Designing AI agents to improve virtual meeting inclusion

Abstract

Widespread shifts to remote work have transformed many workplace meetings from in-person gatherings into computer-mediated interactions, creating opportunities for developing tools to improve them at scale. However, deploying effective technical interventions requires understanding users’ values, design preferences, and outcomes in real workplace contexts. This dissertation investigates how artificial intelligence (AI) agents can be designed to improve inclusion in workplace meetings through three interconnected projects blending qualitative interviews, user-centered design, a lab study, and a field deployment. The first project investigates causes of bias—inequitable barriers to participation—in video conferencing (VC) meeting software. Through a qualitative interview study, I find that VC platforms encourage meeting leaders to exert control over key social and technical parameters, giving them outsize influence over bias. To mitigate bias, I recommend: (1) transferring control from meeting leaders to technical systems or other participants, and (2) supporting leaders in exercising control more equitably. The second project examines how an AI meeting assistant could operationalize these recommendations, shifting focus from bias to inclusion—defined as attendees’ ability to participate to their desired extent. Through user-centered design, I formalize the “Observe, Ask, Intervene” (OAI) framework for AI-based meeting tools. I build a prototype instantiating the framework and evaluate it in a lab study, finding that participants’ stated preference for private interventions made the system ineffective at changing behavior. The third project reports on an OAI-based system that overcomes this limitation by employing the induced hypocrisy procedure. Through a field deployment, I demonstrate that the system can affect behavior and identify new design tensions to consider in future iterations. Major contributions include: (1) a framework for identifying ways VC meeting software might hinder inclusion, (2) the “Observe, Ask, Intervene” (OAI) framework for designing AI meeting agents, and (3) an effective prototype instantiating OAI. Through these contributions, this dissertation forwards our understanding of how AI agents can be used to mediate social and interactional challenges between users in the workplace. I discuss broader applicability of these contributions to AI-supported collaborative work and current challenges in human-AI interaction.