Michigan to test prison classification system that officials say could improve violence predictions
Michigan plans to pilot a new system for assigning prison security levels, a decision that can shape where someone is housed, how closely they are supervised and what programs they can access. State officials say the redesigned process could improve predictions of future violence by about 30%. That is a notable claim, though the state has not yet shared the performance data behind it.
Despite descriptions of the project as prison artificial intelligence, the nonprofit that helped develop the software says AI does not calculate the security scores. Instead, the system applies a set of rules to information already in prison records. Corrections staff review the results and make the final decisions.
The Michigan Department of Corrections oversees roughly 32,000 incarcerated people. At present, staff use a manual scoring process to decide each person’s security classification. The outcome can affect housing and supervision, as well as access to education, vocational training and other opportunities. Michigan says its violence prevention program is available at all general population levels.
The state has four general population security levels: I, II, IV and V. Level I generally means more freedom of movement, while Level V involves tighter supervision. There is no general population Level III.
The pilot is scheduled to begin Dec. 1 at selected men’s prisons. If the trial goes well, Michigan may expand the system to more facilities. One proposed change is to let staff reassess classifications as new information comes in, rather than relying mainly on records gathered at scheduled intervals.
The software was developed with Recidiviz, a nonprofit that works with criminal justice agencies in 19 states. The organization reviewed a decade of Michigan classification data as part of its work on the new process. The system considers factors chosen by state corrections experts and assigns points according to established rules. More recent disciplinary incidents may count more heavily than older ones, while sustained good behavior or completion of rehabilitation programs could help someone qualify for a less restrictive level.
The software can also flag a record for review when new information becomes available. Staff can examine the score and, when circumstances call for it, override the result. Medical needs or personal safety concerns, for example, may affect whether a particular housing assignment is appropriate.
Recidiviz says it uses AI for some other tasks, such as helping staff prepare case notes and organize information from meetings. But the organization says AI does not determine a person’s security classification or release. In this case, the score comes from a formula that staff can inspect—not a system that independently makes the call.
That distinction matters, but it does not mean the process is automatically reliable. A rules-based score can still be affected by errors in the records it uses, and no assessment can predict with certainty what someone will do in the future.
Michigan’s estimate of a 30% improvement in predicting violence also needs more explanation. The state has not described how it arrived at that figure or released the results used to support it. It is not yet clear whether the new system will identify more people who may pose a risk, how often it might incorrectly flag someone, or how its performance will be compared with the existing process.
Those details matter because a classification can affect someone’s everyday life for years. A mistaken or outdated disciplinary record, for instance, could influence a score even if the person’s behavior has since changed. Staff review and the ability to challenge or correct information will be important parts of making the process fair.
There is also a practical limit to what better assessments can accomplish: prisons need suitable space and enough staff to act on them. Byron Osborn, who leads the union representing Michigan corrections officers, has raised concerns about a shortage of higher-security beds. If staff determine that someone needs closer supervision but no appropriate bed is available, the score alone cannot solve the problem.
Michigan says it plans to add 870 high-security beds through its Safe Prisons Initiative. More than 400 are expected to come from two new units at Oaks Correctional Facility and St. Louis Correctional Facility. The state estimates those construction projects will take about two and a half years, while some facility conversions are already underway.
The state has faced serious violence in its prisons, including recent cases in which incarcerated people were suspected of killing cellmates. A more responsive classification process could help staff notice changing risks sooner. But safer housing also depends on having the personnel and facilities to respond when a concern is identified.
Michigan’s pilot may offer a useful test of whether updated records and a clearer scoring process can improve decisions. The reported 30% gain sounds promising, but officials will need to explain how it was measured and show whether it holds up in practice. The results should also make clear how errors are handled and whether people can demonstrate that their circumstances or behavior have changed.
