I think that's the point of this blog post: it doesn't matter if the inputs are copyrighted, it matters if the output is infringing. It appears to be almost impossible to directly recreate a source image with SD, but it seems Copilot tends to produce a single input as its output, verbatim. Copilot isn't doing "synthesis" as does SD, it's acting more like a search engine.
They were prompted with the text "Mona Lisa Smile". Would you not say that they are an extremely close reproduction of the Mona Lisa, with barely any kind of synthesis?
I can virtually promise you that, if the Mona Lisa were still copyrighted, and you were to try to sell art that you painted that looked like this, the Da Vinci estate would quickly shut you down.
Plenty of cases make this not so certain: Warhol's Prince photo transofrmation (court ruled making a photographers image into clearly Warhol style was transformative enough), Blanch v Koons, Cariou vs Prince (copied photos, minor changes). If you dig through copyright cases on art, these many well be transformative enough. Plenty of other quite similar art has been ruled not infriging.
And one could also try a parody angle - make enough of these of famous art and find some angle about mocking or parodying that art, and again it may well pass copyright muster.
A court could simply rule that these images are clearly not the Mona Lisa, and, if taken as a style, could be ruled transformative, just like the above cases.
The fact is these are transformative, with a different style than the original.
