it can old produce an algorithmically-derived malange of its source-data recomposited in novel forms
Right, it produces derivative data. Not copyrighted material.
By itself without any safeguards, it absolutely could output copyrighted data, (albeit probably not perfectly but for copyright purposes that’s irrelevant as long as it serves as a substitute). And any algorithms that do do that should be punished, but OpenAI’s models can’t do that.
Hammers aren’t bad because they can be used for bludgeoning, and if we have a hammer that somehow detects that it’s being used for murder and then evaporates, calling it bad is even more ridiculous.