{"resourceId":"recidiviz-clarifies-michigan-classification-not-ai-generative-ai-safeguards-corrections-20261102","versions":[{"version":"external-f35fed114c00014881201023b371f5acf19d7e1eb4c194c2b451f935d9dd9d60","resource":{"id":"recidiviz-clarifies-michigan-classification-not-ai-generative-ai-safeguards-corrections-20261102","title":"Recidiviz says Michigan’s new prison classification tool is points-based, not AI, and sets out its review and error-ceiling rules for generative AI case notes","organization":"Recidiviz (nonprofit) with Michigan Department of Corrections","sector":"Corrections / classification instruments and generative AI case notes","geography":"Michigan and other U.S. states","publishedAt":"Recidiviz blog, Nov. 2, 2026","publicationDate":"2026-11-02","eventDate":null,"sourceName":"Recidiviz","sourceLabel":"Why We Don’t Use AI to Classify People in Prisons","sourceUrl":"https://www.recidiviz.org/updates/why-we-dont-use-ai-to-classify-people-in-prisons","evidenceClass":"vendor-claim","outcomeClass":"emerging","topics":["governance-procurement","knowledge-work","operating-model"],"finding":"Recidiviz, the nonprofit working with the Michigan Department of Corrections on its classification pilot due Dec. 1, wrote on Nov. 2, 2026 that coverage describing the tool as AI was wrong: both the old and new instruments are rules-based points systems staff can read line by line. It says the redesign analyzed 10 years of Michigan classification data and that MDOC staff experts decided which factors to keep. Recidiviz says it does not use AI to decide custody, parole or release, or to score risk, and that AI only drafts records that staff review and submit. For its generative AI case-note tool, it says staff manually review 100% of outputs in early testing, a second model grades every output later, harmful errors have a strict ceiling, and tools are rolled back if the ceiling is crossed.","sledRelevance":"A provider correcting AI framing of a corrections risk tool, and publishing its own generative AI control rules, which buyers can turn into contract terms.","evidence":"Opened and read the full post. All statements are the provider’s own; no error rates, ceiling values or validation results are disclosed. MDOC’s Oct. 28 release was archived previously and is not reused.","architectureImplications":"Transparent points-based instrument for classification; generative AI limited to draft case notes with a second-model grader.","governanceImplications":"Separates decision tools from drafting tools; human authorship of records; rollback on harmful-error ceiling.","securityPrivacyImplications":"Case notes contain sensitive supervision data; the post does not detail data handling.","caveats":"Provider self-description; the 30% improvement claimed by MDOC remains unexplained; no independent validation.","streamIds":["public-safety"],"roles":{"sales":"Sales framing: Recidiviz shows that a corrections risk tool can be transparent points arithmetic, and that media labels of AI may be wrong. The buyer problem is explaining to legislators and incarcerated people what a tool really does. A bounded engagement could document each tool’s method and decision role. Do not claim the Michigan tool is validated; the 30% figure and error ceilings are not published.","engineering":"Engineering framing: two different systems sit here, a points-based classification instrument and a generative AI note drafter with a second model grading outputs. Pre-sales should ask how the instrument was validated on local data and how drafting errors are categorized and capped. A proof of value would sample drafts against transcripts and report harmful-error rates. Constraints: data quality across 10 years of records and model change control.","delivery":"Delivery framing: the work is keeping decision tools transparent and AI drafting under strict review. Ownership sits with the corrections classification director and the case-management program lead. Acceptance criteria are proposals: published instrument factors, validation by subgroup, a harmful-error ceiling with reported rates, and rollback on breach. Main risks: rubber-stamp review of drafts and unexplained predictive claims."},"roleDetails":{"sales":{"customerProblem":"Corrections agencies face confusion and distrust about whether tools use AI to decide custody or release.","stakeholders":["Corrections classification director","Case-management lead","Legislative oversight","Advocates for incarcerated people"],"discoveryQuestions":["Is this tool AI, rules-based, or statistical?","How was the instrument validated on our data?","What is the harmful-error ceiling for AI drafts?"],"valueHypothesis":"Clear method disclosure and output evaluation build trust in corrections tools.","potentialEngagement":"Tool method documentation and evaluation design for corrections classification and AI drafting.","unsupportedClaims":["Michigan 30% violence prediction improvement","Error rates not disclosed","Independent validation"]},"engineering":{"fit":"Relevant to corrections agencies redesigning classification or adopting AI case notes.","architectureIntegration":"Points-based instrument in the offender management system; generative AI drafting from meeting transcripts with a grader model.","prerequisites":["Historical classification data","Transcripts of supervision meetings","Error taxonomy"],"constraints":["Data quality","Model change control","State law on records authorship"],"security":"Restrict drafting tool data to the case, keep officer authorship, log every output and grade.","proofOfValue":["Subgroup validation of the instrument","Draft accuracy sample against transcripts","Rollback drill"]},"delivery":{"work":"Document instrument factors, validate, and run AI drafting with full output evaluation.","dependencies":["Provider evaluation data","Staff working group","Oversight review"],"ownership":"Corrections classification director and case-management program lead.","skillsAdoption":"Officers need training to edit drafts substantively; staff need to explain scores line by line.","governanceCheckpoints":["Instrument factors published","Validation reviewed","Harmful-error ceiling set"],"acceptanceCriteria":["Subgroup validation reported","Harmful-error rate below ceiling","Rollback completed on breach"],"risks":["Rubber-stamp review","Unexplained predictive claims","Bias in historical data"]}},"retrievedAt":"2026-11-03T03:54:35Z","enrichedAt":"2026-11-03T03:57:16Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Officers remain legal authors and must spend real time editing drafts.","procurementImplications":"Buyers can require disclosure of whether a tool is AI, evaluation of every output, and rollback terms.","operatingModelImplications":"Every model or prompt change treated as a new launch with re-evaluation.","sourceVerification":{"openedUrl":"https://www.recidiviz.org/updates/why-we-dont-use-ai-to-classify-people-in-prisons","referenceExcerpt":"Michigan's classification system doesn't use AI, and AI is never used in Recidiviz tools to make decisions about anyone's security level or release.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}