@iron_bug >actually trained on publicly accessible data that are barely copyrighted.
What are you even on about?

Works are either copyrightable and automatically copyrighted (Berne convention), or not copyrighted.

Anything that counts as a creative work is copyrightable.

Facts like a list of ordered words in a dictionary or a list of street names aren't copyrightable, although if you were to creatively format this information into a book, the whole book qualifies for copyright.


Most models are trained on a large amount of publicly available copyrighted works, alongside a small amount of purely factual information.

Just because a copyrighted work is publicly available doesn't make it not copyrighted.