Not What, But How: A Framework for Auditing LLM Responses across Positioning, Generalization, Anthropomorphism, and Maxims
FRANZ is an automated framework to audit how Large Language Models frame subjective answers by measuring four dimensions: cultural positioning, generalizing language, anthropomorphism cues, and adherence to conversational maxims. Using a large corpus of Reddit questions, the study found that insider framing and anthropomorphism are tightly linked across different countries.