Algorithmic Behavioral Science: Automated Discovery of Human Biases
Science begins with something distinctly non-scientific. Scientists meticulously test hypotheses that themselves come from a very messy place: a mix of creativity, intuition, observation and chance. We argue machine learning can play a more rigorous role here. We illustrate this in a problem that is independently interesting: judges deciding whether to jail someone. A deep learning algorithm trained on past data discovers a striking behavioral bias: the pixels in a defendant’s face alone accounts for 30 to 50% of the explainable variation in whether a defendant is jailed. This finding is not explained by race, skin color, demographics or well-understood facial features from psychology. To make the discovery usable, we develop a procedure that allows the algorithm to communicate what it is seeing in the face. It leads us to identify interpretable facial features, previously not considered, that bias the way judges treat defendants; tests on independent data, unseen by the algorithm, suggest this bias is quantitatively large, bigger for example than the effect of race. We suggest that this technique can be more broadly applied and suggests a new way to generate meaningful hypotheses.
Bio: Sendhil Mullainathan is the Roman Family University Professor of Computation and Behavioral Science at Chicago Booth. His current research uses machine learning to understand complex problems in human behavior, social policy, and especially medicine, where computational techniques have the potential to uncover biomedical insights from large-scale health data. He currently teaches a course on Artificial Intelligence. Full bio here.
