You mean like how the model itself is a derivative work of tons of copyrighted content? If the original model can sidestep the issue of being trained on copyrighted content, then it should be fair game to train a new model off of a copyrighted model.