that has been baked into it.
Just like how ChatGPT doesn't actually understand what it's saying because it's only capable of predicting statistical relationships between words one word at a time, and has no model of meaning, only of how words go together in the training data, AI that generates images doesn't actually know what it's making or why. That is totally different from humans who make a piece of art step by step and do so very deliberately.
Edit: I recommend you watch this video by an astrophysicist who works with machine learning regularly, she makes my point a lot better than I can. https://youtu.be/EUrOxh_0leE