The question you should ask is, what is the easiest way for the neural net to pretend that it has emotions in the output in a way that is consistent with a really huge training dataset? And if the answer turns out to be, "have them", then what?