You can "launder" away the license of any source code you have copied simply by deleting it! No snazzy neural network needed.. The litigants argument is that this is what GitHub CoPilot does. It allows others to publish derivative works of copyrighted works with the license deleted. Given that it apparently is trivial to get CoPilot to spit out nearly verbatim copies of the code that it was trained on, I don't think it satisfies the "transformative" requisite of the (American) Fair use doctrine.
It seems like the ideal way to proceed is to make the AI output unique and creative. Perhaps that requires AGI because currently the model has no understanding of art.
Maybe more importantly, the AI needs the ability to judge when its output amounts to plagiarism, like humans generally are able to. The AI needs to feel bad about ripping off someone else’s work. ;)
