Evidence · DiggingBeagle record
Evaluation awareness beyond English and Indic language refusal rates
Primary controlled multilingual safety-evaluation study distinguishing harmful and benign refusals.
- Source role
- primary disclosure
Evidence record
Primary controlled multilingual safety-evaluation study distinguishing harmful and benign refusals.
Claim-level citations (1)
- supportsModels changed refusal behavior when a safety test was declared in different languages: In a controlled study of four models and six languages, labeling a request as part of a safety evaluation altered refusal behavior, with different changes across languages and benign versus harmful test cases.
Abstract; Experimental design and Results