When Pleasing the User Compromises Care: AI Sycophancy in Caregiver-Mediated Healthcare
An evidence-grounded benchmark is introduced spanning veterinary, pediatric, and older-adult care, and whether AI advisors maintain appropriate care when users express clinically inferior preferences is tested, and the results expose a broader limitation of user-centric alignment in settings with multiple stakeholders.