Michigan Hands Prison Risk Scores to an Algorithm

Threat assessment: high stakes, low transparency.

Michigan is about to let AI help decide how dangerous incarcerated people are. Futurism AI reports that the Michigan Department of Corrections (MDOC) will roll out software from the nonprofit Recidiviz to sort people into housing levels. A pilot starts in December at several men’s facilities. The state says the tool will “optimize placement of incarcerated individuals.”

This isn’t a chatbot pilot or a back-office efficiency test. It’s an algorithm with real influence over people’s daily lives, and the details released so far are thin.

🎯 Situation Report

Here’s what’s on the table:

1. The system. Michigan sorts incarcerated people into five housing levels based on security classification. Until now, a committee of staff appointed by each facility’s warden made those calls.
2. The vendor. Recidiviz started as a data project at Google, with a mission to “uncover how technology could standardize criminal justice data and improve the system nationwide.” Its platform flags people eligible for programs, reclassification or release, and it coordinates access to transitional housing and jobs.
3. The claim. MDOC says the tool will “improve MDOC’s ability to more accurately predict incarcerated individuals’ risk of future violence by approximately 30 percent.” The state hasn’t said how it got that number.
4. The timeline. The pilot starts in December, limited to men’s facilities at first.

⚠️ Why Housing Level Matters

This might sound like a logistics decision. It’s not.

Housing level decides which programs and resources a person can access. It also decides who they live with, sometimes for their whole sentence. One classification can open the door to rehabilitation or close it.

What stands out to me is the shift in who makes the call. A staff committee can be questioned, pushed back on and held accountable. Software that “heavily informs” the decision is much harder to challenge, especially when the people affected can’t see the inputs.

🔍 The Louisiana Warning

We’ve already seen how this can go. Futurism AI points to a ProPublica investigation into a similar risk tool in Louisiana. It found the data used to classify incarcerated people carried built-in racial and socioeconomic bias.

One case says it all. A blind, wheelchair-bound 70-year-old with a clean disciplinary record got labeled a “moderate risk” of reoffending. That label automatically cost him his eligibility for parole.

Michigan’s system targets housing, not parole. But the core problem carries over. A risk model is only as fair as the data it learns from. And criminal justice data reflects decades of uneven policing, sentencing and enforcement.

This isn’t new territory. Risk-scoring tools have shaped bail, sentencing and parole across the US for years, and they’ve faced steady criticism over bias and opacity the whole time. Michigan is adding one more decision point to that list.

📋 Open Questions

MDOC hasn’t answered the questions that matter most:

1. What data goes in? Criminal history, disciplinary records, demographics, zip codes? Each one brings its own risk of bias.
2. Where does the 30% come from? Better than what baseline, and measured how?
3. Who has the final word? Does staff judgment override the software, or just rubber-stamp it?
4. Can people appeal? Will incarcerated people see their scores and be able to contest them?
5. Who audits it? Is there independent review of outcomes across racial and economic groups?

Futurism AI says it reached out to MDOC for comment.

🧭 What It Means for AI Practitioners

If you build or deploy AI, treat this as a case study in high-risk use cases. Prisons, hiring, lending and healthcare all share the same pattern: a model’s output turns into a life-altering decision for someone with little power to push back.

A few takeaways:

1. Accuracy claims need receipts. A vague percentage without a method behind it won’t earn trust, and it shouldn’t.
2. Training data is policy. Historical data bakes historical decisions into future ones.
3. Human oversight has to be real. “Human in the loop” means nothing if people just defer to the score.

🔭 What to Watch

The December pilot is the next checkpoint. Watch whether MDOC publishes its methodology, sets up independent audits or explains how people can contest a classification. If none of that happens, expect civil rights groups and researchers to start asking loudly.

Software alone won’t fix Michigan’s corrections problems. Done carelessly, it could deepen them.

You can find the full story at Futurism AI.