Zombie Predictions and the Future of Bail Reform
63 Pages Posted: 27 Sep 2017 Last revised: 29 Sep 2017
Date Written: September 22, 2017
In the last five years, legislatures in all fifty states have made changes to their pretrial justice systems. Most aim to shrink jails by incarcerating fewer poor, low-risk defendants. The goal is to make risk, rather than wealth, the key determinant of bail decisions for all defendants, reserving incarceration for those most likely to skip bail or to threaten the public if released. At the same time, jurisdictions around the country are exploring new ways to reduce the risks associated with releasing defendants, so that more will become suitable for release. These include text messages to remind defendants of upcoming court dates, which are surprisingly effective at bringing them back to court; ankle-worn GPS monitors that aim to track defendants’ whereabouts without upending their lives; and drug treatment and other supports that can help avert future entanglement with the law.
Predicting risk is the linchpin of these reform efforts, and many jurisdictions are now adopting statistical tools that leverage historical data to forecast which defendants can safely be released. Such tools are often integral to the political bargain that lets other bail reforms proceed, since they boost policymakers’ confidence that dangerous defendants will stay behind bars. But there is a stark and little-remarked problem at the heart of these pretrial risk-assessment tools: They aren’t using the right data.
In most cases, these newly adopted tools base their predictions on static datasets that predate the current wave of reform. The result is zombie predictions, which are blind to any benefit from recent, risk-mitigating policy changes. At the same moment as they reduce the true risks of pretrial release, jurisdictions across the country are also adopting statistical tools that will blindly predict such risks remain as high as ever. This will overstate risk, diminish the apparent impact of genuinely helpful reforms, and needlessly incarcerate people who could safely be released.
To avoid these harms, jurisdictions must tailor their risk assessments based on fresh, local data, comparing predictions against subsequent outcomes to keep their predictions on target. If they fail to do so, then the current attempts at bail reform may become one more chapter in a history of failed reform efforts and unintended consequences that is nearly as old as money bail itself.
Keywords: Bail, Risk Assessment, Algorithms, Machine Learning, Civil Rights
Suggested Citation: Suggested Citation