Abstract

High-quality user feedback is essential for effective human-AI interaction. It bridges knowledge gaps, corrects drift, and shapes system behavior; both during the interaction and throughout model development.  However, despite its importance, in-the-wild human-AI conversations contain sparse low-quality feedback. Given the central role human feedback plays, it necessitates a critical examination of human feedback during interactions with CAs. To understand and overcome the Feedback Barriers that prevent users from giving high-quality feedback, we ran two studies examining the feedback dynamics in human-AI interactions. Our formative study identified four feedback barriers through violations of Grice’s maxims—Relation, Quality, Manner, and Quantity—by both the users and CAs. We then derived design desiderata aimed to minimize maxim violations and found that scaffolds reducing maxim violations encouraged user interaction and higher-quality feedback during our second study. Finally, we propose a call for action to create CAs that foster healthy feedback dynamics.

Advisor

Ziang Xiao