{"resourceId":"michigan-aclu-mdhhs-snap-ai-case-reads","versions":[{"version":"external-ee945a20aeb0d45544a3c9c5dc285bcfdcff072329e6cfd4e0722f115581099c","resource":{"id":"michigan-aclu-mdhhs-snap-ai-case-reads","title":"ACLU of Michigan questions use of AI in assessing SNAP payments across the state","organization":"Michigan Public (reporting on MDHHS and ACLU of Michigan)","sector":"State human services / SNAP payment accuracy","geography":"Michigan, United States","publishedAt":"2026-08-02","publicationDate":"2026-08-02","eventDate":null,"sourceName":"Michigan Public","sourceLabel":"Michigan Public","sourceUrl":"https://www.michiganpublic.org/health/2026-08-02/aclu-of-michigan-questions-use-of-ai-in-assessing-snap-payments-across-the-state","evidenceClass":"independent-reporting","outcomeClass":"cautionary","topics":["governance-procurement","data-security","operating-model"],"finding":"Michigan DHHS told a Senate subcommittee it launched an AI tool for pre-certification SNAP case reads, completing 17,500 in three months; in about 1,500 reads it caught more than $150,000 in potential payment errors. The ACLU of Michigan requested records asking what evidence shows the tool works, whether it errs on complex or second-language applications, and whether federal approval was obtained, citing incentives to deny in error and Michigan's earlier MiDAS unemployment-fraud automation failure ($20M settlement).","sledRelevance":"Independent scrutiny of the same AI pattern Washington DSHS and newly funded projects are pursuing (AI-selected SNAP case review), which makes it newly relevant to October 2026 developments.","evidence":"Agency testimony figures as reported; 'potential' errors are not confirmed errors; ACLU concerns are questions, not findings; MDHHS had not responded at publication.","architectureImplications":"Pre-certification AI case reads sit before payment issuance, so false positives can delay or withhold benefits.","governanceImplications":"Raises federal approval, transparency and accuracy-evidence requirements before deployment; H.R. 1 error-rate penalties create asymmetric incentives.","securityPrivacyImplications":"Involves review of full SNAP case files; transparency demands may conflict with vendor confidentiality.","caveats":"Published Aug 2, 2026, before the October developments; outcome of the records request unknown; no audit of the tool.","streamIds":["health-human-services"],"roles":{"sales":"Sales framing: any SNAP case-review AI offering will meet civil-rights scrutiny like Michigan's. Lead with transparency, documented accuracy and federal approval readiness. A bounded engagement could prepare an agency's evidence package and public disclosure before launch. Do not cite Michigan's $150,000 as savings; it describes potential errors and is unaudited.","engineering":"Engineering framing: pre-certification case reads intercept payments, so the system must log reasons for every flag, measure false positives, and segment performance by application complexity and language. Design for public-records disclosure: model documentation, validation results and change history. Human reviewers must be able to clear a flag quickly to avoid benefit delays.","delivery":"Delivery framing: governance must include federal approval checks with USDA FNS, a published description of the tool, an appeal path, and monitoring for disparities by language. Ownership sits with the SNAP program office and agency legal counsel. Engage advocates early; Michigan's MiDAS history shows the reputational cost of automation errors."},"roleDetails":{"sales":{"customerProblem":"States want AI to cut SNAP payment errors under H.R. 1 penalties but face civil-rights scrutiny over wrongful withholding.","stakeholders":["SNAP director","Agency general counsel","Civil rights / equity officer","Legislative oversight staff"],"discoveryQuestions":["Has USDA FNS approved the process change?","What evidence shows flags are accurate?","How do limited-English and complex cases perform?","What do you disclose publicly?"],"valueHypothesis":"Bounded hypothesis: agencies that document accuracy and safeguards upfront reduce legal and reputational risk from AI case review.","potentialEngagement":"A transparency and validation package (documentation, subgroup testing, disclosure plan) for an AI SNAP case-review tool.","unsupportedClaims":["Michigan's tool is proven accurate","$150,000 is confirmed savings","AI case review is low risk"]},"engineering":{"fit":"Relevant to any pre-payment AI review design.","architectureIntegration":"AI reads cases before certification and flags them for human review; logs and explanations needed.","prerequisites":["Flag reason logging","Validation data","Segmentation by language and complexity"],"constraints":["Federal approval for process changes","Public records obligations","Benefit timeliness standards"],"security":"Case-file access controls; disclosure without exposing client data.","proofOfValue":["False-positive rate on an audited sample","Performance by language and case complexity","Average delay added to flagged cases"]},"delivery":{"work":"Establish governance, disclosure and monitoring around AI SNAP case reads.","dependencies":["Legal review","FNS engagement","Advocate engagement"],"ownership":"SNAP program office with general counsel.","skillsAdoption":"Reviewers trained to clear or confirm flags with documented rationale.","governanceCheckpoints":["FNS approval","Public tool description","Disparity monitoring"],"acceptanceCriteria":["Proposed: documented false-positive rate before scale","Proposed: no unexplained disparity by language","Proposed: flagged cases resolved within timeliness standards"],"risks":["Wrongful withholding","Litigation","Loss of public trust"]}},"retrievedAt":"2026-10-29T22:07:50-05:00","enrichedAt":"2026-10-29T22:20:00-05:00","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Concern that the tool may disadvantage limited-English applicants.","procurementImplications":"Agencies should be ready to disclose model evidence and federal approvals under public records requests.","operatingModelImplications":"Pre-issuance review changes payment timing and requires clear human accountability.","sourceVerification":{"openedUrl":"https://www.michiganpublic.org/health/2026-08-02/aclu-of-michigan-questions-use-of-ai-in-assessing-snap-payments-across-the-state","referenceExcerpt":"The civil rights organization raised concerns about tasking AI with rooting out erroneous food assistance payments, since the technology has been known to make errors.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}