Bayesian reasoning for overcoming over-reliance in AI-assisted decision making

The tendency of users to blindly follow the suggestions of AI-powered decision support systems is a troubling phenomenon that may lead to unjustifiable errors and unreasonable decisions. Known as over-reliance, it is especially problematic when decisions involve high uncertainty and high costs for all parties. In this paper we discuss over-reliance as a factor that worsens the quality of the decision-making process and introduces unwanted and uncontrollable noise, like other cognitive biases. We propose to approach over-reliance within a Bayesian account of rationality and to model it as a form of base rate neglect, a well-known bias violating sound probabilistic reasoning. Finally, based on existing studies both on over-reliance in human-computer interaction (HCI) and on cognitive biases in human reasoning, we suggest a form of interaction that may be useful for mitigating over-reliance.