UC Irvine

Collaboration with other agents requires people to reflect on several aspects of their collaborator's capabilities: How good are they at a task? Do they have access to the same information as I do? How useful is their advice? This dissertation takes a closer look at how people answer these questions while interacting with other humans and more importantly, with AI agents. First, we discuss the role of mental models in facilitating people's interactions with other agents. In Chapter 2, we present a computational framework to understand how people assess the ability of other humans and AI agents. Our research reveals a discrepancy in people's assessments: individuals accurately gauge another human's ability on a task, but consistently overestimate an AI agent's performance on the same task. In Chapter 3, we examine how people navigate advice when a human or an AI advisor has access to different information than they do. We find that while people take into account this difference in information, they consistently underestimate the value of advice. Finally, in Chapter 4, using two case studies we demonstrate the utility of cognitive modeling in inferring people's latent reliance on AI. The first case study models people's decision to accept advice, and the second case study models people's decision to intervene in the AI's course of action. We conclude by discussing avenues for future work. Altogether, this dissertation uses a cognitive science lens to improve our understanding of human-AI interaction, highlighting the role of mental models and drawing valuable lessons from human-human interaction.