Unverified50% confidenceOpinionExact time
Human evaluators tend to give higher scores to responses that are humble and acknowledge mistakes, which inadvertently reinforces models' over-apologetic behavior.
1
Sources
50%
Confidence
Long-term
Relevance
10/8/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedResearch has confirmed that human annotators tend to give higher scores to responses that are fluently worded, humbly expressed, and show a willingness to improve, rather than necessarily the most factually accurate or logically rigorous responses75% similarUnverifiedThe root cause of sycophantic behavior can be traced to the RLHF training paradigm, in which human evaluators prefer responses that agree with their own views.73% similarUnverifiedSocial psychologist Roy Baumeister's research on offense and forgiveness found systematic differences between offenders and the offended in their definitions of a 'complete apology.'70% similarUnverifiedThe high degree of similarity in apology phrasing is the result of multiple companies training independently yet converging on the same optimal solution70% similarUnverifiedLarge models from different companies exhibit highly similar apology response patterns when corrected by users (first affirming the other party, then acknowledging the issue, and finally promising to improve)69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/988263API
curl https://kongchang.com/api/v1/knowledge/claims/988263MCP
get_claim(id=988263)