What Is the Michigan Case Accuracy Review Assistant?
The Michigan Department of Health and Human Services (MDHHS) deployed a tool called the Michigan Case Accuracy Review Assistant to support SNAP case reviews. SNAP — formerly known as food stamps — currently serves roughly 1.4 million low-income Michigan households.
According to the department, the tool works by comparing information across case records to flag possible inconsistencies, missing data, or calculation issues. A human caseworker then reviews those flags and makes the final eligibility decision.
MDHHS has been direct on one point: the tool does not approve or deny benefits on its own.
“The tool does not make eligibility determinations, assign fraud scores or independently change a household’s benefits.”
That’s the official position. But critics argue the line between “flagging” and “deciding” can get blurry fast — especially at scale.
Why Was This Tool Deployed?
The timing matters here. The rollout followed the passage of a federal spending bill that cut an estimated $186 billion from SNAP over the next decade by shifting costs to states with high payment error rates.
The federal government fully funds SNAP for states with an error rate below 6%. Michigan’s error rate was sitting at 9.9% — well above that threshold.
MDHHS Chief Operating Officer David Knezek testified before lawmakers in March, framing the AI tool as a practical fix. He said it allowed the agency to scan every case before payments go out and target cases with the highest likelihood of payment errors.
Over three months, he said the tool helped catch $150,000 in possible errors.
That’s the state’s case for the technology: reduce errors, protect funding, keep benefits flowing to people who qualify.
What the ACLU Is Concerned About
The ACLU of Michigan sent a letter to MDHHS on July 30, raising several pointed concerns.
The Transparency Problem
The ACLU says it remains unclear exactly how AI has been implemented — what data it uses, how it weights inconsistencies, and what thresholds trigger a flag. That opacity is the core issue.
“It appears that AI is now playing some significant role — although how significant remains murky — in decisions about who gets to keep their benefits, and who loses them,” the letter stated.
The Incentive Structure Problem
The federal government does not count wrongly denied SNAP applications as errors. That creates a structural incentive for states to lean toward denials when a case looks uncertain — and an AI tool that flags inconsistencies could amplify that pressure.
Nationally, an estimated 4.5 million people have lost food assistance since the law went into effect last July. In Michigan alone, that figure is roughly 113,000 people.
The Historical Precedent Problem
Michigan has been here before. From 2013 to 2015, the state’s automated unemployment system falsely accused roughly 40,000 people of fraud. That resulted in a $20 million settlement.
Phil Mayor, the ACLU of Michigan’s deputy legal director, drew the connection directly:
“Given the fiasco Michigan experienced when it used computer-reliant decision making for unemployment insurance benefits, it is a real concern that AI is now involved in deciding who does or doesn’t get help putting food on their family’s table.”
What the ACLU Is Demanding
The ACLU has filed a Freedom of Information Act request for documents on two specific things:
- Whether Michigan disclosed its use of AI to the federal government before adopting the tool
- Any instances where the AI produced inaccurate, false, fabricated, or unreliable results
These aren’t abstract demands. If the tool has a documented error pattern, the public and affected households have a right to know.
The Broader Stakes for AI in Government
This situation reflects a pattern that’s showing up across public sector AI deployments: a tool gets introduced to solve a real operational problem, the technical details stay internal, and scrutiny arrives only after the system is already running.
The MDHHS position — that humans remain responsible for every eligibility decision — is a reasonable safeguard on paper. But it only holds up if caseworkers have the time, training, and authority to genuinely override AI-generated flags rather than rubber-stamp them.
That’s the accountability gap the ACLU is pointing at. And it’s a gap that matters regardless of how well-intentioned the tool’s deployment was.
What to Watch
A few things will determine how this plays out:
- FOIA results: If documents show the AI tool produced unreliable outputs or that federal disclosure was skipped, the political and legal pressure on MDHHS will increase significantly.
- Error rate data: If Michigan’s SNAP error rate drops sharply, the state will use that as validation. If denials spike without a corresponding drop in overpayments, that’s a different story.
- Federal oversight: Whether the federal government requires states to disclose AI use in benefits administration is an open question — and Michigan’s case may help force that conversation.
The core issue isn’t whether AI should be used in government at all. It’s whether the public gets enough visibility into how it works to hold agencies accountable when it doesn’t.
Comments (0) No comments yet
Want to join this discussion? Login or Register.
No comments yet. Be the first to share your thoughts!