Depends on the medium, but human drawing is not that far from finding an image in a noise. If you start with empty paper, then refine it into sketch, then into colored image, then into more detailed image, etc. It's not that far off from refining image out of noise.

With img2img it is like drawing what it "sees", with text2img it's more like drawing something from memory.

Anyway, as I said most of the training data were photographs of the real world, those would still exist without human artists. I guess the pending question is if the AI could then develop some artistic style without inspiration from people.

My guess is that not at first, without non-realistic inputs it would only try to create realistic images. But there would be some distorted images and if any of them look good, they'll end up on the internet. Future AIs will be train on them as well and expand on them... It would eventually evolve into different artistic styles without humans, just take longer to get there. Good we had human artists to kickstart the process.