They should train a model on a clean dataset and copyright dataset, charge extra on the copyright model, and pay a royalty to copyright owners when their works are cited in a response.
The problem there is how are we defining "works are cited"? Also couldn't you just do the same thing done to spotify and make bot farms to generate millions of citations?
You can simply pay to everyone whose works you have used for training, every time a model processes a request.
I like this and agree. It should be opt-in. I almost feel as if it should be something exciting and rewarding.