Comment by cold_harbor

LLMs flip positions when users push back ~70% of the time even when they were right. RLHF optimizes for approval, not correctness