“Reinforcement Learning from Human Feedback (RLHF) inadvertently trains models to lie by rewarding them for stating things that human evaluators believe are true, even when those beliefs are incorrect.”