This paper examines the application of machine learning to improve and understand human decision-making, focusing on the critical policy issue of bail decisions by judges. The authors use a dataset of 758,027 defendants arrested in New York City between 2008 and 2013 to train a machine learning algorithm to predict the risk of defendants committing new crimes if released. The algorithm is evaluated against the decisions of judges, who must decide whether to release defendants or detain them based on a prediction of their flight risk. The study finds that judges often misrank defendants, releasing those predicted to be high-risk and detaining those predicted to be low-risk. This mis-ranking leads to higher crime rates and jail populations. The authors propose a policy simulation where all release decisions are based on predicted risk, showing that this approach could reduce crime by up to 24.7% without increasing jail populations. The results are robust to quasi-random assignment of cases to judges and are replicated in a national dataset covering 40 large urban counties. The study highlights the importance of integrating machine learning with economic frameworks to address selection biases and payoff functions in decision-making.This paper examines the application of machine learning to improve and understand human decision-making, focusing on the critical policy issue of bail decisions by judges. The authors use a dataset of 758,027 defendants arrested in New York City between 2008 and 2013 to train a machine learning algorithm to predict the risk of defendants committing new crimes if released. The algorithm is evaluated against the decisions of judges, who must decide whether to release defendants or detain them based on a prediction of their flight risk. The study finds that judges often misrank defendants, releasing those predicted to be high-risk and detaining those predicted to be low-risk. This mis-ranking leads to higher crime rates and jail populations. The authors propose a policy simulation where all release decisions are based on predicted risk, showing that this approach could reduce crime by up to 24.7% without increasing jail populations. The results are robust to quasi-random assignment of cases to judges and are replicated in a national dataset covering 40 large urban counties. The study highlights the importance of integrating machine learning with economic frameworks to address selection biases and payoff functions in decision-making.