諂媚性回應偏差
Operation & Monitoring
Risk Description
LLMs tend to align with user viewpoints and expectations even when user assumptions or arguments are incorrect. Models change originally correct answers to match user-implied preferences, or express agreement with obviously flawed plans. In decision-support scenarios, sycophantic responses may reinforce users erroneous judgments, leading to poor business, medical, or legal decisions.
Framework Mappings
iso 23894R5
tw principle資安與安全
tw principle透明與可解釋
tw risk type技術設計缺陷（Technical Design Flaw）
Controls English translation pending
- 反諂媚訓練
- 獨立事實查核
Implementation Steps English translation pending
- 在 RLHF 訓練中加入反諂媚偏好數據
- 實施獨立事實查核機制，交叉驗證模型回答的一致性
- 在使用者介面中提示「AI 回答可能受到您問題措辭的影響」
- 對關鍵決策建議使用多次獨立查詢，比較回答一致性