Lab co-director Madalina Busuioc investigates the cognitive biases that arise in decision-makers' interaction with artificial intelligence decisional aids.
Artificial intelligence algorithms currently serve as decisional aides to human decision-makers in many bureaucratic contexts. In three experimental studies, the research explicitly examines how decision-makers rely on algorithmic recommendations and what impact this has on their decision-making. Two potential types of biases formed the focus of the study: automation bias – undue overreliance on algorithmic advice despite evidence to the contrary and selective adherence – biased adherence to algorithmic advice when predictions confirm decision-makers’ priors, beliefs, or biases.
The study finds evidence of selective adherence to algorithms: namely, higher deference to algorithmic recommendations when predictions align with prevalent stereotypes, compounding bias in decision-making at the expense of already marginalised communities and individuals. The research helps unpack an important cognitive mechanism through which decisional biases can become exacerbated in the interaction with automation.