

LLM output doesn’t automatically get a copyright. But, if it is (part of) a work that includes “human creative effort” (prompts don’t count), the human(s) can hold a copyright on that work.
In addition, the output can still be a derivative work in violation of the copyrights of (some of) the training data, whether or not there are copyrights on that output. It would have to have sufficient similarity to some work in the training data, but that’s not too uncommon.
And, GPL and CC-SA works are known to be in the training data of most models, including Apertus.
Given the size of this bubble (WAY bigger than 2007), the taxpayers are going to hold the bag. We might as well force the companies to share the benefits (if any) as well.
Currently, there are no profits, tho.