{"resourceId":"michigan-doc-classification-pilot-staff-concerns-not-ai-wcmu-20261106","versions":[{"version":"external-fb9d63e696dbcce06a537e654e924f479d2bf54673705021fb0a08bf6c9b1ac8","resource":{"id":"michigan-doc-classification-pilot-staff-concerns-not-ai-wcmu-20261106","title":"Michigan prison classification pilot starts Dec. 1 amid staffing crisis; MDOC says the score does not use AI and staff make the final call","organization":"Michigan Department of Corrections with Recidiviz (reported by WCMU Public Media)","sector":"Corrections / security classification","geography":"Michigan, United States","publishedAt":"WCMU Public Media, Nov. 6, 2026 (Blace Carpenter)","publicationDate":"2026-11-06","eventDate":"2026-12-01","sourceName":"WCMU Public Media","sourceLabel":"MDOC to pilot new classification system, aims to lower prison violence","sourceUrl":"https://www.wcmu.org/local-regional-news/2026-11-06/mdoc-to-pilot-new-classification-system-aims-to-lower-prison-violence","evidenceClass":"independent-reporting","outcomeClass":"emerging","topics":["data-security","governance-procurement","accessibility-workforce"],"finding":"WCMU reports that from Dec. 1 the Michigan Department of Corrections will pilot a new classification system at select prisons, created with the nonprofit Recidiviz, which was hired to examine historical prison data, hold focus groups and talk with employees. MDOC cites over 300 officer assaults by inmates and over 500 assaults on other prisoners in 2025. After initial coverage assumed AI was involved, MDOC and Recidiviz said AI is not used in classification; MDOC’s Nov. 5 statement says the tool calculates an assessment score that staff review before a final decision. A St. Louis Correctional Facility employee says misclassification has contributed to assaults there, and the report describes staffing vacancy rates over 30 percent at multiple facilities. Pilot sites are not named.","sledRelevance":"Shows how a corrections analytics change can be mislabelled as AI, and how staff voices and staffing conditions shape whether a classification tool is trusted. Useful for corrections buyers separating AI features from rules-based scoring.","evidence":"Public-radio reporting with agency statements; no outcome data, validation method or pilot sites published in the article.","architectureImplications":"A scoring tool that stays rules-based and reviewable by staff needs data feeds from case management and a way to record overrides.","governanceImplications":"Clear public statements on whether AI is involved, plus human final decision, were issued only after press coverage; earlier disclosure would reduce confusion.","securityPrivacyImplications":"Classification uses sentence, history and behavior data that affect movement and programs for incarcerated people; access and error correction matter.","caveats":"Pilot not started; outcome claims are agency expectations; Recidiviz describes AI use elsewhere in its products for meeting transcription; Michigan-specific.","streamIds":["public-safety"],"roles":{"sales":"Start with a labelling lesson: a rules-based classification score was widely reported as AI until the agency and vendor corrected it. Corrections buyers need clarity on what is AI, what is statistics, and who decides. Offer a transparency and requirements engagement, and avoid repeating the agency’s prediction-improvement expectations as evidence. Staffing shortages and officer trust are the real adoption constraints raised here.","engineering":"The tool described is a points-based score reviewed by staff, so the engineering questions are data quality, override logging and refresh frequency rather than model risk. Prerequisites include clean case-management feeds and a record of staff overrides. Proof of value should compare scores with later incident data and override rates during the pilot, using the agency’s own metrics and stated validation plan.","delivery":"Support the pilot with a measurement plan: override rates, reclassification timing, incident trends at pilot sites versus comparison sites, and staff feedback. Assign ownership to the classification unit with facility wardens, brief staff on how scores are produced, and set a review point after the pilot. Acceptance criteria are proposals; the article reports no results yet. This is interpretation from a single source with limited transferability."},"roleDetails":{"sales":{"customerProblem":"Outdated classification instruments contribute to misplacement and staff unease, while AI labelling confusion erodes trust.","stakeholders":["Corrections director","Facility wardens","Classification unit","Staff union representatives","Legislative oversight","Advocacy groups"],"discoveryQuestions":["Is any part of our scoring AI or machine learning, and how do we say so publicly?","How often do staff override the score?","Which data sources feed classification and how current are they?"],"valueHypothesis":"Transparent, reviewable scoring plus override tracking may improve trust and let the agency test accuracy; outcomes are not shown.","potentialEngagement":"Requirements and measurement-plan workshop for a classification pilot, including public explanation materials.","unsupportedClaims":["That the new system predicts violence about 30 percent better","That it reduces assaults","That staffing problems will improve"]},"engineering":{"fit":"Rules-based scoring and data pipeline support; no AI claim.","architectureIntegration":"Connect case management, incident and housing data to the scoring tool with override logging.","prerequisites":["Case management data access","Incident data history","Override capture workflow"],"constraints":["Pilot sites not named","Validation method not published","State-specific policy"],"security":"Restrict access to classification records and log views and changes.","proofOfValue":["Compare score and later incident rates at pilot sites","Report override rates and reasons weekly"]},"delivery":{"work":"Pilot measurement, staff briefing and review checkpoint.","dependencies":["Pilot site selection","Data quality fixes","Staffing levels"],"ownership":"Classification unit lead with facility wardens.","skillsAdoption":"Train staff on score construction and override documentation.","governanceCheckpoints":["Post-pilot review","Public statement on AI use","Legislative briefing"],"acceptanceCriteria":["Override rate and reasons reported monthly","Score reviewed for disparate impact before expansion","Staff survey response collected at pilot sites"],"risks":["Staffing shortages limit review quality","Public confusion about AI","Pilot data too small to judge"]}},"retrievedAt":"2026-11-09T03:57:53Z","enrichedAt":"2026-11-09T04:01:20Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Staffing vacancies above 30 percent and officer concerns about misclassification bear on whether staff can review scores carefully; the source gives no pilot staffing plan.","procurementImplications":"The tool was built through a nonprofit partnership using historical data, focus groups and staff input; buyers should ask vendors to state plainly whether AI is involved.","operatingModelImplications":"Staff make the final classification; the operating model depends on reviewer capacity and override tracking.","sourceVerification":{"openedUrl":"https://www.wcmu.org/local-regional-news/2026-11-06/mdoc-to-pilot-new-classification-system-aims-to-lower-prison-violence","referenceExcerpt":"The updated classification tool does not replace corrections staff’s decision-making abilities","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}