Toward an Integrative Study of Human-AI Interaction

Abstract

As artificial intelligence (AI) systems are increasingly embedded in the workflows of individuals and groups, designers and researchers of human-AI interaction (HAI) navigate a vast design space of possible configurations, making decisions that span algorithmic parameters, interface choice, and interaction protocols. This thesis develops an integrative approach that examines how design factors combine and interact to determine the outcomes of human-AI collaboration.

Chapter 1 synthesizes prior HAI research into a coherent design space framework encompassing algorithms, interfaces, users, and task settings, motivating a research program for systematic exploration of interdependencies between these factors. Chapters 2 and 3 turn to group-AI interaction through large-scale behavioral experiments. Chapter 2 investigates how social information---both direct conversation and peer behavior indicators---affects individual reliance on algorithmic decision support. The study reveals that while social information modulates the effects of performance feedback and model explanations on reliance, it does not improve predictive accuracy, illuminating critical tensions between social mechanisms and system design. Chapter 3 examines large language models as facilitators of group deliberation in hidden profile tasks. While LLM facilitation increased information sharing volume, density, and breadth, it did not improve decision quality, highlighting fundamental challenges in group-AI system design beyond information aggregation.

Chapter 4 advances an integrative approach to HAI research, emphasizing shared design spaces, systematic exploration strategies, and predictive models that generalize across contexts. The chapter provides methodological guidance and a tractable roadmap for advancing this integrative research agenda, laying the foundation for a more context-aware science of human-AI collaboration.