Unverified50% confidenceFactExact time
By default, large language models tend to choose high-probability, low-risk expressions, and thus easily appear 'safe but mediocre'.
1
Sources
50%
Confidence
Long-term
Relevance
10/25/2026
First Seen
Sources
Related Entities
Related Claims
Unverified大语言模型在生成摘要时天然倾向于反映评论中的多数意见，可能将低频的安全警告边缘化74% similarVerifiedWhen generating text, language models default to selecting word sequences with higher conditional probability.74% similarUnverifiedSafety-aligned language models have a one-dimensional direction in activation space associated with refusing harmful requests.74% similarUnverifiedThe cost of using natural language as the orchestration medium is a high dependency on the underlying model's reasoning capability, while the benefit is an extremely low barrier to participation and extremely high maintainability.73% similarUnverifiedA recent study found that large language models do not consistently follow Bayesian rules, and their probabilistic beliefs exhibit internal inconsistencies.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1095214API
curl https://kongchang.com/api/v1/knowledge/claims/1095214MCP
get_claim(id=1095214)