I disagree with your analogy, as I find it overstates the active involvement of the driver (prompter) during the drive (actual image generation).

Preparation is it’s own process, whether you’re curating art you made yourself/stole from non-consensual artists, or have been finding references as an artist. Different skillset. They help the process of making the final image, but they are not a direct part of that process.

And let’s not kid ourselves about theses datasets. There’s no accountability so there’s no way to ensure that any dataset you’re getting from other people aren’t comprised of, at least partially, stolen art.

ControlNet let’s you add visuals to your prompt for greater control, but you’re still generating the image externally, and leaving the vast majority of the decision making to the model you’re using. Even if someone is happy with the result they get from a generative model and find it visually pleasant, that doesn’t make it art. The model is doing the work and the model cannot have artistic intent, so it cannot make art. It can make images and people can enjoy those, but those images aren’t something new.

They are amalgamations of most basic common denominator of existing things. It is much more like a really advanced collage that is great at hiding the seams.