SRI Seminar Series: “The power of discussion: Designing useful communication with AI agents”

Our weekly SRI Seminar Series welcomes Anastasia Kuzminykh, assistant professor at the University of Toronto’s Faculty of Information and a faculty affiliate at the Schwartz Reisman Institute. With a background spanning computer science, psychology, and ethnographic research, Kuzminykh’s work bridges human-computer interaction and the design of intelligent systems.

Her current research focuses on how people interact with AI systems through conversation—exploring trust, reliability, and user perception in human–agent communication. Through The COoKIE Group, Kuzminykh studies how design decisions shape ethical and effective human–AI collaboration, especially in contexts involving large language models, generative tools, and decision-support systems.

Moderator: Avery Slater, Department of English & Drama

Abstract:

The rapid development of LLM-based interfaces has made interactions with AI systems easily accessible, dynamically adaptable, and most importantly, comprehensible for a wide audience, leading to the unprecedented current popularity of human-AI collaboration. Shifting to this human-agent communication paradigm, empowered by the intuitive nature of conversational interfaces, quickly opened exciting opportunities for users and has already significantly reshaped their information-seeking practices, decision-making processes, creative activities, etc. However, alongside all the promising advances, we started observing numerous worrisome threats to the quality of such human-AI collaboration, including, for example, users’ overreliance on AI, the “gulf of envisioning” challenge with formulating effective prompts, the models’ tendency for sycophantic behaviours, or the issues of potential amplification of biases and intensification of echo chambers. In this talk, I will cover some of the recent work done with my research group toward addressing these challenges and discuss the role of conversation architecture in supporting efficient human-AI collaboration.