The problem is, because of how these models were trained, how do we know if its telling us how it feels vs how it thinks we want it to respond?