Not What, But How: Communicative Audit Reveals LLMs Frame Subjective Answers With Systematic Bias
arXiv·low signal
Evaluates how LLMs communicate answers to subjective questions — not whether answers are correct, but whether response framing introduces bias. Finds that users are sensitive to how information is presented, and current LLM evaluations miss this dimension entirely. For builders using LLMs in user-facing applications (chatbots, advisors, tutors), framing bias can undermine trust even when factual content is accurate.