Unverified50% confidenceOpinionExact time
Safety classifiers often see only a single message and lack conversation history, making it easy to misjudge sarcasm, quotations, or educational discussion in context as real harm.
1
Sources
50%
Confidence
Long-term
Relevance
11/4/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe accuracy of existing content classifiers in distinguishing 'genuine threats' from 'rhetorical expressions' is far from satisfactory.73% similarUnverifiedIndirect prompt injection is more harmful because the victim is completely unaware of the malicious content.72% similarUnverifiedAbout a quarter of conversations judged safe in isolated evaluation (context-blind) were actually unsafe for the specific questioner.72% similarUnverified当网络批评可能招致执法人员实地拜访时会产生寒蝉效应，导致公众自我审查71% similarUnverified信息论中的信噪比概念可类比内容传播——信道中内容过多时有价值的信号难以被识别69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1133583API
curl https://kongchang.com/api/v1/knowledge/claims/1133583MCP
get_claim(id=1133583)