2026. Preferred but Less Safe? Human Preference–Safety Conflict in Paired AI Comparison Data. AI and Wisdom. 1, 3 (Sep. 2026).