Training is transformative use.

All of this is just statistical modeling. There is no “opt-out” for published art. Copyright only exists as a financial incentive to give the public useful works. Well… this is a use. And the end result is novel, reflecting minimal portions of each work. The more things it’s trained on, the less each one matters. A verbatim copy of Honky Chateau does not serve the same purpose as a robot that can play a rude jingle about your boss, by name.

The worst thing we could do about AI is make copyright even worse.