Make the First Step Act a smarter step by opening the risk assessment black box

In the fraught landscape of national politics, criminal justice reform is providing hope that Republicans and Democrats can agree on something. The First Step Act  — which Mitch McConnell has promised to bring to the senate floor, following passage in the House and the President’s endorsement — would expand good time credits, strengthen bonds between prisoners and their families, better prepare inmates for reentry, and retroactively reduce the harsh and unjust disparity between crack and cocaine sentences.

In other words, the act would make important and common sense (if exceedingly modest and overdue) reforms.

Why then have some of the justice groups that potentially have the most to benefit from the bill opposed it?

Much of the answer lies in the deep distrust that underlies and pervades many aspects of our criminal justice system. Time and time again, common-sense sounding policies — like money bail to prevent flight risk and sentencing guidelines to increase consistency — have instead entrenched wealth and race-based inequities.{mosads}

One little-noticed provision in the bill has raised just such alarm. The Act specifies the use of algorithmic risk and needs assessment to decide which prisoners get what programming, incentives, and rewards, including shortened sentences. If passed, the Attorney General would be required to develop and adopt a risk and needs assessment system that would be used to provide prisoners with tailored treatment based on whether they were minimum, low, medium, or high risk.

While we applaud the use of evidence to reduce recidivism, the reliance on factors like a person’s history, educational background, and other demographic factors to classify them risks exacerbating and further embedding historical and institutional patterns of bias, particularly against individuals of color. The exclusion of juvenile lifers is also an issue.

Similar concerns are being raised by the civil rights community about the application of risk assessment tools in pretrial detention contexts in dozens of jurisdictions and of sentencing tools in 28 states. These automated systems, often built on incomplete and biased data, are part of what Michelle Alexander has called the Newest Jim Crow and Virginia Eubanks has called the “digital poorhouse.”

In the context of prison risks and needs, the Senate version of First Step acknowledges the problem. It requires the Attorney General to consult with an Independent Review Committee when developing its tool. It requires regular reporting not only on overall recidivism — which, at 83 percent at 9-years at the state level, is unacceptably high — but also unwarranted disparities among similarly classified prisoners of different demographic groups. Importantly, it requires the system to be based on dynamic, not just static, factors within the prisoner’s control.

And yet, ultimately, it falls short. The Attorney General is not accountable to the Independent Review Committee, membership in which is highly restricted. There is no mechanism for sharing prisoner-level data, nor for disclosing details about what factors will be relied upon by the risk assessment system or how. Finally, it is unclear whether prison administrators will be required to follow the tool’s recommendations — studies of similar systems show that judges often deviate, leading, in the case of Kentucky for example, to judges overruling the presumptive default to deny release without bail to persons of low or moderate risk two-thirds of the time.{mossecondads}

Lawmakers should add — or the DOJ should implement — provisions specifying that de-identified data, including metrics of interest to stakeholders will be aggregated and collected in a consistent and standardized way across systems and shared with researchers and relevant communities (along with details about the resulting system). Members of the civil rights community should be invited to serve on the Independent Review Committee.

These steps would invite collaboration and innovation in designing fairer assessment systems — an admittedly difficult task. Greater transparency would also create trust and accountability. Where “trust me” doesn’t work, “see for yourself” just might.

The standardized collection and release of clean, de-identified, federated, prisoner-level, rather than aggregate data would empower not just the government, but researchers, faith-based organizations, and justice-involved groups to discover what works. It would build the thus far neglected data infrastructure of evidence that the President’s staff has called for, shoring up the foundation for serious work on how to best balance the interests of public safety and the risks and needs of all prisoners, including violent offenders. It would also fortify the ability of Congress and the GAO to exercise meaningful oversight.

These small changes — prior to passage or in implementation — would go a long way to operationalize the checks and balances badly needed by our criminal justice system. They would spur data-driven innovation and discovery of what works. They would pave the way for not just a First Step but a new direction for our criminal justice system.

Colleen V. Chien is professor of law at Santa Clara University and a fellow at the Stanford Computational Policy Lab; formerly she worked in the White House Office of Science and Technology Policy. Follow her on Twitter @colleen_chien

Clarence Wardell III, PhD, is director of City Solutions at Results for America; formerly he was co-lead in the White House Police Data Initiative and a Public Interest Technology Fellow at New America. Follow him on Twitter @cwardell