Return to Article Details
Preferred but Less Safe? Human Preference–Safety Conflict in Paired AI Comparison Data