It's not meant to be a useful use-case, but a proof that the training data contains GPL code and the model is capable of reproducing copyrighted code. And yes,

It's not meant to be a useful use-case, but a proof that the training data contains GPL code and the model is capable of reproducing copyrighted code. And yes, the implication is that a different less explicit prompt could still emit copyrighted code.