I think they are saying the training part itself doesn’t. It seems to be kind of a fine point, otherwise you could implement a license laundering AI by trainin

I think they are saying the training part itself doesn’t. It seems to be kind of a fine point, otherwise you could implement a license laundering AI by training it on a single codebase and asking it questions about it. If you end up with a byte by byte replica sans a LICENSE file presumably you wouldn’t get away with actually _using_ the model to clone licensed software even if training it was technically ok.