Effects of Uncertainty and Knowledge Graph on Perception of Fairness

Abstract

AI-informed decision making consequential to individuals raises critical concerns on fairness. Fairness introduced by AI models and/or data is perceived by humans (also refers to perception of fairness) implicitly or explicitly. It is a central component of maintaining satisfactory relationships with humans in AI-informed decision making. Furthermore, model uncertainty and knowledge on training data play a crucial role in AI-informed decision making. This paper introduces model uncertainty and knowledge on training data represented by knowledge graphs into AI-informed decision making. We aim to investigate what uncertainty level and knowledge graph, and how they affect user perception of fairness in AI-informed decision making. A user study on judging the recidivism rate of prisoners found that uncertainty of model prediction of recidivism rate can benefit user perception of fairness, but only under low and medium uncertainty conditions. However, we did not find significant effects of knowledge represented by knowledge graph on user perception of fairness. These findings have wide implications in the user interface design of AI-informed decision making applications.

Ähnliche Arbeiten