Security classification in prison is not an administrative formality. It shapes where incarcerated people are housed, how freely they move, and whether they can access the education and reentry programs that bear on life after release.
Michigan’s Department of Corrections is redesigning how it makes those calls. Starting December 1, 2026, MDOC will pilot a classification tool developed with assistance from Recidiviz at selected men’s facilities, with the system flagging cases that may warrant reclassification and surfacing them for staff review.
Recidiviz is a nonprofit criminal-justice technology organization. The tool does not make final decisions autonomously; MDOC spokesperson Jenni Riehle told the Detroit Free Press that staff review an assessment score before any classification change is made.
MDOC projects the redesigned system will improve its ability to predict incarcerated people’s future violence risk by approximately 30%. That figure comes from the department’s own announcement, and it arrives without a baseline, a validation sample, a statistical method, or a confidence range. Treat it as MDOC’s aspiration, not an established result.
“Classifying prisoners at the security level that fits their current behaviors and needs keeps everyone safer , the individual, other incarcerated people, and correctional staff.” Heidi E. Washington, Director, Michigan Department of Corrections.
Director Heidi E. Washington described the tool as “dynamic, objective” and said she expects it to strengthen facility security. The word “objective” is where the scrutiny needs to start. A classification model trained on historical corrections records inherits whatever patterns those records contain. Those patterns can reflect inconsistent enforcement, documentation gaps, and demographic disparities.
The human-oversight claim deserves careful reading. A software flag shapes which cases receive attention, what evidence feels most salient, and how quickly someone could be moved to a different security level. Decision support still shapes decisions.
MDOC operates four general-population security levels: Level I, Level II, Level IV, and Level V. Lower levels mean more movement and access to programming; higher levels impose progressively tighter controls. Dynamic reassessment could work in an incarcerated person’s favor, updating their classification faster when behavior genuinely warrants a change. It could also accelerate a move toward more restrictive conditions when a data-entry error or incomplete record triggers the wrong flag.
The pilot announcement does not specify what data fields the tool uses or how it weights recent behavior against older disciplinary records and convictions. It also does not state whether performance will be measured separately across racial, ethnic, age, or disability groups. The cited announcement and reporting do not describe any process by which incarcerated people or their attorneys could challenge an inaccurate record that elevated a classification score.
Credible accountability requires specific disclosures once the pilot runs: error rates, false-positive and false-negative rates, override frequency, and demographic performance breakdowns. MDOC also needs a clear, public process for correcting bad data. Staff need training that treats the score as advisory, not authoritative, and the pilot’s scope needs to stay defined. Its stated purpose is housing and supervision classification, not parole eligibility or release decisions.
Getting this right means publishing the pilot’s actual results, including its failures, not just optimistic projections. One press release and a launch date are not a transparency plan.
