Evaluating the Impact of Anthropomorphism and Verbal Interaction on Users’ Perception of a Virtual Agent in Facilitating Self-Disclosure

Virtual agents, computer-controlled virtual characters, have been used in several systems for mental health purposes. Some research on their most effective features has shown that anthropomorphism improves trustworthiness and adherence to treatment. However, it has been suggested that non-humanoid agents, or even non-embodied ones, can be used in tasks involving personal disclosure or information recall. After our literature review, some questions arose regarding the best features for an agent in the role of an active listener for women to disclose complex life events: Would a human-like or non-humanoid visual be better received? And if representing a human, how relevant would it be for the agent to speak? To find answers, we created a speaking humanoid, a mute humanoid, and a speaking non-humanoid agent and compared them in a personal-disclosure context. In a repeated-measures experiment, 21 women roleplayed the disclosure of challenging situations with each agent. We measured their emotional responses, affinity, empathy, social presence, social closeness, and likability for the agents. Our results showed that having a human-like appearance had a significant effect on affinity (p<0.009), social presence (p<0.025), social closeness (p<0.001), and likability (p<0.031) but not on their emotional responses or levels of empathy. The humanoids’ speaking ability did not significantly affect any measures but slightly improved feelings of social closeness, identifiability, and perceived awareness. Our results suggest that female users prefer speaking human-like agents for disclosing complex issues, but some could still prefer speaking non-humanoid agents.