The Fraud Algorithm That Accused Tens of Thousands: Michigan’s MiDAS Reckoning and What the Refunds Actually Settled

Start with the design, not the dollar figure

Michigan’s unemployment fraud detection system did not fail because its model was slightly miscalibrated. It failed because the process around the model had already decided the outcome before any human being looked at a case. That distinction matters for anyone who sits in a procurement meeting or an oversight hearing and is told that the next system will be more accurate.

The Michigan Integrated Data Automated System, known as MiDAS, was built by contractor Fast Enterprises starting in 2011 for $47 million and rolled out in 2013, according to a case study published by the Beeck Center’s Benefits Tech Accountability project. During its first two years in use, the system falsely accused 40,000 residents of fraud—a five-fold increase over expected numbers. The University of Michigan’s Science, Technology, and Public Policy program summarizes the same figure: approximately 40,000 Michiganders wrongly accused between 2013 and 2015, with an error rate reaching 93%.

Those numbers are the headline. The mechanism is the lesson.

How a data mismatch became a fraud finding

MiDAS flagged any data discrepancy as fraud, no matter how trivial, and required the applicant to clarify within 10 days, according to the Benefits Tech Accountability case study. The system also averaged an applicant’s entire income rather than examining individual paychecks, which produced discrepancies between reported and system-determined income—and more false fraud flags.

Each of those design choices is a decision someone made. None of them required a sophisticated model. They required a process that treated the claimant as guilty until proven otherwise, and then made proving otherwise nearly impossible.

The notice problem compounded the math problem. Applicants received notices only on the unemployment portal, not in the mail—a portal they had no reason to check if they were no longer receiving benefits. The dispute questionnaire was pre-filled with responses that would flag the applicant for fraud. A person who did manage to find the notice and open the form was handed a document designed to confirm the accusation.

This is the part of the MiDAS story that gets lost when the conversation turns to algorithmic accuracy. A system that flags any discrepancy, hides notice in a portal, and pre-fills the dispute form against the claimant has already decided the outcome. Better model performance would not have fixed it. A human-review requirement before adverse action would have.

The error rate is the accountability metric

In 2016, the Michigan Auditor General reviewed 22,000 MiDAS fraud determinations and found that 93% did not actually involve fraud, according to the Benefits Tech Accountability case study. That is the number procurement officers should be required to publish—not the fraud-detection rate, not the dollar value of claims flagged, but the false-positive rate measured against an independent review.

A fraud system’s value is bounded by its false-positive cost. MiDAS externalized that cost onto claimants through tax refund interception, wage garnishment, and repayment demands. The University of Michigan explainer documents the downstream harm: lost homes to foreclosure, seized tax returns, repossessed vehicles, destroyed credit, bankruptcy, and unhoused claimants. The system did not merely deny benefits. It pursued people for money they did not owe.

When a public agency buys a fraud-detection tool, it is buying a claim about error rates. The MiDAS record shows what happens when no one requires that claim to be tested before collections begin.

What the courts actually settled

The legal outcomes are narrower than the harm, and practitioners should be precise about the difference.

In a 2017 settlement in Zynda v. Zimmer, Michigan’s unemployment agency agreed to stop using MiDAS’s automated functions without human review and to reverse and refund certain fraud determinations, according to the Benefits Tech Accountability case study. That settlement addressed the process defect directly: no automated adverse action without a human in the loop.

The class action Bauserman v. Unemployment Insurance Agency produced the larger reckoning. In April 2019, the Michigan Supreme Court held that the class action was timely filed but that only claimants whose money was first collected on or after March 9, 2015 could be included in the potential class, according to the Michigan Attorney General’s October 2022 announcement. In July 2022, the Court held that the plaintiffs could seek monetary damages from the state based on the alleged violation of the state constitution.

The state announced a $20 million settlement in October 2022. The University of Michigan explainer reports that the Court of Claims approved the settlement in late January 2024, that the class consisted of approximately 3,000 Michiganders, and that the state will refund plaintiffs over 100% of what the agency falsely claimed they owed.

Read those numbers together. Roughly 40,000 people were wrongly accused. The settlement class was approximately 3,000, narrowed by the 2019 ruling to those whose money was first collected after March 9, 2015. The settlement is a partial remedy, not a full accounting. It compensates a subset of the people the system harmed, and it does so years after the harm.

The Cahoo v. SAS Analytics litigation produced a different kind of outcome. In 2023, the Sixth Circuit reversed the lower court’s denial of qualified immunity to agency officials named in the suit, according to the Benefits Tech Accountability case study. The court acknowledged problems with the MiDAS program but found the remaining claims concerned an internal fraud red flag that did not deprive benefits without additional process, and that the officials lacked sufficient connection to the pre-deprivation process defects. A dissent flagged factual disputes about the protections individuals received and the role of state officials in decision-making.

That ruling is a statement about individual liability, not about whether the process was adequate. Qualified immunity outcomes should not be read as vindication of the system. They are a reminder that when a public agency deploys an automated system, the people who designed the process may never be held personally accountable for it—which is precisely why the contract and the procurement record have to carry the accountability that courts may not.

The replacement system and the questions that should precede it

Michigan has contracted Deloitte to design and implement a new automated system, with the agency aiming for it to be operational by 2025, according to the University of Michigan explainer. The agency’s director has described reforms including overpayment waivers, backlog reduction, and a human-centered focus, per the Attorney General’s announcement.

The useful oversight questions for the replacement are concrete and answerable before go-live:

• Who signs off before an adverse action? A contract clause requiring a human adjudicator to review and sign any fraud finding before garnishment or refund interception would have blocked the MiDAS harm chain at its first link, regardless of how the model scored the case.
• How is notice delivered? Portal-only notice failed claimants who had no reason to log in. The contract should specify mail or multiple channels for adverse actions.
• Are dispute forms tested with claimants? A pre-filled form that flags the applicant is a design defect, not a training issue. Usability testing with people who have actually navigated the benefits system should be a contract deliverable.
• Is the error rate reported publicly on a schedule? The Auditor General’s 93% finding came after the harm. A contractual requirement to publish false-positive rates measured against independent review, at defined intervals, would surface the problem while it is still fixable.
• What triggers a halt? The Zynda settlement stopped unreviewed automated functions only after litigation. A contract should specify the error-rate threshold or audit finding that suspends automated adverse action automatically.

None of these clauses requires new technology. They require the agency to decide, in writing and before launch, that the burden of proof sits with the government rather than the claimant.

What the refunds settled—and what they did not

The Bauserman settlement returned money to people the system wrongly pursued. It did not return the years, the credit history, the housing, or the certainty that the next system will be different. The 2019 ruling that narrowed the class means many people harmed by MiDAS were outside the settlement entirely.

The MiDAS record is often told as a cautionary tale about algorithmic accuracy. It is more usefully told as a procurement story. A $47 million system was deployed with a process that treated any discrepancy as fraud, hid notice where claimants would not see it, and pre-filled the dispute form against them. The Auditor General later found that 93% of reviewed determinations did not involve fraud. The state paid $20 million and refunded a subset of what it collected. The officials named in the companion litigation were granted qualified immunity.

That is the full ledger. The next time a vendor presents a fraud-detection system and the conversation turns to model performance, the question to raise is not how accurate the model is. It is who reviews the finding before the garnishment notice goes out, how that person’s decision is documented, and what error rate would cause the agency to stop.

Questions readers can raise in the next oversight meeting

• What is the false-positive rate of the current fraud-detection system, measured against independent human review, and when was it last published?
• Does the contract require human review and sign-off before any adverse action, including garnishment and refund interception?
• How are adverse-action notices delivered, and has the agency tested whether claimants actually receive and understand them?
• What error-rate threshold or audit finding triggers suspension of automated adverse action under the contract?
• For the replacement system, what human-review, notice, and error-reporting requirements are written into the contract, and are they enforceable?

The MiDAS reckoning produced refunds, a settlement, and a replacement contract. Whether it produces a different process depends on what the next contract says before the system goes live.