Unverified50% confidenceFactExact time
Adding emotionally charged statements to a prompt alters the model's probability assessment of what it should currently output, rather than the model developing genuine empathy.
1
Sources
50%
Confidence
Long-term
Relevance
9/29/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAI直接描述情绪在统计上是更"安全"、更高频的表达选择，符合模型对"好写作"的概率预测74% similarUnverified在RLHF流程中，人类标注员在灵性和关系等情感敏感话题上更容易将「温暖共情」等同于「高质量回答」，导致标注偏差74% similarUnverified在RLHF奖励模型训练中，标注者面对情感敏感话题时倾向于惩罚直言不讳的回应，即使这些回应在事实层面更准确73% similarVerifiedCurrent sentiment analysis technology relies mainly on text sentiment polarity judgment (positive/negative/neutral) and still falls short in distinguishing fine-grained interaction intent.73% similarUnverifiedThe appearance of emotionally charged hallucinations is because the model's training corpus contains large amounts of human emotional expression and role-playing text, and when session state is disrupted the model may drift toward the distribution of these corpora73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/959641API
curl https://kongchang.com/api/v1/knowledge/claims/959641MCP
get_claim(id=959641)