I suppose that's a relatively easy thing to fix, technically. It proves, however, that th underlying LLM is trained on copyrighted data.

I'm not sure the problem goes away simply if the LLM in question (or any other one) gets some "no verbose regurgitation" filter.