Article
Thirty-Four Thousand False Fraud Verdicts
Catching a possible fraud case for a person to examine and deciding that person is guilty are not the same power, even though it is easy for an organization to hand a system both without ever choosing to. Once that line disappears, the complaints that follow can be explained away one at a time for y
- Published
- August 25, 2026
- Updated
- August 19, 2026
- Reading time
- 10 min
The Global Signal
Michigan's Unemployment Insurance Agency deployed an automated system called MiDAS in October 2013 to detect unemployment insurance fraud. Rather than flagging suspected cases for a caseworker to review, MiDAS was configured to adjudicate fraud determinations automatically, issuing findings, quadruple penalties, and wage garnishments without a human reviewing the case first. Between October 2013 and September 2015, MiDAS adjudicated 40,195 fraud cases; state audits and subsequent litigation findings put the system's error rate at 85 percent or higher, meaning the substantial majority of those automated fraud findings were wrong. The resulting class-action litigation, Bauserman v. Unemployment Insurance Agency, produced a $20 million settlement, announced October 20, 2022, preliminarily approved by the Michigan Court of Claims on January 23, 2023, and granted final approval on January 29, 2024, covering more than 3,000 claimants wrongly accused of fraud, subjected to wage garnishment, seized tax refunds, and in some documented cases, bankruptcy.
Error rate found in MiDAS's automated fraud adjudications, per subsequent audits and litigation findings
Applies to the 40,195 fraud cases MiDAS adjudicated between October 2013 and September 2015; the $20 million settlement was announced by the Michigan Attorney General's office in October 2022.
Catching a possible case for a person to review is a different act than deciding that person is guilty and issuing the penalty automatically.
What changes when flagging and adjudicating are not the same act
A system built to catch a possible case can quietly be handed the authority to decide it, and nothing about its own metrics will show the difference.
Drawing an explicit line between flagging and adjudicating, and tracking reversal rate as a standing metric from day one, catches an error pattern that otherwise takes years and litigation to surface.
Why the Visible Metric Misleads
A fraud-detection system's own internal metrics, cases flagged, cases adjudicated, penalties issued, describe how much work the system is doing, not whether that work is correct. An 85 percent error rate is not a metric MiDAS itself was built to report; it only became visible once independent audits and litigation compelled a review of outcomes against the cases the system had actually adjudicated. The more revealing measure sits one layer beneath adjudication volume: the rate at which the system's own findings are overturned on appeal or reconsideration, tracked as a standing operational number from the moment automated adjudication begins, not discovered years later through litigation.
The Leadership Move
The right move is not to abandon automated fraud detection, which can genuinely help an understaffed agency identify cases worth a caseworker's attention. It is to draw an explicit, non-negotiable line between a system that flags a case for human adjudication and a system that adjudicates the case itself, and to track that system's own reversal rate as a standing metric from its first day of operation, not as something an outside audit eventually has to reconstruct.
- Ownership
The technology team that builds an automated fraud-detection system typically owns whether it functions correctly against its own test cases. Agency leadership owns the separate, consequential decision of whether that system is authorized to adjudicate outcomes autonomously or only to flag them for a person. When leadership treats early complaints as evidence the system needs defending rather than evidence the system needs reviewing, the distinction between flagging and adjudicating stops being examined at all.
- Tradeoff
Requiring human adjudication of every fraud flag is slower and more expensive than full automation, particularly for an agency processing a large caseload with limited staff. The alternative, Michigan's case shows starkly, an 85 percent error rate compounding across roughly two years of automated adjudication, cost far more, in the $20 million settlement, in the documented bankruptcies and garnishments, and in the two years during which the pattern went unreviewed as a pattern.
- Human consequence
Claimants wrongly accused of fraud experienced wage garnishment, seized tax refunds, and in some cases bankruptcy, based on a determination no person had reviewed before it was issued. The complaints reaching unemployment lawyers within months of the system's launch were real people experiencing that consequence directly, well before the state's own leadership treated the pattern as something other than a system to be defended.
Implication for Operators
Any organization that automates a consequential determination, not just a flag for a person to review, should assume that determination's error rate is invisible to standard operational metrics unless a specific, independent mechanism tracks reversals on appeal from the system's first day of operation. The practical shift is drawing the flag-versus-adjudicate line explicitly before deployment, and treating early complaints about a new automated system as evidence to review, not evidence to defend against.
MiDAS was not secretly unreliable from the state's perspective; the reliability question was never being asked in a form that would have surfaced the answer. A system built to flag possible fraud was allowed to adjudicate it instead, and the gap between those two authorities produced an 85 percent error rate that took roughly two years and a lawsuit to become visible as a number, rather than a string of individually explained complaints.
A fraud-detection system making mistakes explains little. A flag handed the authority to adjudicate, with no standing check on how often that adjudication was wrong, while the complaints it generated were treated as something to defend against rather than review, explains everything.
When evidence contradicts the current plan, how often does leadership review the decision rather than explain the evidence away?
FAQ
Did MiDAS flag cases for a person to review, or decide them automatically?
It decided them automatically. Fraud determinations, penalties, and wage garnishments were issued by the system directly, without a caseworker reviewing the case before the determination was finalized, which is the specific design choice this article's argument turns on.
How high was the system's actual error rate?
Subsequent audits and litigation findings place it at 85 percent or higher across the 40,195 fraud cases MiDAS adjudicated between October 2013 and September 2015, meaning the substantial majority of automated fraud findings during that period were incorrect.
Were there warning signs before the state halted the system in 2015?
According to subsequent reporting, complaints from affected claimants reached unemployment lawyers within a relatively short period after the October 2013 launch, and agency staff raised internal concerns that reporting describes leadership as stubbornly defending the system against, rather than reviewing.
What did the $20 million settlement cover?
The $20 million settlement, announced October 2022 and granted final court approval in January 2024, compensates more than 3,000 claimants wrongly subjected to fraud findings, wage garnishment, and seized tax refunds under the automated adjudication system, resolving the Bauserman v. Unemployment Insurance Agency litigation.
Does this mean states should not automate fraud detection at all?
No. The distinction this case makes clear is between automation that flags a case for a person to decide and automation that decides the case itself, with the second requiring a standing reversal-rate check that MiDAS never had until litigation forced one.
Related intelligence
Article
The Data One Company Published and the Other Didn't
Two autonomous-vehicle companies experienced the same category of incident: a real-world edge case their systems had not been explicitly designed to handle. One tried to manage the disclosure. The other built continuous, third-party-audited disclosure into how it operates, and that second choice is
Article
Insurance Renewal Confusion Is a Dead End Hiding in Plain Sight
A renewal notice arrives, coverage or pricing may have changed, and the explanation is often written in language most policyholders don't have time or expertise to fully parse. Here's why that's a dead end, not just paperwork.
Article
The Algorithm and the Randomized Trial
Two education systems built algorithms to help make decisions about individual students at scale. One imposed its algorithm nationally, on results day, with no pilot and no override. The other tested its tool for years, on a smaller population, before trusting it broadly, and built it to route stude