Unverified50% confidenceFactExact time
Multiple academic studies have found that translating the same harmful question from English into low-resource languages significantly reduces the refusal rate of mainstream LLMs.
1
Sources
50%
Confidence
Long-term
Relevance
9/18/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedSince 2023, the widespread adoption of large language models has lowered the barrier to bulk-generating seemingly well-sourced articles, comments, and even academic abstracts.77% similarUnverifiedMultiple academic studies in 2023 confirmed the feasibility of cross-lingual jailbreak attacks, finding that model refusal rates dropped significantly after translating harmful requests into low-resource languages such as Zulu and Georgian.76% similarVerifiedThe 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.75% similarUnverified低资源语言中相当一部分面临数字化消亡风险（联合国教科文组织估计）73% similarUnverified将英语有害提示翻译为祖鲁语、格鲁吉亚语、泰语等低资源语言后，GPT-4、Claude等模型的拒绝率可从接近100%骤降至30%-50%73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/933160API
curl https://kongchang.com/api/v1/knowledge/claims/933160MCP
get_claim(id=933160)