appropriate refusals

**Appropriate refusals** is the **safety behavior where models refuse genuinely harmful requests while correctly allowing benign requests that use similar language** - appropriateness depends on intent-aware contextual interpretation. **What Is Appropriate refusals?** - **Definition**: Correct refusal decisions that align with policy and user intent rather than keyword triggers alone. - **Context Requirement**: Interpret domain meaning, ambiguity, and legitimate technical usage. - **Decision Quality**: Refuse when risk is real, assist when request is allowed. - **Common Challenge**: Lexical overlap between harmless and harmful contexts. **Why Appropriate refusals Matters** - **Safety Accuracy**: Avoids harmful compliance while reducing unnecessary denials. - **Usability Preservation**: Technical and educational users need valid non-harmful responses. - **Trust Building**: Consistent contextual judgment improves user confidence. - **Fairness Improvement**: Reduces over-blocking of legitimate speech patterns. - **Operational Efficiency**: Fewer mistaken refusals lower support and escalation burden. **How It Is Used in Practice** - **Intent Classification**: Combine semantic models and policy rules for context-aware decisioning. - **Ambiguity Handling**: Ask clarifying questions when harmful intent is uncertain. - **Evaluation Design**: Test on paired benign and harmful prompts with similar wording. Appropriate refusals is **a high-precision safety goal in LLM systems** - context-sensitive refusal behavior is essential to balance robust harm prevention with useful assistant performance.