Michigan is testing a new prison classification system that could change how security levels are set, how often they are reviewed, and how quickly staff respond to shifting risks. State officials say the approach may improve predictions of prison violence, but the biggest surprise is that it is not really artificial intelligence making the call. It is a rules-based scoring system that still leaves the final judgment in human hands.
The stakes are obvious. In a prison setting, a security level affects where someone lives, how closely they are monitored, and whether they can access education or job training. Michigan’s Department of Corrections oversees about 32,000 incarcerated people, and the difference between a lower-security placement and a tightly controlled unit can shape daily life in a major way.
Michigan currently uses four general population security levels, ranging from Level I to Level V. Lower levels usually mean more movement and more opportunities, while higher levels bring tighter supervision and fewer freedoms. Programs like violence prevention remain available across those settings, but officials want a system that better matches current behavior instead of relying too heavily on old snapshots of a person’s record.
The new system is set to begin testing on Dec. 1 at selected men’s prisons. If the pilot goes well, the state could expand it to other facilities later. One of the main changes is that classifications can be reviewed more continuously, using updated prison records instead of waiting for a fixed schedule to roll around.
Recidiviz, the nonprofit behind the software, worked with Michigan corrections staff after reviewing a decade of classification data. The formula weighs factors chosen by state experts, and recent disciplinary problems can carry more weight than stale incidents from years ago. Good behavior over time may help someone move to a less restrictive setting, while fresh concerns can trigger another look.
That part matters because prison records are not just paperwork. They are the backbone of decisions that shape housing, supervision, and access to programs. If a report is inaccurate, or if a person has genuinely changed, the system needs enough flexibility to catch that before a bad score locks in the wrong outcome.
There is also a big clarification worth keeping straight. Recidiviz says Michigan’s classification tool does not use AI to determine security levels. The software follows a points-based formula, the kind of thing many people would call an algorithm, and prison staff can review the calculations and override them when needed.
That distinction matters because AI has become a buzzword that gets attached to almost anything with a computer inside it. Here, the process is more old-school than the headlines suggest. It is still software, still data-driven, and still capable of mistakes if the underlying information is incomplete or wrong.
Michigan says the redesign could improve predictions of future violence by about 30%, but the number comes with plenty of questions. The announcement does not explain how that figure was reached, what benchmark was used, or how often the system could miss a person who truly needs closer supervision. Those details matter when a score can influence someone’s housing and safety.
There is another practical problem that no algorithm can solve by itself. Even if the system identifies someone as needing a higher-security placement, the prison still needs a bed available in the right unit. Michigan has acknowledged a shortage of higher-security space and says it plans to add 870 beds as part of its Safe Prisons Initiative, but some of those projects will take years to finish.
That leaves corrections leaders juggling safety, staffing, and space at the same time. A better scoring system might help identify problems sooner, but it does not build new housing on its own or put more officers on the line overnight. In a system already dealing with violence, overcrowding, and resource limits, the real test is whether the new process can make decisions that are both faster and fairer without pretending technology can do everything.
Michigan’s pilot will show whether a more responsive classification model can actually improve life inside the state’s prisons. If the system helps staff spot changes sooner, rewards real progress, and keeps people in the right setting, it could become a meaningful tool. But the pressure will be on to prove that the scores make sense in the real world, not just on a spreadsheet.
