Paper Type

Complete

Abstract

Communicating uncertainty is essential for effective human–AI collaboration because it shapes user trust and reduces both overreliance and unwarranted aversion. Although AI predictions are inherently probabilistic, the uncertainty surrounding them can stem from different sources, including randomness in the data (aleatoric uncertainty) and limitations in the model’s learned knowledge (epistemic uncertainty). This study examines how these distinct forms of uncertainty differentially affect trust and decision behavior. We conduct a two-stage controlled experiment (N = 240) using a high-stakes health insurance cost prediction task. Results show that both forms of uncertainty significantly reduce behavioral reliance on AI predictions compared to a no-uncertainty baseline. However, uncertainty arising from data randomness produces a stronger reduction in behavioral reliance than uncertainty from knowledge limitations. At the same time, transparent communication of uncertainty modestly improves perceived satisfaction and system usefulness for both uncertainty types. These findings highlight that the source of uncertainty is not interchangeable in its effects on trust, and inform the design of more effective uncertainty communication strategies in high-stakes AI applications.

Paper Number

1986

Comments

SIG HCI

Share

COinS
 
Aug 15th, 12:00 AM

More Than Confidence Scores: Do Different Types of Uncertainty Matter for Trust in AI?

Communicating uncertainty is essential for effective human–AI collaboration because it shapes user trust and reduces both overreliance and unwarranted aversion. Although AI predictions are inherently probabilistic, the uncertainty surrounding them can stem from different sources, including randomness in the data (aleatoric uncertainty) and limitations in the model’s learned knowledge (epistemic uncertainty). This study examines how these distinct forms of uncertainty differentially affect trust and decision behavior. We conduct a two-stage controlled experiment (N = 240) using a high-stakes health insurance cost prediction task. Results show that both forms of uncertainty significantly reduce behavioral reliance on AI predictions compared to a no-uncertainty baseline. However, uncertainty arising from data randomness produces a stronger reduction in behavioral reliance than uncertainty from knowledge limitations. At the same time, transparent communication of uncertainty modestly improves perceived satisfaction and system usefulness for both uncertainty types. These findings highlight that the source of uncertainty is not interchangeable in its effects on trust, and inform the design of more effective uncertainty communication strategies in high-stakes AI applications.