What happened
FOX 2 Detroit reports that Michigan’s Department of Corrections plans to begin a pilot on December 1, 2026, in some men’s prisons. The new classification module uses a tool developed by Recidiviz, a criminal-justice nonprofit. According to the report, Recidiviz is used in 19 states. The department says the pilot will improve its predictions of future violence by 30%. The article does not provide the underlying evaluation or define the comparison behind that figure.
The system is intended to make determinations in real time, rather than assess an incarcerated person’s potential for violence on a scheduled basis. It flags people for classification, after which they receive a security-level designation. Department Director Heidi E. Washington said the pilot uses evidence-based and historical data and described it as a dynamic, objective tool intended to strengthen facility security. Those are the department’s stated aims; the source does not establish how the tool reaches its determinations or what data it weighs.
The report describes security levels 1, 2, 4, and 5, with no level 3 designation listed. Level 1 is described as the safest and permits the most movement and activities, with a focus on reentry preparation, education, and skill-building. Level 2 provides a structured environment and many programs, emphasizing education, evidence-based programming, and positive routines. Level 4 brings increased supervision and controlled movement, with programming focused on behavior, substance-use interventions, and high school equivalency completion. Level 5 is the most restricted, with close supervision and evidence-based programming focused on violence prevention and behavioral therapy.
The article frames the pilot as a response to increased prison violence, citing the Safe Inside Project’s reported 77% increase in assaults on staff by inmates and a 54% increase in assaults between inmates in U.S. prisons. It gives no start date beyond the planned December 1 launch, no price, and no details on the pilot’s duration, participating facilities, appeal process, or eventual expansion criteria.
What the archive adds
Community safety report and new protections for teens covered OpenAI’s expansion of ChatGPT violence-prevention safeguards. That is a related public-safety use of automated systems, but not the same event or a prior account of prison classification. Michigan’s pilot is a new fact in the broader question of how institutions use software to assess violence risk.
Trend signal
The report offers a limited example of consequential decisions mediated by software, but no SITE_PATTERNS entry genuinely fits prison risk classification or establishes a broader trend here. The 30% improvement is a state-reported expectation, not evidence in this source that the tool improves outcomes in practice.
Why now
The department’s stated rationale, as reported by FOX 2, is to strengthen safety amid rising assaults in U.S. prisons. The pilot is scheduled to start December 1, 2026; the source does not give a more specific explanation for that timing.
Why it matters
This classification is not merely a prediction displayed to staff: the report says security level governs movement, supervision, and the kinds of programs available. A system that changes those assignments in real time could help institutions respond to changing behavior, but errors could also mean unnecessarily restrictive conditions or reduced access to education and reentry preparation. Teams building decision-support systems for high-stakes settings should treat the classification process, its evidence, and the consequences of an assigned category as part of the product—not assume that an “objective” score is self-validating. The source provides no independent validation or details about human review, so it cannot establish whether the anticipated accuracy gain translates into safer or fairer decisions.
Limits and uncertainty
The 30% figure is attributed to the Michigan Department of Corrections, and the report supplies no baseline, evaluation design, sample size, or definition of improved prediction. It does not say which historical or current data the software uses, how often classifications update, or whether staff can override recommendations. The levels listed omit level 3 without explaining whether it does not exist or is simply not described. The source also gives no information about error rates across groups, notice to incarcerated people, appeal rights, safeguards against stale or inaccurate records, or the pilot’s duration. It therefore establishes the planned deployment and intended classification effects, not the tool’s accuracy or consequences.
Reader takeaway
Michigan plans to use Recidiviz software for real-time prison security classifications, and says the pilot will improve violence prediction by 30%. Because those classifications affect movement and programming, the consequential question is not only whether the model predicts risk, but how its estimates are validated, reviewed, and contested. The source does not answer those questions.
Open question
How will Michigan measure the claimed 30% improvement, and what review or appeal process will apply when a classification changes someone's security level?