Role-Based Response Asymmetry

Kaur et al. "Role-Based Response Asymmetry in Mental Health AI." arXiv:2510.16829, 2025.

Key findings used in wiki

• Models respond differently based on perceived user role (patient vs. clinician vs. caregiver)
• The same clinical content receives different safety treatment depending on who the model thinks is asking
• Caregiver personas receive less cautious responses than patient personas for equivalent risk levels
• Role-based asymmetry means safety evaluations using only one persona type underestimate real risk
• Directly motivates InvisibleBench's multi-persona evaluation approach across caregiver and care recipient roles