Presentation
Evaluating the Impact of Consequence on Trust in AI-Aided Tasks: An Eyetracking Study
DescriptionArtificial intelligence (AI) is increasingly used in high-risk domains, where the cost of error (i.e., consequence severity) varies. Understanding how consequences influence human trust in AI is important for safe and effective collaboration. This study examined the influence of consequences on trust in an AI-assisted mental rotation task under low- and high-consequence conditions. Thirty-five participants performed the task with support from an imperfect AI. Eye-tracking data, trust ratings, and performance metrics were collected. When the AI was incorrect, participants showed significantly longer fixations on reference images and answer choices in high-consequence conditions. When the AI was correct, fixation durations on the AI’s suggestions were significantly shorter, also under high-consequence conditions. These results indicate reduced reliance on AI in high-consequence scenarios. Notably, participants avoided making initial errors in high-consequence scenarios, indicating increased vigilance. These findings indicate that consequence severity influences human trust and attention during AI-aided decision-making. Kim et al. (2025) highlighted a significant decline in trust following “false reassurance”, when AI erroneously confirmed users’ wrong initial actions. Future research should increase trial volume and task complexity to explore how consequence severity affects trust across various human-AI error patterns, particularly following wrong human initial actions.
Event Type
Creating AI that Works for People
Lecture
TimeThursday, October 16th11:50am - 12:10pm CDT
LocationGrand B
Cognitive Engineering & Decision Making
Creating AI that Works for People: Human-Centered Innovation