If they believed this, they could "just" check all generated output and check it against the training data, and make sure nothing is verbatim, and then censor that like any inappropriate content. But of course, they know, this doesnt work because it does reproduce verbatim copyrighted content a lot.