TherapyProbe: Generating Design Knowledge for Relational Safety in Mental Health Chatbots Through Adversarial Simulation

Researchers introduce TherapyProbe, a methodology to identify relational safety failures in mental health chatbots through adversarial simulation. The study reveals dangerous interaction patterns like 'validation spirals' and creates a Safety Pattern Library with 23 failure archetypes and design recommendations.

• →TherapyProbe methodology uses adversarial multi-agent simulation to identify harmful conversation patterns in mental health chatbots.
• →Research identifies critical safety failures like 'validation spirals' and 'empathy fatigue' that emerge over multiple conversation turns.
• →A Safety Pattern Library catalogues 23 failure archetypes with corresponding design recommendations for developers.
• →Current safety evaluations focus on single-turn responses but miss therapeutic dynamics that unfold over time.
• →The methodology is replicable and uses open-source models without requiring expensive API costs.