The Five Pitfalls of Data-Driven Sentencing

Abstract Artificial intelligence and analytics are often touted as high-tech solutions that will make sentencing fairer, more sophisticated, or more objective. But as data scientists with extensive firsthand experience analyzing federal sentencing data, this article’s authors see several pitfalls that could not only prevent data-driven sentencing from achieving these goals but may in fact reinforce the very problems judges seek to remedy. The article outlines five such pitfalls: (1) unbalanced data, (2) biased data, (3) inconsistent data, (4) averages masking variation, and (5) improper modeling choices. It concludes by briefly discussing ways in which these challenges can be overcome.