Measuring Racial Discrimination in Algorithms

Algorithmic decision-making can lead to discrimination against legally protected groups, but measuring such is often hampered by a fundamental selection challenge. We develop new quasi-experimental tools overcome this challenge and measure algorithmic in pretrial bail decisions. show that the reduces of four moments, which be estimated extrapolating variation across as-good-as-randomly assigned decision-makers. Estimates from New York City both sophisticated machine learning algorithm simpler regression model discriminate Black defendants even though defendant race ethnicity are not included training data.