MDHHS Deploys AI in SNAP Reviews Sparking Concerns Over False Positives
MDHHS publicly announced the deployment of an AI-assisted SNAP case reader using Vertex AI, with experts warning of potential false positives and drawing parallels to MiDAS-era errors. Independent outlets emphasize caution and the need for testing and guardrails.
The SNAP AI tool could generate mass false positives by conflating innocent errors with fraud, echoing MiDAS-era failures.
Key facts
- What
- MDHHS publicly announced the deployment of an AI-assisted SNAP case reader using Vertex AI, with experts warning of potential false positives and drawing parallels to MiDAS-era errors.
- Incident date
- Mar 17, 2026
- Who
- Michigan Department of Health and Human Services (MDHHS)
- Failure mode
- Hallucination
- AI surface
- Copilot
- Severity
- High
What happened
MDHHS announced the deployment of an AI-assisted SNAP case reading tool to review all cases and improve payment accuracy before disbursement. The system reportedly runs on Google Vertex AI, with officials describing it as a way to scale review. Experts warn that such tools can misclassify innocent clerical errors as fraud, referencing Michigan's MiDAS-era experiences as a cautionary parallel.
What broke inside the model
- 01 · TriggerA user asks for a fact, a citation, or a figure.
- 02 · Model stepThe model writes a fluent, confident answer.
- 03 · Control gapNothing ties the claim back to a real source.
- 04 · FailureA fabricated fact ships as if it were verified.
- 05 · ConsequenceThe false claim reaches a customer, a court, or the public.
Confidence holds, and even spikes, as the claim detaches from any source.
The mechanism failed when the AI incorrectly labeled non-fraudulent clerical errors as fraud, potentially flagging large numbers of SNAP cases. Critics argue the deployment occurred with limited testing and insufficient safeguards, increasing the risk of widespread false positives.
What it cost
Sources
- PressMichigan's Health and Human Services Department Deploys AI to Process SNAP Applicationsmichiganadvance.com
- PressSNAP benefits AI could mislabel fraud; lawsuits raise concernsmichiganindependent.com
Cite this entry
https://failureindex.ai/failures/mdhhs-deploys-snap-reviews-sparking-concernsAI Failure Index. "MDHHS Deploys AI in SNAP Reviews Sparking Concerns Over False Positives" (FI-0200). Realm Labs. https://failureindex.ai/failures/mdhhs-deploys-snap-reviews-sparking-concerns (indexed Jun 5, 2026).Data fields CC-BY 4.0, prose citation permitted. Incident ID FI-0200. Full dataset at /data.
Note from Realm Labs, the Index steward
How Realm would have caught this
- Prism
- OmniGuard
- AI Detection & Response (AIDR)
A runtime layer that watches the model's internal state can flag the moment a model commits to a claim it has no support for, and hold or reroute the response before it reaches a user. Realm reads those signals in real time rather than grading the transcript after the fact.