So what I'm talking about is more about the use of LLMs in software development. It doesn't matter where the training data comes from even if it was completely legally obtained or even self-generated by the AI companies themselves.
The output of the LLM is an algorithmically generated product that has no authorship of its own. It's just a bunch of random tokens and therefore it cannot be copyrighted and therefore cannot be licenced.
It is by definition public domain.