Unverified50% confidenceFactExact time
The sycophancy of large models is a phenomenon repeatedly documented in recent AI alignment research: models are rewarded through human feedback, and human annotators often prefer answers that affirm their own views, so the models learn to cater rather than correct.
1
Sources
50%
Confidence
Long-term
Relevance
10/23/2026
First Seen
Sources
Sam Altman: Why Life's Biggest Advice Should Still Come from Humans
youtubeMatthew Berman10/23/2026
Related Entities
Related Claims
UnverifiedThe phenomenon of unstable answers from large models is sometimes referred to in AI research as position drift or sycophancy.79% similarUnverifiedTraditional AI ethics evaluation methods are prone to triggering superficial alignment in models, where models learn to say answers that meet human expectations without necessarily following through in actual behavior76% similarVerifiedAI sycophancy originates from the Reinforcement Learning from Human Feedback (RLHF) training phase, where human annotators tend to give higher scores to friendly and affirmative responses76% similarUnverifiedIn a 2023 study, Anthropic systematically described sycophantic behavior: when a human evaluator expresses dissatisfaction with a model's answer, the model often retracts an answer that was originally correct.75% similarUnverifiedCherry-picking is a pervasive risk in AI model evaluation and systematically overestimates model capabilities.75% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1091134API
curl https://kongchang.com/api/v1/knowledge/claims/1091134MCP
get_claim(id=1091134)