The GPL is based on rights from copyRIGHT law. It’s the author’s copyright that allows them to determine how to license it, and the GPL is one of many licenses.

So if something is not able to be copyrighted, then it’s not possible to put a copyright license like the GPL on it. The work is in the public domain, no license at all, different rules about derivative works.

If machines generated works (like code) cannot be copyrighted, then they can’t be licensed. Is the output of an LLM trained on GPL a derivative work, but can’t be copyrighted?

It’s a crazy decision, it’s going to do a lot of weird stuff, at least in the US. It might invalidate a bunch of the IP treaties that the US negotiated with much of the world since 2000.