Michigan’s Department of Corrections plans to pilot an AI system designed to assess which incarcerated people may pose a risk of violence. The assessments will place prisoners in one of five security levels, which can affect their housing and access to programs and other resources.
The department expects the system to make risk predictions 30% more accurately than current methods. The pilot is scheduled to begin in early December 2026 at several men’s prisons. The stated goal is to address a recent increase in attacks on staff and violence between incarcerated people.
The tool was developed with Recidiviz, a nonprofit organization that originated from Google. Department director Heidi E. Washington said the pilot uses scientific and historical data to create what she described as a dynamic, objective tool to improve facility safety. She did not specify which data the AI system will use.
That information matters because the system’s assessments could influence prisoners’ security classifications, placement and available resources. An investigation by ProPublica into a similar risk-assessment program in Louisiana found that the data used there contained racial and socioeconomic biases, raising concerns about the consequences of errors or uncorrected bias in such systems.
Sources: t3n.