The "it's stealing" arguments have quieted down. Especially since there are weights trained on fully licensed materials by Adobe and others.
Now that code models can do it, it's a moot point. Data can be found anywhere, and models are pretty good at generalizing out of domain, not unlike human brains.
Engineers finding out these models are good is drawing a lot of the same "slop" / "clanker" arguments that non-artists have been using. These arguments are much less interesting than arguments about copyright and control.
