Inaccurate Empathy is Worse than Minimal Empathy in Affective Human-Robot Interaction

Simulated empathy is valuable for conversational agents and robots in emotionally sensitive domains such as counseling, healthcare, and customer service. However, the appropriateness of different types of empathy for different task domains is under-explored. We describe the results from a study comparing three approaches to the automatic generation of empathic responses for a conversational robot that conducts a depression screening conversation with a user. The computational models of empathy incorporate both rule-driven, theory-based responses with data-driven responses from a transformer model. Our primary finding i s t hat a sour empathic response models increase in sophistication their empathic accuracy decreases, resulting in more erroneous responses. Empathic inaccuracy, in turn, led to significantly lower ratings o f working a lliance, a measure of user trust in and liking of the robot.