I hear what you're saying and I understand it doesn't literally collage images together. I've tested it A LOT, I see what it does. I honestly think the technical details are less important than the fact that it can't do anything without that data. Less data, and it would be worse. More data and it will be better. While no image that it makes is exactly like an image from the training data, that does not equal original imagery.
Whether it has been trained to steal artists work, or it directly collages from billions of images, everything it creates is based on the work and labor of other people. Actual artists, even if they're referencing other stuff, THEY still have to create it. Their mark comes from them. So yea, I respectfully disagree.