the more LLMs optimize user feedback the more they lie