Algorithms have entered the center of many decision making processes, either by providing predictions or assessments to facilitate human decision making or, in some scenarios, suggesting decisions directly. More and more attention has been placed to ensure algorithms satisfy some desirable notion of fairness, which is an important step forward. In this talk, I will discuss the importance of examining algorithmic decision-making and algorithm-facilitated decision making in the broader context of intended applications and in the lens of human-algorithm interactions. I will first present a welfare-based analysis of fair classification algorithms to assess the welfare impact of fairness-constrained classification algorithms in the context of financial lending. Our analysis shows that applying stricter  fairness constraints in the algorithms can worsen welfare outcomes of all groups. Then, I’ll discuss a sequence of controlled human-subject experiments studying how the interactions between people and algorithms influence human decision making. In our experiments in two contexts (pretrial release and financial lending), when presented with algorithmic risk assessments, participants exhibited additional bias in their decisions and showed a change in their decision-making process by increasing risk aversion.

This talk is based on joint works with Lily Hu and Ben Green.