Verified75% confidenceFactExact time
The safety alignment of many models degrades significantly in non-English languages, and low-resource languages are especially easy to exploit because the human annotation data for RLHF alignment training is highly concentrated in English.
3
Sources
75%
Confidence
Long-term
Relevance
8/6/2026
First Seen
Sources
Related Claims
UnverifiedSafety alignment training data is most comprehensive in English, and may have significant coverage gaps in languages such as Arabic and Urdu, constituting a multilingual jailbreak vulnerability.76% similarUnverified绝大多数RLHF标注工作以英语为主，导致不同语言间对齐质量存在结构性失衡76% similarUnverifiedExisting multilingual models such as mBERT and XLM-R perform far worse on low-resource languages than on English.73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/695656API
curl https://kongchang.com/api/v1/knowledge/claims/695656MCP
get_claim(id=695656)