Human-AI Collaboration in Groups
Conversational agents that work with several people at once
Most conversational agents are built for one user at a time. Groups break that assumption: turn-taking is contested, and whatever the agent says to one person is heard by everyone else. We study how agents should behave when several people are in the conversation, and what design choices are available to someone building one.
In a systematic review of 53 studies of group conversational agents [Yeo, S.*, Zhang, T.*, Bateman, S., Hsieh, G., Kim, Y., Perrault, S., Li, J., and Tang, A. (2026). Group Conversational Agents: A Review of Designs that Support and Shape Group Interaction. In Proceedings of the 2026 ACM Designing Interactive Systems Conference (DIS '26).] , we characterised how these systems intervene in group-level processes, and identified where designers currently work without guidance.
Shaping how groups talk
An agent that joins a group conversation changes who speaks and what gets said. In two user studies, with 20 and 200 participants, we asked whether multimodal reflective nudges improve the quality of online deliberation [Yeo, S., Jiang, Z., Tang, A., and Perrault, S. (2025). Enhancing Deliberativeness: Evaluating the Impact of Multimodal Reflection Nudges. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, 515.] . We also built TeamWise, an on-screen avatar that joins video onboarding meetings and generates its turns from what participants say. A formative study followed a team working with it [Obilisetty, V., Elleby, M., Tang, A., and Wang, A. (2026). TeamWise: Exploring Virtually Embodied AI Facilitation for Video-Based Team Onboarding. In Extended Abstracts of the 2026 CHI Conference on Human Factors in Computing Systems (CHI EA '26).] .
Agents in relationships
In some group settings the relationship between the participants matters more than the task at hand. In scenario-based design interviews, we had ten pairs of family members design chatbot interventions for intergenerational conflict [Zhang, T., Poh, E., Hou, Y., Lee, Y., Zhang, R., Li, J., and Tang, A. (2026). ``Grandpa, Can You Speak Nicer?'': Envisioned Chatbot Roles and Design Tensions in Intergenerational Communication Conflicts. In Proceedings of the 2026 ACM Designing Interactive Systems Conference (DIS '26).] . An agent asked to soften one party’s message is negotiating between people who have to keep talking afterwards. We also interviewed 29 daters about AI-mediated self-disclosure in online dating [Tsai, P., Zhang, T., Poh, E., Tang, A., and Chang, Y. (2026). Not Too Early, Not All at Once: Design Tensions in AI-Mediated Self-Disclosure in Online Dating. In Proceedings of the 2026 ACM Designing Interactive Systems Conference (DIS '26).] , where the design questions concern the timing and pacing of what gets revealed.
Who else is listening
An agent that speaks aloud in a shared space has an audience it did not choose. We ran a cross-cultural vignette study with 944 participants in Germany and Singapore, testing how bystander relationships, location, and topic shape what people want a smart speaker to say [Warin, L., Tang, A., Aurelia, E., Misra, A., and Reinhardt, D. (2026). ``Alexa, Do Not Say That in Front of my Boss!'' A Cross-Cultural Comparison of User and AI Preferences for Privacy-Aware Smart Speaker Interactions Across Contexts. In Proceedings on Privacy Enhancing Technologies (PoPETs), 724--739.] .
Publications
Acceptance: 21% - 248/1154. Honourable Mention - Top 5% of all submissions. *Co-first authors.
Acceptance: 27.0% - 1198/4444.
Acceptance: 38.2% - 789/2068.
Acceptance: 21% - 248/1154.
Acceptance: 21% - 248/1154.