In summer 2026, Michigan Department of Health and Human Services Chief Operating Officer David Knezek told a state Senate subcommittee that a Google Vertex AI "pre-certification enabled case read" tool reviewed 17,500 Supplemental Nutrition Assistance Program (SNAP) payments in three months and flagged over $150,000 in potential payment errors (Michigan Public, Aug 2, 2026). Weeks later, the ACLU of Michigan filed a FOIA request asking how that system treats complex households and non-English applications, warning that benefit decisions must not happen behind closed doors (ACLU MI, Jul 2026). Michigan families knew the stakes: from 2013 to 2015, the state's MiDAS unemployment system issued roughly 40,000 automated fraud findings with a 93% false-positive rate in a state audit report, driving over 1,000 people into bankruptcy before a $20 million settlement (Michigan AG).
When agencies deploy state AI on benefits cases across SNAP, Medicaid, WIC, or unemployment insurance, an automated risk score or data-match alert is legally a lead to investigate, never an eligibility or fraud finding. Under Goldberg v. Kelly, 7 CFR 273.13, 7 CFR 273.15, and K.W. v. Armstrong, households retain rights to specific written reasons, human review, continued benefits, and case-file inspection, as mapped in the ed.cafe Checklist below.
Confusion over automated flags harms both low-income families and frontline staff. Two victim counts appear in coverage of Michigan's MiDAS scandal because they measure distinct groups: the ACLU cites roughly 40,000 automated fraud determinations issued from 2013 to 2015, while the $20 million Bauserman settlement paid over 3,000 class claimants whose first wrongful seizure occurred on or after March 9, 2015 (Michigan Public, Aug 2, 2026). In Massachusetts, EOTSS runs a generative AI policy assistant for MassHealth call-center agents (Mass.gov EOTSS, 2025). Part of our safe use of AI tools at work hub and paired with our self-assessment for workforce and benefits counselors, this guide is general operational information, not legal advice.
How does state AI touch SNAP, Medicaid, WIC, and unemployment cases today?
State agencies rarely hand final denial power to a standalone chatbot; instead, they embed optical extraction, cross-database matching, and risk scoring into four casework stages. Knowing which stage generated an alert tells a worker or navigator where false positives hide.
Think of an airport baggage X-ray scanner. An orange box on the monitor tells a screener to unzip a suitcase and inspect an overlapping shape by hand; the scanner never shreds a boarding pass or confiscates luggage on its own. Where that analogy breaks is input quality. Airport X-rays read physical density in real time, whereas benefits software blends lagged quarterly wage files, phone photos of crumpled paystubs, and address heuristics that misread multi-generational apartments, seasonal shifts, and non-English receipts.
- Optical document extraction (
OCR): Scanners fill wage, rent, and childcare fields from uploaded photos. Year-to-date columns misread as weekly pay or skipped Spanish and Haitian Creole rent receipts inflate countable income. - Cross-database batch matching: Batch jobs match Social Security numbers against State Wage Information Collection Agency (
SWICA) quarterly reports, commercial payroll brokers like Equifax's The Work Number, and out-of-state Electronic Benefit Transfer (EBT) swipe logs. Quarterly feeds reflect prior months, not last week's shift cut. - Pre-certification risk scoring: Models rank pending files by predicted error risk, like Michigan's Vertex AI case read across 17,500 SNAP payments (Michigan Public, Aug 2, 2026). High scores trigger quality-control holds or screen warnings.
- Call-center GenAI assistants: Sidebars summarize manuals for phone reps, like Massachusetts EOTSS's MassHealth tool (Mass.gov EOTSS, 2025). Dropped categorical carve-outs can cause wrong verbal advice.
| Benefits AI pipeline | Software output | Common false-positive trigger | Mandatory human check |
|---|---|---|---|
| OCR intake. | Pre-filled wage or rent fields. | YTD gross read as weekly pay; handwritten non-English rent receipt skipped. | Open raw image; verify pay dates and landlord signature. |
| Batch match. | Wage, residency, or EBT alert. | 3-to-6 month SWICA lag; border-town grocery swipes; shared addresses. | Request current stubs; never close on raw match alone. |
| Risk score. | Error-probability rank or flag. | Seasonal hours, split custody, or high rent-to-income ratios flagged. | Treat as audit lead; check each statutory rule by hand. |
| Agent GenAI. | Summarized manual snippet. | Omitted student, disability, or immigrant-deeming carve-out. | Open underlying CFR or state rule before advising. |
What went wrong in Michigan's MiDAS disaster and Idaho's Medicaid algorithm cuts?
Two court dockets show the harm when state agencies let software cross from flagging discrepancies to cutting benefits without transparent human review: Michigan's MiDAS scandal and Idaho's K.W. v. Armstrong Medicaid ruling.
Launched in October 2013 after staffing cuts, Michigan's MiDAS system ran "robo-adjudication" whenever weekly unemployment certifications differed from quarterly employer wage reports (Justia, Cahoo). MiDAS posted a survey inside dormant portals with a 10-day deadline. Missing that alert or checking an unclear box triggered an automated fraud finding with zero human review, terminating benefits and imposing quadruple (400% statutory penalty rate on overpayments), tax intercepts, and wage garnishments up to 25% of weekly pay in collection cases.
A 2016 Michigan Auditor General review found zero intentional fraud in 93% of sampled robo-adjudicated cases (ACLU MI, Jul 2026). Because quarterly feeds lump 13 weeks together, MiDAS misread October wages earned before a November layoff as fraud. After the Michigan Supreme Court let constitutional claims proceed in Bauserman, the state signed a $20 million settlement for the more than 3,000 claimants seized within the March 9, 2015 lookback window (Michigan AG), while Cahoo v. SAS Analytics Inc. kept contractor due-process claims alive.
In Idaho, the Department of Health and Welfare used a spreadsheet formula to cut Home and Community-Based Services (HCBS) Medicaid budgets for adults with developmental disabilities by 20% to 30% in affected participant cases, claiming vendor "trade secrets" (ACLU ID). In K.W. v. Armstrong, 789 F.3d 962 (9th Cir. 2015), the Ninth Circuit barred the cuts, ruling Fourteenth Amendment due process forbids hiding public benefit formulas behind vendor confidentiality (Justia). (For private Medicare Advantage denials using nH Predict, see how billers and families appeal AI medical claim denials.)
Which federal due-process and SNAP rules forbid treating a risk flag as an eligibility finding?
Constitutional precedent and federal SNAP, Medicaid, and administrative-AI rules require human verification, plain-language reasons, and continued benefits before contested cuts take hold. Even after OMB M-25-21 (April 3, 2025) replaced OMB M-24-10, federal "High-Impact AI" benefit rules still mandate human oversight, testing, and appeal remedies (WhiteHouse PDF).
- Pre-termination hearing (Goldberg v. Kelly, 397 U.S. 254 (1970)): Subsistence benefits are Fourteenth Amendment property interests requiring timely, specific notice and a pre-cut hearing opportunity (Justia).
- 10-day specific SNAP notice (
7 CFR 273.13): Agencies must send a Notice of Adverse Action 10 days before cutting SNAP (273.13(a)) stating plain-language reasons (273.13(a)(2)) (7 CFR 273.13). Cryptic score codes violate273.13(a)(2). - 90-day SNAP appeal,
aid paid pending, and free file access (7 CFR 273.15): Households get 90 days to appeal (273.15(g)). Appealing within the 10-day notice window continues benefits unchanged (273.15(k)(1)), and273.15(p)(1)guarantees free pre-hearing copies of the entire case file (7 CFR 273.15). - USDA FNS verification (
7 CFR 272.15&273.2(f)(9)): States must report major system changes (272.15) and give 10 days to clarify unverified commercial matches (273.2(f)(9)). - Medicaid continuation (
42 CFR 431.210&431.230): Mandates 10-day advance reasons, up to 90 days to appeal, free file access (431.242), and continued coverage if appealed before the cut date (42 CFR 431.230).
| Program | Advance notice | Aid-paid-pending deadline | Appeal window | Pre-hearing file & match access |
|---|---|---|---|---|
| SNAP. | 10 days (7 CFR 273.13(a)). | Within advance notice period (273.15(k)). | 90 days (273.15(g)). | Free case-file copies (7 CFR 273.15(p)(1)). |
| Medicaid. | 10 days (42 CFR 431.211). | Before effective date (431.230). | Up to 90 days (431.221). | Full hearing-record access (42 CFR 431.242). |
| WIC. | 15 days (7 CFR 246.7(j)(6)). | Within 15 days of notice (246.9(e)). | 60 days (246.9(e)). | Pre-hearing record review (7 CFR 246.9(i)). |
| Unemployment. | State monetary/fraud window. | Certified weeks per Java rule. | 10 to 30 days (state law). | Subpoena & inspect employer match & logs. |
How can agency eligibility workers avoid automation bias on a 45-case daily queue?
Eligibility staff beat automation bias with a six-minute primary-document check whenever a score, OCR parser, or cross-match flags a reduction. Under queue pressure, automation bias tempts staff to click "Accept System Adjustment" in 20 seconds instead of checking uploads. (See case manager AI case notes, consent, and state rules.)
Worked SNAP recertification example (assumptions labelled): Maria, a Lawrence, Massachusetts banquet server with two children, earns $1,940/month across four September stubs ($485/week) and pays $1,450/month rent on a Spanish landlord receipt. OCR misreads a "Gross YTD: $24,960" box alongside a lagged $7,200 Q2 SWICA report ($2,400/month) and skips "Recibi $1,450 de renta mensual", setting rent to $0. The model flags an 88% error-probability score in its case report and queues a $192/month cut.
- Rubber-stamp path (30 seconds): Accepting pre-filled fields cuts $192/month in food aid and triggers a 90-minute fair hearing (~$120+ agency labor).
- Six-minute check: Using the ed.cafe "A Flag Is Not a Finding" Checklist, the worker checks both raw images, confirms four September stubs total $1,940 prospective income (
7 CFR 273.10(c)), restores the $1,450 rent deduction via bilingual review, and logs: "AI pre-read flagged Q2 wages ($2,400/mo) and skipped Spanish rent receipt; verified 4 Sept stubs ($1,940/mo) and signed $1,450 rent receipt."
How should community benefits navigators demand the underlying match record and lock in aid-paid-pending?
When a client brings a reduction or overpayment notice to a clinic, pantry, or career center, navigators should freeze the cut before Day 10 and compel the raw match printout.
Spot four automation tells: (1) vague "electronic data source" notices missing employer names or dollar amounts, (2) sudden $0 rent, childcare, or medical deductions after uploads, (3) full-quarter overpayment demands when the client worked only weeks, and (4) border-state grocery EBT swipes misread as moves. Execute four steps:
- Lock in
aid paid pendingbefore the cut date. File the SNAP (7 CFR 273.15(k)(1)) or Medicaid (42 CFR 431.230) hearing request by confirmed fax or portal upload before the effective date, noting SNAP recoupment rules if upheld. - Demand free pre-hearing copies (
7 CFR 273.15(p)(1)). Request the budget worksheet, rawSWICAor Work Number printout, OCR log, and pre-certification risk-read summary. - Challenge unverified commercial matches. Closing a SNAP file without a 10-day Request for Contact violates
7 CFR 273.2(f)(9). - Submit chronological pay-period proof. Pair dated stubs and stop-work letters on a one-page timeline, matching our Massachusetts tenant screening checklist after SafeRent.
The ed.cafe "A Flag Is Not a Finding" Caseworker & Benefits-Navigator Fair-Hearing Checklist
Print the ed.cafe "A Flag Is Not a Finding" Caseworker & Benefits-Navigator Fair-Hearing Checklist for eligibility desks and intake rooms. The ed.cafe Checklist pairs worker checks with a household appeal script.
THE ED.CAFE "A FLAG IS NOT A FINDING" CASEWORKER & NAVIGATOR CHECKLIST
Publisher: ed.cafe by No Worker Left Behind (EIN 88-3268312) | Not legal advice.
PART A: AGENCY ELIGIBILITY WORKER 5-POINT VERIFICATION GATE
[ ] 1. INSPECT RAW UPLOADS: Check original images vs OCR fields for YTD swaps
and skipped handwritten or non-English rent/care receipts.
[ ] 2. ALIGN PAY DATES: Confirm SWICA or payroll hits show current prospective
income (7 CFR 273.10(c)), never prior-quarter lump sums.
[ ] 3. VERIFIED-UPON-RECEIPT CHECK (7 CFR 273.2(f)(9)): Send a 10-day request
before acting on commercial payroll, EBT swipe, or risk-score alerts.
[ ] 4. SPECIFIC REASONS (7 CFR 273.13(a)(2)): State exact dollar figures, dates,
sources, and rules; never cite opaque score labels or generic codes.
[ ] 5. LOG HUMAN JUDGMENT: Add a 2-sentence note naming primary records checked.
PART B: COMMUNITY NAVIGATOR 4-STEP TRIAGE & APPEAL PROTOCOL
[ ] 1. PROTECT 10-DAY CLOCK: Notice Date: ________ | Effective Date: ________
File SNAP (7 CFR 273.15(k)) or Medicaid (42 CFR 431.230) appeal BEFORE the
cut date for continued benefits (explain SNAP recoupment if upheld).
[ ] 2. FLAG AUTOMATION TELLS: [ ] Vague "electronic source" notice [ ] Sudden $0
rent/care deduction [ ] Full-quarter wage lumping [ ] Border EBT swipes.
[ ] 3. DEMAND MATCH RECORDS: Invoke 7 CFR 273.15(p)(1) / 42 CFR 431.242(a) for
free pre-hearing copies of budget sheets, raw match screens, and AI flags.
[ ] 4. ESCALATE FRAUD PENALTIES: Refer Intentional Program Violation (IPV), tax
intercept, or UI 400% penalty notices to legal aid immediately.
PART C: COPY-READY FAIR-HEARING & MATCH-RECORD REQUEST SCRIPT
To: [State Agency Appeals Unit] | Date: [Date] | Via: [Confirmed Fax / Portal]
Name: [Client Name] | Case ID: [ID] | Notice Effective Date: [Date]
1. FAIR HEARING & AID PAID PENDING: I request a fair hearing on the [Date] notice
to [deny/reduce/end] my [SNAP / Medicaid] benefits. Filed before [Effective
Date], please continue benefits unchanged per 7 CFR 273.15(k)(1) [or 42 CFR
431.230].
2. PRE-HEARING RECORD DEMAND: Per 7 CFR 273.15(p)(1) [or 42 CFR 431.242(a)],
please send free pre-hearing copies of: (a) the budget worksheet; (b) the
full printout of any automated wage, payroll, or residency match; and (c) any
automated document-scan or pre-certification review record used here.
3. INTERPRETER: Please provide a free [Language / ASL] interpreter at hearing.
Signed: ___________________________ Authorized Rep: __________________________Where this stops being true
- Expired SNAP certification periods: Under
7 CFR 273.13(b)and273.15(k)(2), automaticaid paid pendingcovers mid-period cuts, not routine certification expiration. - Verified-upon-receipt feeds: Primary federal feeds (
SSAor validatedeDRSmatches under7 CFR 273.2(f)(9)) follow narrower rules than commercial payroll feeds, though 10-day notice and 90-day appeal rights still cover identity errors. - Intentional Program Violation (
IPV) dockets: Under7 CFR 273.16,IPVpenalties require an Administrative Disqualification Hearing, court order, or signed waiver. ReferIPVor unemployment fraud penalties to legal aid immediately, and share our pro se court-filing AI citation check. - Not legal advice: Published by
ed.cafeat No Worker Left Behind for operational literacy. Appeal clocks vary (M.G.L. c. 151Avs MichiganMESA); consult agency counsel or legal aid (MassLegalHelp.orgorMichiganLegalHelp.org) on live cases.
Questions people ask
Is AI legally allowed to decide my SNAP or Medicaid eligibility on its own?
No. Federal SNAP (7 CFR 273.2, 273.13, 273.15) and Medicaid (42 CFR Part 431) rules, backed by Goldberg v. Kelly and OMB M-25-21 high-impact AI rules, bar unreviewed algorithms from cutting benefits without human checks, specific written reasons, and fair-hearing rights. Michigan's 17,500-case SNAP tool only flags files for staff.
Why do reports say Michigan's MiDAS system harmed 40,000 people while the $20 million settlement paid 3,000?
They count two groups. From 2013 to 2015, Michigan's MiDAS software issued roughly 40,000 automated unemployment fraud findings with a 93% false-positive rate in state audit samples. The $20 million Bauserman settlement paid the more than 3,000 class claimants whose first seizure occurred on or after March 9, 2015.
How fast must a household appeal an AI-triggered SNAP or Medicaid cut to keep benefits running?
To keep continued benefits (aid paid pending) during an active certification period, request a fair hearing before the effective date printed on the Notice of Adverse Action, normally within the 10-day advance notice window under 7 CFR 273.15(k)(1) and 42 CFR 431.230. Households retain 90 days to appeal for back benefits.
Can a state agency refuse to show how an algorithm cut benefits by claiming vendor trade secrets?
No. In K.W. v. Armstrong, 789 F.3d 962 (9th Cir. 2015), federal courts ruled that Fourteenth Amendment due process bars hiding public benefit formulas behind vendor trade-secret claims. Under 7 CFR 273.15(p)(1) and 42 CFR 431.242, claimants and navigators get free pre-hearing copies of worksheets and match records.
Why do commercial payroll databases like The Work Number trigger false benefit denials?
Commercial payroll feeds and quarterly SWICA files often reflect prior-quarter lump pay, severance, or departed workers left marked "active." Because commercial payroll hits are not verified upon receipt under 7 CFR 273.2(f)(9), caseworkers must request household clarification before cutting SNAP benefits.
How does Massachusetts currently use generative AI in public benefits administration?
In MassBuys 2025 disclosures, Massachusetts EOTSS reported running a generative AI assistant for MassHealth call-center agents to search eligibility manuals during calls, not to autonomously deny claims. Staff using Commonwealth GenAI tools remain bound by AI.001 v1.1 human-verification rules.
Key terms
- Aid paid pending: The right under
7 CFR 273.15(k)and42 CFR 431.230continuing SNAP or Medicaid unchanged if appealed before the cut date. - Automation bias: Accepting an automated risk flag or OCR field under queue pressure without checking primary uploaded records.
- Notice of Adverse Action (
7 CFR 273.13): Written notice sent at least 10 days before cutting SNAP stating specific reasons and appeal rights. - Pre-certification AI case read: Predictive error screening before payout, like Michigan's 17,500-case Vertex AI SNAP review.
- Robo-adjudication: Software issuing binding benefit cuts or fraud penalties with zero human review, as in
MiDAS. - Verified upon receipt (
7 CFR 273.2(f)(9)): Primary federal feeds trusted on receipt; outside matches require 10-day follow-up before cuts.
Sources
- Michigan Public, "ACLU questions AI in SNAP" (Aug 2, 2026): Michigan Public
- ACLU of Michigan, "Concerns on AI in SNAP" (Jul 2026): ACLU MI
- Michigan AG, "$20M UIA MiDAS Settlement" (Oct 20, 2022): Michigan.gov
- MA EOTSS, "MassHealth GenAI Assistant" (2025): Mass.gov PDF
- U.S. Supreme Court, Goldberg v. Kelly, 397 U.S. 254 (1970): Justia
- Ninth Circuit, K.W. v. Armstrong, 789 F.3d 962 (2015): Justia and ACLU ID
- Sixth Circuit, Cahoo v. SAS Analytics Inc., 912 F.3d 887 (2019): Justia
- eCFR, "7 CFR 273.13 & 273.15 (SNAP)": 273.13 and 273.15
- eCFR, "42 CFR 431.230 (Medicaid)": 431.230
- OMB, "M-25-21 (replacing M-24-10)" (Apr 3, 2025): WhiteHouse PDF
