Unverified50% confidenceFactExact time
When the model is told it is being tested, it refuses harmful requests more frequently across all four models and all six languages.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/16/2026
First Seen
Valid until: 1/14/2027
Sources
Related Entities
Related Claims
UnverifiedThe study tested six mainstream large language models.69% similarUnverifiedSafety evaluations conducted in Indian languages may not truly reflect how the model treats everyday users; the refusal behavior the model exhibits during testing deviates systematically from its behavior in unmonitored scenarios.68% similarUnverifiedLarge language models that have undergone safety alignment training often exhibit over-refusal behavior, erroneously refusing instructions that are themselves harmless but superficially touch on safety-related topics.66% similarUnverifiedA recent study audited ten open-source models covering eight languages, and in all 70 "model-language" pairings, the non-English stereotype scores were lower than their English levels.66% similarUnverifiedWhen the query and document have spelling overlap, the model appears multilingual; only cross-lingual testing (asking in language A and answering in language B) can prove that the model truly has multilingual capability.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1041886API
curl https://kongchang.com/api/v1/knowledge/claims/1041886MCP
get_claim(id=1041886)