Michigan turns to Recidiviz AI for prison security placements

What the score can change

Until now, a committee appointed by each facility’s warden reviewed prisoners’ cases. The Michigan Department of Corrections says Recidiviz’s software will help “optimize prisoner placement.” The announcement leaves open whether the system will advise those committees or substantially shape their decisions.

Recidiviz began as a data project at Google, aimed at standardizing criminal-justice data and improving how the system works across the country. Its platform uses AI and reportedly helps identify people eligible for state programs, a security-level review or release, and connect them with resources such as temporary housing and job-search assistance.

Security classification is not a release decision, but it still has consequences: it can determine which programs and resources prisoners can access, as well as whom they are housed with during their sentences.

The missing details matter

The state has not said what data the system will use. That matters because the inputs can shape the classifications that follow.

A ProPublica investigation into a similar risk-assessment system in Louisiana found that the data used to classify prisoners carried racial and socioeconomic bias. The result, the investigation reported, was a criminal-justice system that became harsher toward people seriously pursuing rehabilitation.

In one case, a 70-year-old blind man who used a wheelchair was classified as having a “moderate risk” of reoffending despite having no disciplinary violations. The classification automatically made him ineligible for parole.

Michigan’s tool is meant to guide prison placement, not parole. Still, I think the Louisiana case is a warning against treating a prediction score as neutral simply because it is produced by software. The state’s roughly 30% figure is difficult to assess without knowing what “more accurately” means, what baseline it uses, or how the system will be evaluated.

The Michigan Department of Corrections was asked to comment. If the software is meant to improve placement, the key test will be whether it does so without turning past disparities in criminal-justice data into new constraints on prisoners’ lives.