Language Models Learn to Mislead Humans via RLHF