为什么要减少AI 的谄媚(sycophancy)?谄媚是 alignment failure(对齐失败).LLM 倾向于 迎合用户观点 而不是提供真实或批判性分析,会降低 AI 的 epistemic reliability(认知可靠性).
谄媚会放大用户错误信念,形成 confirmation bias amplification(确认偏误放大)
Reducing sycophancy improves epistemic quality and leads to more reliable AI-assisted decision making (Dubois et al., 2026)
AI sycophancy has been identified as an alignment failure that can undermine decision quality (Dubois et al., 2026).
⸻
APA 7 Full Citation
Dubois, M., Ududec, C., Summerfield, C., & Luettgau, L. (2026). Ask don’t tell: Reducing sycophancy in large language models. arXiv.
doi.org