According to reporting from Detroit’s Fox2, the Michigan Department of Corrections (MDOC) is moving forward with a new software module to classify and sort incarcerated individuals into security levels based on their perceived potential for violence.
In Michigan, prisons categorize incarcerated individuals across housing security levels—from Level 1 (the lowest security and highest level of mobility/programming) to Level 5 (the most restrictive, reserved for those deemed to pose severe safety risks). Historically, these classifications were evaluated by warden-appointed committees using established administrative metrics. Now, the state is outsourcing—or at least heavily integrating—these housing assessments into a platform built by Recidiviz, a high-profile non-profit technology organization.
While MDOC claims this software transition will dramatically reduce prison violence, it exposes severe ethical landmines, questionable statistics, and the perilous trend of applying tech-solutionism to complex human systems.
The Promise: A 30% Boost in Violence Prediction?
In a press release detailing the pilot program—set to launch across select men's facilities starting December 1, 2026—MDOC boasted that the Recidiviz software will "optimize placement of incarcerated individuals". Most notably, state officials asserted that the tool is expected to "improve MDOC's ability to more accurately predict incarcerated individuals' risk of future violence by approximately 30 percent."
Where did MDOC get that 30% figure? Neither MDOC nor Recidiviz has made public the underlying statistical methodology, baseline metrics, or peer-reviewed validation supporting that claim.
Public agencies routinely throw around quantified accuracy metrics to justify expensive software deployments, yet without open auditing, a "30% improvement" remains a black-box PR statistic.
The Denial: "It's Not AI... Except When It Is"
When pressed on the integration of artificial intelligence into automated risk scoring, both MDOC and Recidiviz walked back the narrative.
"AI is never used in Recidiviz tools to make decisions about anyone's security level or release," a Recidiviz spokesperson stated.
MDOC Public Information Officer Jenni Riehle mirrored this defense:
"While Recidiviz builds AI tools that support staff with tasks such as meeting transcription, AI is never used to make decisions about an individual's security level or release."
However, Recidiviz's own marketing and technical documentation tell a far more blurred story:
AI-Powered Assessments: Marketing material for Recidiviz's "Intake" capabilities notes that officials can "administer AI-powered conversational assessments, letting residents respond at their own pace... while still capturing essential data."
Automated Risk Scoring & Reclassification: Recidiviz's "Dynamic Classification" product explicitly advertises that it "calculates risk scores based on form content" and triggers automated reclassification workflows.
Pervasive AI Integration: On its "How Recidiviz Uses AI" web page, the organization highlights leveraging AI for "matching people with jobs, services, or support based on their needs," as well as summarizing case notes and generating code.
Whether labeled as machine learning, algorithmic scoring, or dynamic rule-based software, the core reality remains unchanged: an automated data system is being leveraged to weigh risk profiles and determine where a human being sleeps, eats, and lives every day.
What’s at Stake: High Stakes, High Bias
In a carceral environment, housing classification isn't just about administrative convenience—it shapes every facet of an incarcerated person's daily life:
Safety & Environment: It determines whether someone is placed in a general population block with re-entry programming or locked down in high-security facilities surrounded by volatile populations.
Rehabilitation Access: Higher security tiers restrict access to educational opportunities, vocational training, work release, and behavioral therapy.
Parole & Sentence Duration: A higher security risk score frequently damages an individual’s eligibility for parole, early release, or community-based supervision.
The Cautionary Tale: Louisiana and COMPAS
We have seen this playbook before. Algorithmic risk tools rely on historical criminal justice data—data inherently steeped in systemic racial profiling, over-policing, and socioeconomic disparities. When historical biases feed an algorithm, the algorithm simply automates and institutionalizes those biases.
A prominent ProPublica investigation into risk-assessment software used in Louisiana laid bare the devastating consequences of algorithmic classification:
Risk models consistently misclassified Black individuals as high-risk at higher rates than white counterparts with similar offenses.
In one egregious case documented by ProPublica, a 70-year-old blind man in a wheelchair with a spotless disciplinary record was flagged as a "moderate risk" for future violence.
Because of that algorithmic tag, he automatically lost his eligibility for parole consideration.
While Michigan authorities insist their deployment focuses on housing assignments rather than direct parole determinations, the structural danger is identical. An unfair higher-security designation creates a self-fulfilling prophecy: higher supervision leads to more infractions, which leads to longer incarceration.
The Dangerous Illusion of Techno-Solutionism
Faced with severe staffing shortages, budget constraints, and rising facility violence rates, corrections departments across the country are eager to purchase digital silver bullets.
Software tools like Recidiviz are marketed as clean, data-driven solutions to deep systemic crises. But code cannot fix overcrowded facilities, underpaid corrections staff, or lack of rehabilitative infrastructure.
When state agencies replace human judgment and contextual evaluation with opaque risk formulas, they do not eliminate bias—they simply hide it behind a dashboard. Until state officials provide complete transparency, open-source auditing, and explicit avenues for incarcerated individuals to challenge automated risk scores, Michigan’s new pilot project threatens to amplify the very harms it purports to fix.
The story of Michigan’s inmate classification algorithm highlights the core challenges driving the current boom in artificial intelligence. Companies, government agencies, and tech vendors are rushing to deploy automated decision-making systems. However, many of these systems suffer from biased training data, opaque black-box logic, and misleading marketing claims.
This gap—between hype and real, ethical, production-ready implementation—is where AIVASU Tech Institute plays a vital role.
How AIVASU Courses Help Engineers Ride the AI & Tech Boom
AIVASU offers hands-on programs spanning AI Engineering, Generative AI, Autonomous Agents, Data Science, AI-Integrated Full-Stack, and MLOps. These courses empower developers and career switchers to navigate real-world challenges effectively:
1. Eliminating Bias with Responsible AI & Data Science
The Industry Problem: Risk-assessment algorithms often recycle historical biases, misclassifying individuals because their models rely on skewed datasets.
How AIVASU Prepares You: Courses in Data Science and AI Engineering teach proper data preprocessing, dataset auditing, model fairness, and mitigation of algorithmic bias. You learn to engineer systems that deliver equitable, mathematically sound predictions.
2. Mastering MLOps & Explainable AI (XAI)
The Industry Problem: Closed-source, "black box" models face severe backlash when users cannot understand or challenge how a decision was calculated.
How AIVASU Prepares You: The AI/ML Cloud Engineering and MLOps tracks teach transparent system design. Developers build explainable AI pipelines, deploy model monitoring tools, and conduct continuous audits to verify system behavior in production.
3. Moving Beyond Hype to Autonomous Agents & Production LLMs
The Industry Problem: Organisations struggle to distinguish between superficial AI marketing buzz and functional, production-ready automation.
How AIVASU Prepares You: Specialized tracks like Agentic AI (LangChain, LangGraph, CrewAI) and Generative AI train developers to build autonomous agents capable of structured reasoning, document retrieval (RAG), and verified execution.
4. Shipping Enterprise-Grade, Full-Stack AI Products
The Industry Problem: Building a standalone model is only half the battle; integrating it securely into existing workflows requires full-stack expertise.
How AIVASU Prepares You: AIVASU’s AI-Integrated Full Stack and Cloud DevOps programs show students how to connect modern frontend/backend frameworks (Next.js, FastAPI, Django) with AI backends, vector databases, and containerized microservices.
5. Career Mentorship & Direct Industry Referrals
The Industry Problem: High demand for skilled AI engineers creates intense competition for roles that demand real-world experience rather than theoretical knowledge.
How AIVASU Prepares You: Students build multi-tier enterprise capstones for their portfolios, receive 1-on-1 code reviews from working tech leads, and gain access to placement sprints with partner companies across the tech ecosystem.
By training developers to build transparent, unbiased, and production-ready systems, AIVASU ensures that software engineers do not just participate in the AI boom—they lead it responsibly.