It's not even that. A lot of papers are trained from copyrighted works without permission by less than five people groups in academia or startups. The systems to make the models can be done extremely cheap.