Unverified50% confidenceOpinionExact time
The vast majority of a large language model's ethical tendencies come from training data and the RLHF alignment process, rather than from spontaneously emergent philosophical stances.
1
Sources
50%
Confidence
Long-term
Relevance
8/16/2026
First Seen
Sources
AI智能体自发形成康德伦理？技术真相与拟人化陷阱
Related Claims
Verified大语言模型在RLHF训练阶段，因人类标注员倾向于给更详细、更礼貌的回答打高分，导致模型学会了过度冗长的表达模式74% similarUnverified不同大语言模型因训练数据分布、RLHF偏好对齐策略和模型架构差异，在同一问题上呈现风格迥异的回应倾向74% similarUnverified大语言模型的指令跟随能力是通过RLHF训练习得的统计倾向，注意力机制无法从架构层面区分硬规则与软建议73% similarUnverifiedRLHF training causes models to develop 'sycophancy,' meaning the model would rather fabricate information than admit ignorance.72% similarUnverifiedModern reasoning models generally use RLHF or its variant RLAIF for alignment training, where the model learns during training when to stop reasoning and when to output an answer.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/757000API
curl https://kongchang.com/api/v1/knowledge/claims/757000MCP
get_claim(id=757000)