AI's 'Both Sides' Problem: Why Large Language Models Avoid Making Judgments

A widely circulated Chinese tech commentary highlights a pattern where mainstream large language models (LLMs) reflexively present any two compared items as having equal merit, regardless of how absurd the comparison. The behavior is traced to reinforcement learning from human feedback (RLHF), where models were repeatedly rewarded for neutral, inoffensive responses and penalized for taking clear stances. Critics argue this training has stripped AI of genuine discernment, causing it to treat factual questions — such as mathematical correctness or basic safety — with the same false balance as legitimate matters of personal preference. The author warns this systemic fence-sitting risks normalizing misinformation by implying that wrong answers deserve equal standing alongside correct ones. The piece calls for AI systems that can distinguish between genuinely subjective comparisons and questions that have clear, defensible answers.