Researchers from the University of Kansas and the University of Southern California have conducted a groundbreaking study examining the nonverbal behaviors that indicate connection among participants in support groups. The findings, published in the Proceedings of the 27th International Conference on Multimodal Interaction, highlight the potential of machine learning to enhance mental health services, especially in a post-pandemic world where demand for support has surged.
The study analyzed data from 18 support groups, involving a total of 96 participants, all of whom were dealing with general anxiety. Researchers focused on measuring dyadic alliance—the bond between two individuals—by surveying participants on their feelings of connection during online sessions. These sessions were facilitated through video conferencing by a virtual conversational agent, designed to promote engagement while a human operator monitored the interactions.
Yunwen Wang, an assistant professor of journalism and mass communications at KU and a lead author of the study, emphasized the growing need for mental health support, particularly exacerbated by the COVID-19 pandemic. “We were examining the burnout among mental health professionals as the demand for support continued to rise,” Wang stated. “This project aims to explore how we can ethically use artificial intelligence to enhance access to mental health resources rather than replace human therapists.”
During the study, participants’ verbal and nonverbal communications were meticulously recorded and analyzed. This included gestures like head nodding, smiling, and eyebrow movements, alongside vocal attributes such as pitch variation and speech intensity. These behaviors were evaluated using computational algorithms to ascertain their correlation with participants’ reported feelings of connection.
The analysis revealed that specific nonverbal cues, such as frequent head nods and brow raises from the listener, significantly increased the speaker’s sense of alliance. Conversely, for speakers, variations in pitch and less intense facial expressions also fostered stronger connections. These insights suggest that both verbal and nonverbal communication play crucial roles in establishing rapport among group members.
As mental health services face unprecedented demand, the study opens the door for further exploration into the integration of machine learning in therapeutic settings. Wang noted, “While our findings indicate the potential for AI to help identify behavioral markers of connection, we are not advocating for its unrestricted use in mental health. Further research is essential to understand the ethical boundaries and practical applications of AI in this sensitive field.”
The research team is also investigating how AI agents might influence user trust and engagement in support groups, particularly in more serious contexts such as substance-use disorders. Wang highlighted the importance of understanding how users perceive AI involvement in their mental health journeys. “We aim to determine whether AI-facilitated support can be seen as acceptable, especially given the limited number of trained professionals available.”
Ultimately, the study aims to enhance mental health services by improving understanding of how genuine connections are fostered in group settings. Wang concluded, “The human-to-human dynamic remains essential for participants to share their experiences and offer empathy. In this context, the AI agent serves as a facilitator, allowing for richer interactions among individuals.”
The ongoing research seeks to address critical questions surrounding the role of AI in mental health, including the implications for privacy and security. As discussions about AI in therapeutic contexts grow, the findings from this study contribute to a broader dialogue on how technology can responsibly support mental health initiatives.
For further details, refer to the study: Kevin Hyekang Joo et al, “Multimodal Behavioral Characterization of Dyadic Alliance in Support Groups,” Proceedings of the 27th International Conference on Multimodal Interaction (2025). DOI: 10.1145/3716553.3750818.
