Imo, I'd prefer a "contamination" approach where the strictest licenses in the training data applies to all outputs. I doubt such a rule would get through big business filters but it would maximize the public good and any country that does manage something like that would probably gain the most benefits from these companies.