You're saying that the generative model isn't a derivative work of all this copyrighted material?
Definitely not - these are high dimensional AI models that learn billions of weights from billions of different copyrighted training inputs.
If anything, it's a *gradient* work!