Isn't the narrative that AI does not infringe on copyright literally, demonstrably false? One of the 'goodness' measures of AI training is the loss function, meaning its ability to reproduce the training set. If it has low loss, and some copyrighted content is included in its training set, like the Harry Potter novels, it is by definition able to reproduce the novels contents more or less verbatim, perhaps with some minor inconsistencies. This is like storing a compressed version of a movie on my server, and claiming that it doesn't infringe copyright because the copy is lossy.
The fair use argument is retarded.