These models are built to generate text that sounds like what a person would say — that’s the key thing. So they’re definitely not built to be truthful.