Your conflating a lot of issues here that just muddy the waters.
First, whether particular features like watermarks end up in the end product or not is not terribly relevant. The particular model implementation that produces that behavior doesn't understand that watermarks are not a desirable part of the transformed end products, whereas humans do. In that way, the AI is certainly a little more 'honest' about what it's doing. It would be trivial to make the AI understand this, and stop reproducing watermarks.
You claim that the model performing this work is somehow different because AI doesn't 'think' about problems, as evidenced by strange artifacts, whereas presumably humans do think about them because of the absence of these artifacts. I'm curious why this makes a difference. If a model was sufficiently advanced that you were unable to differentiate a painting made by a human along certain themes and an AI along certain themes would it be somehow less objectionable? Why or why not? If the end product is of identical quality why should it matter the route you take to get there, in the eyes of the law?
You bring up scale, but scale is also not relevant. Say that I create a school devoted to training legions of artists to produce paintings in the style of a particular artist, while maintaining a transformative aspect. Is this illegal because I'm doing it at scale? No. If that's not illegal, why is it illegal to do it with code? Because it's more efficient? In what other domain is producing creative work more efficiently by transforming existing work illegal?