From the Public Safety edition of October 28, 2026
Michigan Department of Corrections will pilot a Recidiviz-built, data-driven prisoner classification tool on Dec. 1, claiming about 30% better prediction of violence risk
Michigan Department of Corrections / Recidiviz · Corrections / risk classification and placement · Michigan, United States
- Publisher
- Michigan Department of Corrections
- Original publication
- Michigan Department of Corrections press release, Oct. 28, 2026
- Source retrieved
- 2026-10-29
- Event date
- 2026-12-01
What happened
The Michigan Department of Corrections (MDOC) announced on Oct. 28, 2026 that it will pilot a redesigned prisoner classification system at selected men’s facilities from Dec. 1, 2026, developed with Recidiviz, a nonprofit technology organization that works with criminal justice agencies in 19 states. MDOC says the system is expected to improve its ability to predict incarcerated individuals’ risk of future violence by approximately 30% and will allow ongoing classification reviews using real-time data rather than inputs collected on a schedule. Director Heidi Washington described it as using evidence-based and historical data to create a tool that is “dynamic, objective”. MDOC supervises about 32,000 people across four general-population security levels; classification determines movement, supervision and access to programming such as vocational training and higher education. The department says the pilot will test how well the tool performs before wider use.
Why it matters
Classification scores decide daily liberty and program access for thousands of people; a claimed predictive gain announced without method, baseline or fairness testing is exactly the kind of corrections algorithm buyers and oversight bodies need to probe.
Evidence and measured results
curl returned HTTP 403; retrieved the press release with the WebFetch tool and read it in full. No technical report, validation study or fairness analysis was linked or found.
Limitations and uncertainty
Agency press release; the 30% figure is a projection with no stated baseline, metric or validation method; no fairness or disparate-impact testing is described; the tool is not described as AI, so it is classified here as an algorithmic risk instrument.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-10-29; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- A corrections department must show that a new data-driven classification tool improves violence prediction fairly before scores shape placement for 32,000 people.
- Stakeholders
- MDOC director and classification office
- Recidiviz
- Facility wardens and classification staff
- Legislative oversight
- Advocates for incarcerated people
- Discovery questions
- What does the 30% improvement measure, and against which baseline?
- How do error rates differ by race, age and facility?
- Can staff override, and are overrides logged?
- How can a person contest a classification?
- Value hypothesis
- Interpretation: an independent validation plan could make the pilot’s results credible to oversight bodies.
- Potential engagement
- A pilot evaluation design covering validation metrics, fairness analysis and override monitoring.
- Unsupported claims
- The 30% figure as an independently measured result
- That the tool is unbiased
- That it will reduce assaults
Pre-sales engineering
Role takeaway
- Fit
- Relevant to state corrections agencies revising classification instruments.
- Architecture and integration
- Scoring tool fed by real-time incident, behaviour and programming data; outputs drive security-level placement.
- Prerequisites
- Data-quality checks on each feed
- Defined metrics and baseline
- Override workflow
- Constraints
- Pilot at selected men’s facilities only
- No published technical documentation
- Projection, not observed result
- Security
- Protect incarcerated people’s data, log score changes and overrides, and restrict model changes to governed releases.
- Proof of value
- Back-test on historical cohorts
- Compare error rates by group
- Audit a sample of overrides
Delivery
Role takeaway
- Work
- Run an evaluable pilot with fairness analysis and a contest route.
- Dependencies
- Recidiviz model documentation
- Data feeds
- Evaluation partner
- Ownership
- MDOC classification director’s office
- Skills and adoption
- Training classification staff to interpret scores and record override reasons.
- Governance checkpoints
- Pre-registered evaluation metrics
- Mid-pilot fairness review
- Expansion decision gate
- Acceptance criteria
- Published validation with group error rates before expansion (proposed)
- Override reasons logged for 100% of overrides (proposed)
- Contest route available to every pilot participant (proposed)
- Risks
- Data errors raising security levels
- Opaque scoring
- Expansion before evaluation
Implementation considerations
Lighthouse Advisory interpretation across the operating dimensions a public-sector buyer must settle before this evidence becomes a design. Each note answers the question under its heading for this specific source.
Architecture and integration
What must connect, and where does the AI sit in the workflow?
Shifts classification from scheduled inputs to continuous review on real-time data, implying data pipelines from incident, behaviour and program systems into the scoring tool.
Governance
Who approves, reviews and stays accountable for outcomes?
Pilot before wider use is stated, but evaluation criteria, independent validation and review rights for incarcerated people are not described.
Security and privacy
What data, permissions and controls need testing?
Real-time behavioural data feeding classification raises data-quality and due-process questions; the release does not describe data sources.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
Classification affects access to programming and barrier-free housing; staff will need to understand and override scores.
Procurement
What should contracts, pricing and exit terms secure?
Developed with a nonprofit partner rather than a commercial vendor; ownership of the model and data is not stated.
Operating model
Which teams own the service once it runs?
Continuous re-classification could change movement and programming decisions more often than scheduled reviews.
Publication history
- 2026-10-28Public Safety · Issue 527 resources
Stable resource ID: michigan-doc-recidiviz-data-driven-classification-pilot-violence-risk-20261028