The AI is learning. Why is one of those things stealing and not the other?
The definition of a derivative work is agnostic to the means to create it. The final product is what is being evaluated.
If the answer is "but the Bob Ross paintings that are derivative works by amateurs are never sold and only hung at home - it would only be an issue if they sold them or tried to display them publicly as described https://www.law.cornell.edu/uscode/text/17/106 " - then I would agree.
If I create an image through any means - be it paint on canvas, stylus on tablet, or prompt and ML and don't violate 17 U.S. Code § 106 then there shouldn't be any issue.
If, however, I publish that image - again, no matter how it is created - then there is an issue. But it isn't the maker of the paints, or Adobe, or the creators of whatever ML model generated it. A program cannot hold a copyright and cannot itself create a derivative work. ( https://www.smithsonianmag.com/smart-news/us-copyright-offic... ) and so the issue of creating and publishing a derivative work is upon the human who did it - not the computer.
If you want to sue someone for publishing an image generated as "create a landscape of San Francisco in the style of Pixar" then have Pixar go after the person who published that image. It doesn't matter what process created that image.
Even after following along and gaining the confidence to paint on their own, they continue to use the same approach to painting. Are those paintings afterwards derivative works and to be treated the same way as an AI painting asked to create a scene with certain elements in the stye of Bob Ross?