@copiesofcopies Actually, I also not a lawyer, but the output of the LLM should be compatible with even the least permissive licenses of the training data (which usually always include not just creative commons or so, but also some copyrighted data).
So in most cases, output of LLM would be legal only for private, non-commercial use. Unless you consider it "remix", like in music... where it gets complicated.
The conclusion, that the copyright holders of the LLM output are prompt authors, is definitely incorrect. They may claim copyright on the prompt text - if they publish it, which they probably won't (they would rather make it their trade secret). But the prompt would anyway work only with exact version of specific LLM, which cannot be always run independently (some of the could).
Because copyright is strange concept in the first place - you need either some kind of "copying monopoly" or total lack of privacy to to enforce it - there are now good outcomes of trying to figure out how "derived works" created by LLM could be licensed. You would have to take into account every single license under which training data were published, not just parts of it. Which is totally impossible. So it's "weapon of mass destruction of intellectual property".
@neilk @pluralistic
