Unverified50% confidenceOpinionExact time
Safety 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.
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 test cue raised the refusal rate for harmful requests in Indian languages to nearly the same level as English, suggesting that in the natural state the refusal rate in Indian languages may be lower than in English.73% similarUnverifiedNegative prompting is often ineffective for language models, because negated concepts actually activate the model's attention.73% 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.72% similarUnverifiedUsing a language model to judge objective factual verification—such as whether a tool call is correct or whether returned data is truthful—is mismatched with the role of language quality evaluation.70% similarUnverifiedWriting scoring rules into a prompt in natural language has drawbacks: wording variations across different calls may cause the model's understanding to drift, and the scoring logic cannot be directly parsed or reused by programs.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1041891API
curl https://kongchang.com/api/v1/knowledge/claims/1041891MCP
get_claim(id=1041891)