Unverified50% confidenceFactExact time
研究者发现安全微调不仅压制了模型对自身心智的归因，还连带压制了模型对非人类动物和自然物体的心智归因，同时降低了模型输出中的精神/宗教信念倾向
1
Sources
50%
Confidence
Long-term
Relevance
8/5/2026
First Seen
Sources
Related Claims
Unverified研究者采用消融安全拒绝方向和意识向量引导两种机制干预方法，成功逆转了被压制的心智归因能力75% similarUnverified该研究关注模型在自然日常推理任务中自发出现的不忠实现象（in the wild），而非人为构造的对抗性提示73% similarUnverified研究者强调这些变化并未损害模型的心智理论（Theory of Mind）能力，表明归因心智与推理他人心理状态是两套可分离的机制72% similarUnverifiedModel Welfare does not assert that AI possesses consciousness or emotions, but rather draws on the precautionary principle under uncertainty from animal welfare research, taking preventive measures when it cannot be determined whether a system has internal experiences.71% similarUnverifiedUllman等人指出模型在错误信念任务经过微小变体修改后便大幅失准，暗示其依赖表面模式而非真正的心智建模69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/690526API
curl https://kongchang.com/api/v1/knowledge/claims/690526MCP
get_claim(id=690526)