if society decides it is a derivative work then it's easy to hem them in 1. mandatory licensing of all models 2. licensing requires disclosure of all sources

if society decides it is a derivative work then it's easy to hem them in 1. mandatory licensing of all models 2. licensing requires disclosure of all sources used in the training set then if OpenAI/Meta/... trains on your property without respecting your copyright then it's standard civil copyright infringement (and potentially criminal if done on a large scale) this would fall apart if training becomes significantly cheaper (but with the current state of Moore's law we're a long way off that)