That’s a legal point of contention because the nature of language/image models is hard to fit into the existing copyright framework. That only really applies to cleanroom-ish one shot requests where the inference input doesn’t contain the copyrighted material in question.

It’s a lot easier to argue that it’s a derivative work when you feed the copyrighted code directly into the context and ask it to port it to another language. If the copyrighted code is literally an input to the inference request, that would not escape any judge’s notice. The law may not have any precedent for this technology but judges aren’t automatons beholden to trivially buggy code that can’t adapt.