Executive Summary
- Algorithmic Prison Sorting. Michigan will pilot an AI classification tool from tech nonprofit Recidiviz to sort inmates into housing levels based on predicted violence.
- Replacing Staff Committees. The software shifts housing decisions away from warden-appointed teams toward automated computer scores.
- Unverified Accuracy Claims. MDOC claims the tool will improve its ability to predict prison violence by 30 percent, but the state has not shared its validation data.
- High-Stakes Housing Impact. Security levels (from Level I to Level V) control cell conditions, safety, job eligibility, and visiting hours for prisoners.
- Lessons from Louisiana. Reports on similar tools, like Louisiana’s TIGER algorithm, show how automated scores trap low-risk people in harsh cells based on biased data.
Michigan is set to deploy a major new AI system across prisons statewide. The software is designed to classify incarcerated people based on how dangerous the state believes they are.
That news was first reported by Detroit channel Fox 2. The report revealed that the Michigan Department of Corrections, known as MDOC, is plowing ahead with an inmate categorization algorithm to sort people into various housing levels.
Table of Contents
- The Shift from Warden Committees to Predictive Algorithms
- MDOC December Pilot and the Claimed 30 Percent Gain
- The Flawed Data Problem: Lessons from Louisiana’s TIGER Algorithm
- Comparison Factsheet: Human Committees vs. Algorithmic Risk Tools
- Automation Cannot Substitute for Structural Reform
- Frequently Asked Questions
The Shift from Warden Committees to Predictive Algorithms
In Michigan, incarcerated people are separated into five housing levels according to their security classification. Level I facilities offer lower security with dormitory housing, while Level V facilities impose maximum security and restricted movement.
In the past, screenings were performed by a committee of staff appointed by the Warden at each facility. Now, it seems the state wants to outsource that decision, or at least have it be heavily informed by, a non-profit tech company called Recidiviz.
Recidiviz began as a data organization project at Google, meant to uncover how technology could standardize criminal justice data and improve the system nationwide. Its AI-integrated software platform is said to perform features like identifying individuals eligible for state-offered programs, reclassification, or release, and coordinate access to resources like transitional housing and employment.
The transition mirrors growing private interest in deploying autonomous algorithmic decision systems to manage operational workflows, but applying predictive models to human custody raises intense legal concerns.
MDOC December Pilot and the Claimed 30 Percent Gain
In a press release, MDOC said the Recidiviz software will be used to optimize placement of incarcerated individuals, starting with a pilot program in a number of men’s facilities beginning in December.
The state agency notes that the AI tool is expected to improve MDOC’s ability to more accurately predict incarcerated individuals’ risk of future violence by approximately 30 percent, though how the state arrived at that vague figure is unclear.
For incarcerated people, housing designations are extremely consequential. They determine which programming and resources an incarcerated person is eligible for, and what kind of population they will be housed with for the duration of their sentence.
Placement in a higher security tier restricts job eligibility, limits family visitation schedules, and increases exposure to disciplinary lockdowns. When an algorithm handles that sort, mistakes create immediate daily harm.
The Flawed Data Problem: Lessons from Louisiana’s TIGER Algorithm
Setting the state’s rationale aside, the MDOC-Recidiviz partnership leaves some major questions unanswered.
What kind of data will be collected, for example, will matter a great deal with regard to the AI system’s accuracy. As a ProPublica investigation into a similar risk-analysis program in Louisiana found, the kind of data being used to classify incarcerated people contained inherent racial and socioeconomic bias.
As a result, Louisiana’s carceral system has become even more cruel and harmful to those earnestly working toward rehabilitation. In one case reported by ProPublica, a blind and wheelchair bound 70-year-old was determined to be a moderate risk of reoffending despite having a clean disciplinary record, a classification which results in an automatic loss of eligibility for parole.
This disconnect highlights how static metrics, such as arrest age and historical policing density, punish individuals long after they have reformed. As highlighted in recent legislative debates over state-level accountability rules for AI risk, deploying algorithmic systems without external auditing creates unhedged hazards for the public.
Comparison Factsheet: Human Committees vs. Algorithmic Risk Tools
The factsheet below contrasts traditional staff-led classifications with algorithmic risk scoring platforms across correctional settings.
| Classification Area | Traditional Staff Review (MDOC PD 05.01.130) | Recidiviz Predictive AI Platform |
|---|---|---|
| Decision Maker | Committee of staff appointed by facility Warden | Algorithmic scoring models and automated data feeds |
| Core Inputs | Misconduct logs, program completion, and time served | Standardized historical records and predictive metrics |
| Review Cadence | Scheduled periodic reviews and milestone screenings | Dynamic, ongoing calculations across facilities |
| Promised Gain | Contextual human assessment and direct observation | Claimed 30% boost in predicting future violence |
| Key Failure Mode | Inconsistent reviews between individual prisons | Hardcoded racial bias and unverified statistical math |
Automation Cannot Substitute for Structural Reform
While Michigan’s system is geared toward housing assignments rather than parole decisions, it seems highly unlikely that software alone will solve the state’s ongoing corrections crisis. And if Louisiana’s embrace of AI is any example, it could very well make matters far, far worse.
Across Michigan facilities, officers face mandatory overtime, dangerous staffing shortages, and high turnover rates. An algorithm cannot hire guards, fix plumbing, or provide medical care.
Until state officials open Recidiviz’s algorithms to public scrutiny and independent validation, incarcerated individuals will face life-altering classifications dictated by automated code.
Explore Engineering Utilities and AI Solutions
Test data formatting, payload conversions, and document parsing scripts using our suite of Free Online Developer Tools.
Auditing algorithmic bias, building transparent data pipelines, or evaluating enterprise AI risk? Consult with TeckScaler’s enterprise IT specialists →
Frequently Asked Questions
What is the new AI tool being tested in Michigan prisons?
The Michigan Department of Corrections is partnering with tech non-profit Recidiviz to test an AI classification platform. The tool predicts an incarcerated person’s potential for violence to sort them into housing tiers.
When will the Michigan prison AI pilot program begin?
MDOC plans to launch the Recidiviz classification pilot in several men’s facilities starting in December.
How does Michigan classify prisoners today?
Michigan divides prisoners into five security levels (Level I to Level V). Under policy directive 05.01.130, a screening committee of facility staff appointed by the Warden conducts reviews based on misconduct and program milestones.
What are the main concerns with AI prison classification?
Data scientists and civil rights groups warn that predictive tools rely on arrest data reflecting systemic racial bias. In Louisiana, a similar algorithm trapped a disabled 70-year-old with a clean record in a moderate risk tier, blocking his parole.
