Michigan prisons will begin testing a new technology-assisted system on 1 December 2026 to predict prisoners' risk of future violence and help determine their security classifications. State officials expect the tool to improve prediction accuracy by approximately 30%.
But what happens if it wrongly identifies someone as dangerous?
The Michigan Department of Corrections (MDOC) announced the pilot on 28 October, saying it aims to improve safety for prisoners and staff. However, the department has not publicly explained how the system calculates risk or whether prisoners can challenge inaccurate information before it affects their classification.
New Technology Could Change Prisoners' Security Levels
The pilot will begin at selected men's correctional facilities before officials consider expanding it statewide. Michigan developed the system with Recidiviz, a non-profit technology organisation supporting criminal justice agencies in 19 states.
According to FOX 2 Detroit, the technology uses artificial intelligence to help assess prisoners' potential for violence. However, MDOC's official announcement does not explicitly describe the system as AI or explain its predictive methods.
The system will analyse historical information alongside more frequently updated records, allowing ongoing classification reviews. Michigan currently uses four general-population security levels: I, II, IV and V. Level I allows greater freedom of movement, while Level V involves the strictest supervision.
These classifications determine housing arrangements, daily activities and access to rehabilitation programmes. An inaccurate assessment could place someone under tighter restrictions than necessary.
Michigan Predicts 30% Improvement in Violence Assessments
The planned rollout comes amid concerns about safety inside American prisons.
According to figures from the Safe Inside Project cited by MDOC, assaults against prison staff nationwide increased by 77% between 2019 and 2024. Assaults between prisoners rose by 54% over the same period.
Corrections Director Heidi E. Washington said assigning prisoners to security levels that reflect their behaviour and needs could make facilities safer. She also expressed optimism about using historical and evidence-based information.
MDOC expects the technology to improve its ability to predict future violence by approximately 30%. However, its announcement does not explain how that improvement was calculated or provide supporting validation results.
The projection does not mean prison violence will fall by 30%. Nor does it establish how accurately the system will assess individual prisoners.
A 70-Year-Old Prisoner's Case Raises Questions
A case in Louisiana illustrates how computerised risk assessments can affect prisoners' lives.
In April 2025, ProPublica and Verite News reported on Calvin Alexander, a nearly blind, wheelchair-using prisoner who was then 70 years old.
Alexander had completed rehabilitation programmes and maintained a clean disciplinary record. He was preparing for a parole hearing when Louisiana's TIGER assessment system classified him as a moderate risk of reoffending.
Under state law, that classification made him ineligible to appear before the parole board. The assessment relied on historical factors without accounting for his subsequent rehabilitation efforts.
Michigan's proposed system serves a different purpose. It concerns security placements within prisons, not parole eligibility. There is no evidence that Michigan's pilot has produced comparable problems.
Nevertheless, Alexander's experience shows how risk assessments can have serious consequences when authorities rely on them to make decisions about prisoners.
Could Incorrect Predictions Limit Rehabilitation?
Michigan's security classifications influence more than housing and supervision. MDOC acknowledges that prisoners at lower security levels may have broader access to vocational training and higher education. Although some rehabilitation programmes operate across all general-population levels, other opportunities depend partly on classification.
An inaccurate risk assessment could therefore restrict access to activities intended to prepare prisoners for release.
Recidiviz has separately acknowledged the dangers of inaccurate AI-generated information. In a June 2026 article, the organisation explained that errors in correctional records could affect programme eligibility, housing decisions and personal liberty.
That discussion concerned AI applications in correctional settings generally. It did not establish that Michigan's classification system contains a defect.
Recidiviz also described phased testing and human review procedures for its AI products. Whether those specific safeguards apply to Michigan's pilot remains unclear.
What Happens When the System Gets Someone Wrong?
Michigan oversees approximately 32,000 incarcerated people statewide, although the initial pilot will operate only at selected men's facilities.
Officials say they will evaluate the system before considering wider implementation.
Under Michigan Administrative Code Rule 791.4401, institutional committees are responsible for most security classifications. Prisoners can also file grievances against reclassification decisions.
However, MDOC has not explained how much influence the new technology will have over those decisions or whether prisoners can review and correct the information it uses before their classification changes.
The technology could help officials identify genuine safety risks earlier. More frequent reviews might also recognise improvements in prisoners' behaviour, potentially allowing movement to less restrictive environments.
But even a system that performs better overall can make mistakes about individual people.
As Michigan prepares for its December pilot, its success will depend on more than improved predictions. An equally important question is whether prisoners will have a meaningful opportunity to challenge inaccurate information before it affects their lives.
