Machine Learning as a Tool for Hypothesis Generation.

This article presents a novel, semi-automated procedure for hypothesis generation that integrates machine learning algorithms with human interpretation to discover new, interpretable hypotheses from high-dimensional data. The procedure is illustrated through an application to judicial pretrial detention decisions in Mecklenburg County, North Carolina, where a deep learning model using only defendants’ mug shot images predicts judge decisions with substantial accuracy, capturing up to half of the predictable variation. By generating synthetic morphed images that vary in predicted detention risk and having independent human subjects identify distinguishing facial features, the study uncovers two novel facial characteristics—“well-groomed” and “heavy-faced”—that correlate with judges’ detention decisions beyond known demographic and psychological factors. These features are empirically plausible, not previously recognized by criminal justice practitioners, and suggest new avenues for causal testing, while the general procedure offers a replicable framework applicable to other domains involving rich, high-dimensional data.