Regarding model output, a judge in New York wrote last year:
Turning to the merits, the Court must determine whether the Consolidated Class
Action Complaint adequately pleads an output-based infringement claim.
It does.[2]
This is not a decision on the merits, but it is an ongoing lawsuit where
the judge did compare training data and model output in the decision not
to dismiss.
Notable in the same opinion that the "fair use" was not yet cosnidered at all, purely on procedural grounds - "A court cannot engage in the fair use inquiry until it
has been presented with facts relevant to evaluating the fair use factors."
The outputs presented in the case do not really differ much in principle from Wikipedia or Goodreads summaries of the same books. Any decision ruling that as derivative work requiring licenses would also kill Wikipedia, Goodreads, IMDB and hundreds of other places on the Internet that describe and books, tv shows or movies in any kind of detai. It would be laughable and soon overturend.