I don't have access to Bard. I'm skeptical that's a valid approach. Language models are famously incapable of introspection. Does it change the answer if you tell it it's wrong, for example?