Algorithmic Predictions and Bureaucratic Behavior Adjustment

Abstract

The use of algorithms for policy is surrounded by high expectations but little is known about how their predictions are factored into decisions by the deciding bureaucrats. We introduce a framework of the interplay between algorithms and bureaucrats and motivate several mechanisms of behavior adjustment. We then provide evidence on the adjustment mechanisms from a field intervention in public employment services where caseworkers received access to algorithmic predictions of jobseekers' unemployment duration in a randomly selected number of cases. We find that caseworkers changed their behavior but not in intended ways. Caseworkers increased the meeting frequency among jobseekers with already good predicted re-employment prospects. The behavior could be driven by performance incentives that were not aligned with the intervention and an avoidance of blame.