Random Mechanical Turkers out-predicted the crime algorithm used in US court sentencing

A Dartmouth study published in Science Advances found that COMPAS, the recidivism-risk algorithm used in US courts to inform bail and sentencing, predicted reoffending correctly only about 65 percent of the time. Random Amazon Mechanical Turk workers did better, at a median 67 percent, and so did a simple model using just a defendant’s age and criminal record. The developer called the findings highly misleading, but the study warned that flawed risk scores can push low-risk defendants toward longer sentences.