It's only a copyright issue if I know what the original source code looks like. If I don't know what it looks like, and my autocomplete writes it, how could I possibly know it's stolen?
I think the line b/w derivative work and new work might be different.
I mean if llms are trained on it ... and a lot of other things and then LLM can output the source code from a input ... then wouldn't it be open source / public domain