An Algorithm Is Replacing Bail Hearings in New Jersey

February 23rd, 2017

Awake Goy

New Jersey is trying a new algorithm to fix its broken bail system, a flashpoint for criminal justice advocates who argue that court-assessed fines can discriminate against low-income and highly policed communities—most often, people of color.

Guidelines for how judges set bail vary across the country, but generally use a combination of a bail schedule, which prices out fees for specific offenses, and their own assessment of whether the defendant will appear at their hearing or commit a crime before their trial. If you can’t pay up, you stay in jail until your trial date, sometimes for up to a month.

On January 1, New Jersey replaced its bail system with an algorithm designed to mathematically assess the risk of defendants fleeing or committing a crime—particularly a violent one—before their trial date. The algorithm, called the Public Safety Assessment, was designed by the Texas-based Laura and John Arnold Foundation, a nonprofit that tries to fund innovative solutions to criminal justice reform.

New Jersey isn’t the first state to use algorithms to help judges suss out high-risk defendants. Counties across the country have tried using computer-based techniques to flag those who should continue to be detained until trial, and those who are flight risks.

But the algorithms are not without flaws. Last year, investigative reporting by ProPublica revealed that these programs had in-built racial biases, too. The software assessed risk based on data points gleaned from interviews with defendants, including questions about ZIP codes, educational attainment, and family history of incarceration—all of which can serve as proxies for race.

This system is different, according to Matt Alsdorf, vice president of the foundation’s Criminal Justice Initiative. The initiative assembled a dataset of more than 100,000 individual cases, and looked for factors re-offenders had in common. His team found that the data points that were the most closely correlated with race weren’t actually terribly useful. However, the foundation has yet to release the dataset for public analysis.