Michigan Prison AI Will Sort Inmates by Violence Risk in December Pilot

Michigan is about to let an artificial intelligence system help decide how dangerous its prisoners are. From December 1, 2026, the Michigan Department of Corrections will pilot a new classification tool, built with the nonprofit technology organisation Recidiviz, at selected men's facilities, sorting incarcerated people into security levels based on their predicted risk of future violence. The department says the Michigan prison AI system should predict that risk about 30% more accurately than the current process, at a time when assaults inside American prisons have surged.

Key Points

• Michigan's new prisoner classification pilot begins December 1, 2026, at selected men's facilities, with wider use to be considered only after the pilot is tested.
• The system, developed with Recidiviz, replaces scheduled, committee-based assessments with reviews that use real-time data and can be updated far more often.
• The department expects the tool to improve predictions of future violence by approximately 30%, a figure whose underlying methodology has not been published.
• Michigan supervises about 32,000 incarcerated people and uses four general-population security levels: Level I, Level II, Level IV and Level V.
• A person's security level shapes daily movement, supervision and access to programmes, so the classification decision carries real consequences for daily life inside.

What Happened

The Michigan Department of Corrections announced the redesigned classification system on October 29, 2026, framing it as a safety measure rather than a punishment measure. The pilot will test how well the new tool performs before the department considers using it more widely, and it sits inside the department's Safe Prison Initiative, where prisoner classification and bed space are described as a key area for safety success.

Under the new system, incarcerated people will be placed at the security level that fits their current behaviour and needs, using evidence-based and historical data that can be analysed quickly. The crucial change from the old way of working is timing. At present, classification inputs are collected on a scheduled basis and assessments are carried out by a committee of staff appointed by the warden at each facility, working from manual worksheets. The new system allows ongoing classification reviews incorporating real-time data, so a person's risk profile can be reassessed as new information arrives rather than waiting for the next scheduled review.

Director Heidi E. Washington put the department's case in the announcement itself: "Classifying prisoners at the security level that fits their current behaviors and needs keeps everyone safer — the individual, other incarcerated people, and correctional staff. I am optimistic that this pilot's use of evidence-based and historical data has created a tool that is dynamic, objective, and will strengthen the security of our facilities."

The pilot is deliberately limited. It begins at selected men's facilities on December 1, and the department has not named the full list of pilot sites in the announcement. Women's facilities are not part of this first phase, and the department says the pilot exists precisely to test performance before any wider rollout is considered.

Background and Context

Michigan's prison population is large, and its violence problem is not new. The department supervises about 32,000 incarcerated people statewide. From 2019 to 2024, prisons across the United States saw a 77% increase in assaults on staff and a 54% increase in assaults between incarcerated people, according to the Safe Inside Project figures cited by the department. Michigan's own experience has included at least four recent incidents in which inmates are suspected of killing their cellmates, according to the Detroit Free Press, which also reported that assaults have risen even as the prison population has fallen.

Classification is the machinery that sits behind those numbers. A security level is not an abstract label: it determines daily movement, the amount of supervision a person is under, and the activities they can access each day. At higher security levels, movement is more limited, in order to reduce violence and manage potential negative behaviour. Programming such as Violence Prevention Programming is available at all general-population security levels, but people at lower levels may have access to a wider range of opportunities, such as vocational training or higher education, because of their classification. Getting that call right matters for safety, and getting it wrong can restrict a person's daily life and prospects for years.

The department has been building towards this change for some time. Since January 2026, it has announced the addition of 870 high-security beds, increases to barrier-free accessible rooms, and Protection Unit changes, all presented as part of the same safety push. The classification pilot is the data-driven half of that programme: better beds and better information, in the department's telling, should together make facilities safer for staff and for the people held in them.

Recidiviz, the partner on the project, is a 501(c)(3) nonprofit technology organisation that supports criminal justice agencies in 19 states. The department engaged it last year to help develop the new system. Recidiviz began as a data organisation project inside Google, aimed at standardising criminal justice data and improving outcomes across the system. Its platform is used for tasks including identifying people eligible for state-offered programmes, reclassification and release planning, and coordinating access to resources such as transitional housing and employment. The Michigan Department of Corrections and the Department of Technology, Management and Budget signed a $1 million contract with Recidiviz in 2021 to build a criminal justice data platform, state records show, and that contract has since been extended three times and is now valued at just over $4 million, according to the Free Press. The classification module is the latest piece of that larger platform, which also includes restrictive housing reviews and prison transfers.

Michigan Prison AI: How the New Classification Works

The headline claim is the 30% figure. The department says the redesigned system "is expected to improve MDOC's ability to more accurately predict incarcerated individuals' risk of future violence by approximately 30%." That is an expectation, not a result: the pilot has not yet run, and the department has not published the methodology behind the figure, a gap noted in reporting on the announcement. Readers should therefore treat 30% as the department's target for the pilot to prove, rather than a measured outcome.

What is clearer is how the tool is meant to work in practice. It draws on numerous factors, including the nature of a person's crimes, assessments by department experts, and how they have behaved in prison, and it produces a numerical assessment that feeds into the security-level decision. The department stresses that the system is designed to support staff, not to replace them. A department spokesperson, Jenni Riehle, told the Free Press that artificial intelligence is never used to make decisions about an individual's security level or release; instead, it calculates an assessment score, which staff then use. That distinction matters, because it is the difference between an algorithm deciding where a person lives and an algorithm informing a human decision about where a person lives.

The levels themselves will be familiar to anyone who knows Michigan's system, with one quirk that often confuses outsiders. Michigan uses four general-population security levels: Level I, Level II, Level IV and Level V. There is no Level III designation. Level I is for people deemed the safest, and Level V is the highest security level, where movement is most restricted. The pilot does not change those levels; it changes how people are assigned to them, and how often that assignment can be revisited.

The real-time element is the most consequential technical change. A scheduled review can leave a person in a level that no longer fits them for months, whether they have become safer or more dangerous in the meantime. A system that updates as new data arrives can, in principle, move someone to a lower level sooner when their behaviour improves, opening up programmes and movement, or to a higher level sooner when risk rises, protecting staff and other incarcerated people. Whether it does either of those things fairly depends entirely on the quality of the data it is fed, which is where the debate begins.

Reactions and Quotes

The department's own message has been consistently about safety. Director Washington's statement frames classification as a safety tool first: the right level for the right person at the right time keeps the individual, other incarcerated people and correctional staff safer. That framing is echoed in the Safe Prison Initiative language, which presents classification and bed space as safety infrastructure rather than as administrative paperwork.

Reporting on the announcement has been more sceptical, and the scepticism centres on transparency. Futurism, which first drew wide attention to the plan, noted that how the state arrived at its approximate 30% figure is unclear, and that the department was contacted for comment. The same reporting points to the precedent that worries civil-liberties advocates: a ProPublica investigation into a similar risk-analysis system in Louisiana found that the underlying data contained inherent racial and socioeconomic bias. If the data fed into a risk tool reflects biased policing, sentencing or prison discipline patterns, the tool can reproduce those patterns at scale, however neutral its mathematics appear.

The department's answer, in effect, is scope. Michigan's system is geared towards housing assignments rather than parole decisions, and the department says the AI calculates a score rather than making the decision itself. Supporters of that distinction argue that housing is a safety decision that must be made somehow, and that a data-driven score, regularly updated, is fairer than a committee working from paper worksheets on a fixed schedule. Critics reply that housing is not a minor decision: it determines who a person lives with, what programmes they can reach, and how much of each day they spend locked in, and a scoring error therefore compounds over a sentence.

No independent evaluation of the Michigan tool has yet been published, because the pilot has not started. That absence is itself a fact worth stating plainly: at this stage, the public has the department's announcement, its 30% expectation, and the experience of similar systems elsewhere, but not yet any Michigan results to judge the tool by.

Why It Matters

This story matters beyond Michigan for three reasons. First, scale: Michigan holds about 32,000 people, and a classification system that works there will be watched by corrections departments in other states, several of which already use Recidiviz tools. A successful pilot could become a template; a failed one, with documented bias or error, could become a cautionary tale.

Second, it tests a boundary that runs through almost every public-sector use of artificial intelligence: the line between a tool that informs a human decision and a tool that effectively makes it. Michigan insists its system stays on the informing side of that line, with staff retaining the decision. How that works in practice — how often staff overrule the score, in which direction, and with what reasons recorded — will determine whether the distinction is real or cosmetic. Those override rates are the numbers to watch once the pilot is running.

Third, it lands in a year when artificial intelligence in public institutions is under intense scrutiny, from classrooms to courtrooms. The same week Michigan's pilot was announced, senators were pressing Google about AI in schools, and AI systems were at the centre of election-integrity disputes. Prisons are a less visible setting than schools or elections, but the stakes for the people affected are at least as high, and the oversight is thinner, because incarcerated people have far less ability to challenge a decision about them than a student, a voter or a job applicant.

What Happens Next

The pilot begins December 1, 2026, at selected men's facilities. The department says it will test how well the new tool performs before considering wider use, which means the first months should produce the evidence that the announcement currently lacks: how accurately the tool predicts violence, how often its scores are overruled, and whether its errors fall unevenly on any group of incarcerated people.

What to watch for in that period is specific. First, whether the department publishes pilot results, including the methodology behind the 30% claim and any validation of the tool against real outcomes. Second, whether it discloses which data fields the tool uses, and whether any of them act as proxies for race or socioeconomic background. Third, whether people classified at a higher level under the new system have a clear, timely route to challenge their classification, and whether reclassification in both directions actually happens as new data arrives.

If the pilot works as the department hopes, the next step would be wider use across Michigan's facilities, and a model other states could adopt. If it does not, the department will have to explain why a system built to be dynamic and objective produced results it cannot defend. Either way, December 1 is the date when Michigan's prison AI stops being an announcement and starts being a record — and that record, unlike a press release, can be checked.

Frequently Asked Questions

When does Michigan's prison AI pilot start? December 1, 2026, at selected men's facilities. The department will test the tool's performance before considering wider use.

What does the system actually do? It helps classify incarcerated people into security levels — Level I, Level II, Level IV and Level V — using evidence-based and historical data, updated in real time rather than only on a scheduled basis.

Does the AI decide where a prisoner lives? The department says no. It says the AI calculates an assessment score, and that artificial intelligence is never used to make decisions about an individual's security level or release. Staff retain the decision.

Why is Michigan changing its system now? The department points to rising prison violence: from 2019 to 2024, US prisons saw a 77% increase in assaults on staff and a 54% increase in assaults between incarcerated people, according to Safe Inside Project figures it cited.

What are the main concerns? Transparency and bias. The methodology behind the 30% improvement claim has not been published, and a ProPublica investigation found inherent racial and socioeconomic bias in the data behind a similar Louisiana system. Michigan's system affects housing and programme access, not parole.

Related coverage: Google classroom AI inquiry: senators press Pichai for answers by Nov. 2 and Claude Georgia vote-data scandal: AI bypassed paywall to reach election counts.

Sources: Michigan Department of Corrections press release, October 29, 2026; Detroit Free Press, October 29, 2026; Fox 2 Detroit, October 2026; Futurism, November 1, 2026; NewsBytes, November 2, 2026. The featured image is an illustrative photograph of a prison housing corridor, not a photograph of a Michigan pilot facility.