Description:

Title : Human–AI Collaboration: Performance, Uncertainty, and Human Preferences

Abstract :

This talk presents an overview of recent research on human–AI collaboration, focusing on the conditions under which collaboration improves performance and the challenges that arise as AI systems become increasingly capable. I begin by reviewing several collaborative workflows, including non-interactive statistical aggregation, AI-assisted decision-making in which humans receive AI advice, and agentic settings where humans and AI systems jointly act in dynamic environments. Across these paradigms, I review empirical findings on when and why collaboration improves or fails.

A central challenge is achieving complementarity—cases in which joint human–AI performance exceeds that of either humans or AI alone. I discuss empirical evidence on when complementarity arises and key factors that limit it. As AI systems increasingly outperform humans, opportunities for complementarity diminish, shifting the problem toward managing asymmetric collaborations. In these settings, effective uncertainty communication is essential so that humans can appropriately rely on AI outputs.

I review recent evidence showing that large language models maintain internal signals of uncertainty but often fail to communicate this uncertainty effectively to users, leading to overconfidence and overreliance. Behavioral studies demonstrate substantial gaps between model confidence and human perceptions of confidence, which are further amplified by features such as explanation length. However, targeted fine-tuning can improve uncertainty communication by aligning verbal expressions of confidence with internal reliability signals, improving both calibration and discrimination.

Finally, I discuss how human subjective evaluations shape collaboration, especially in agentic contexts. Beyond accuracy and efficiency, people value AI collaborators that behave cooperatively, allow meaningful human contribution, and are enjoyable to work with. Together, these findings highlight the importance of aligning performance, uncertainty communication, and human preferences in the design of collaborative AI systems.