Michigan spent nearly two years accusing jobless workers of unemployment fraud with no human examiner

The letter that said Grant Bauserman owed $19,910

Grant Bauserman collected unemployment benefits in Michigan from September 2013 into March 2014. On December 3, 2014, the state's Unemployment Insurance Agency sent him redeterminations, official notices that reopen a closed claim, finding that he had been ineligible and had intentionally misrepresented his situation. He was told he owed $19,910. He protested. On June 16, 2015, the state intercepted his tax refund and applied it to the debt. Fifteen months into the case the agency reversed itself. On September 30, 2015 it declared the redeterminations null and void, and Bauserman was repaid.¹

No caseworker had interviewed him. The questionnaire that opened the case had gone only to his account on the agency's website, with no letter or email, and the finding that he had committed fraud was produced by software. The system was the Michigan Integrated Data Automated System, MiDAS, and from October 2013 to August 2015 it decided fraud on its own. Teddy Broe, the other named plaintiff, was assessed penalties and interest of more than $8,000 the same way.¹ The two of them stood in for tens of thousands.

The examiner MiDAS removed

Michigan launched MiDAS in October 2013 to modernize how its Unemployment Insurance Agency processed claims. The program was large and ordinary. In fiscal year 2014 the agency paid about $1 billion in benefits across 611,503 claims, and the state's own performance auditor confirmed the system was fully implemented that October.² Trade reporting put the contract at roughly $47 million and the automated window at October 1, 2013 through August 7, 2015.³

The name MiDAS is shorthand for three contracts. Fast Enterprises, a Colorado firm, agreed around August 2011 to build and run the MiDAS platform, the software that carried a claim from filing through collection. SAS Institute, of North Carolina, contracted around December 2012 to build a separate Enterprise Fraud Detection Software, the engine that generated the fraud analysis, wired into MiDAS. CSG Government Solutions, in Chicago, had run project oversight since about January 2010. The fraud logic that flagged claimants was not one company's work, which is why the question of who is to blame, the vendor or the agency that configured it, has more than one defendant.⁴

Under the old way of working, a data mismatch, a disagreement between what a claimant reported and what an employer or another agency reported, was the start of an inquiry, not the end of one. A human examiner looked at the discrepancy and decided what it was: real fraud, an employer's clerical error, or a good-faith dispute about dates or amounts. MiDAS kept the flag and removed the person. It searched claimant records against employer, state, and federal data reaching back six years, treated any discrepancy as a possible fraud, and never checked whether an employer had simply filed the numbers wrong.⁵

How a data mismatch became a fraud finding

One step manufactured fraud out of arithmetic. MiDAS used an income-spreading formula that averaged a claimant's quarterly earnings evenly across the weeks of the quarter. A person who earned money early in a quarter and correctly reported no income for a later week would, after the averaging, appear to have hidden earnings. The system read that manufactured discrepancy as intentional misrepresentation.⁵

On a flag, MiDAS sent a multiple-choice questionnaire to the claimant's online account, gave ten days, and stated no basis for the suspicion. An answer it read as an admission, or no answer inside ten days, produced what the courts came to call a robo-adjudicated fraud finding.⁵

Removing the examiner is the change that scaled the harm. Data-matching on its own produces false alarms at whatever rate innocent mismatches occur, and that rate is never zero. Employers misfile, dates shift, a formula averages away a true answer. The examiner was the step that caught those before they became findings. Take the examiner out, and every innocent mismatch becomes a fraud determination. Two more choices turned that stream of false findings into ruin. The claimant, not the state, had to notice a questionnaire in a website account many had not opened since their last claim, and answer within ten days, so silence read as a confession. And the penalty was four times the benefits, the maximum Michigan law allowed, charged even when the claimant had received no benefits at all.⁵ Assessments typically ran from $10,000 to $50,000, and some passed $187,000. A wrong flag did not produce a correction. It produced a five-figure debt, collected by garnishing wages and seizing tax refunds for up to six years.⁵

Michigan stopped the fully automated mode in August 2015. A settlement in an earlier case, Zynda v. Arwood, formalized the retreat in February 2017, when the state agreed to suspend automated-system collection and barred fraud findings based solely on computer-identified discrepancies.⁴

When the state counted, the findings collapsed

The scale of the error is where a headline most wants a single number, and the record does not supply one. It supplies two, from two reviews, measuring different things.

In December 2016 the agency's own review of its automated fraud cases became public. Of 22,427 determinations the system had made, more than 93 percent were overturned as not fraud.⁴ Contemporary state reporting described it as a new state review, counted 53,633 cases the system had flagged for fraud, and noted the state had already refunded $5.4 million.⁶ This is the famous "93 percent" figure, and it is easy to hang on the wrong office. Both federal opinions in the litigation credited it to the Michigan Auditor General.⁴ The only performance audit the Auditor General actually ran on MiDAS, released in February 2016, examined the system's security and access controls and produced no fraud-error rate at all.² The Michigan Supreme Court later placed the figure where it belongs, with a study conducted by the agency itself.¹

The fuller count came in August 2017, when the Talent Investment Agency finished reviewing every case from the automated period. That review covered 49,910 people and refunded more than $20.8 million, and it separated the cases by how each was first decided.⁷

Both rows sit inside 62,784 penalty cases that were never appealed, and they teach different things. The 85 percent is the rate for findings a computer made with no human review, and it is the number that matters most. The 44 percent is what happened after the state added a human on top of the same flags. Bolting a reviewer onto bad flags still left nearly half the findings wrong. The December 2016 "93 percent of 22,427" and the August 2017 "85 percent of 40,195" count different populations from different reviews. Neither is "the error rate."

Two courts, two opposite endings

The litigation gives the affair two endings that point in opposite directions. In July 2022 the Michigan Supreme Court held that wrongly accused claimants could sue the state itself for money damages for the due-process violation. The right they invoked was enforceable, the court reasoned, and no other adequate remedy existed.¹ In October 2022 the state agreed to a $20 million settlement of the class action. A limitations ruling had already narrowed the class to people whose money was first collected on or after March 9, 2015, and a court approved the settlement in late January 2024.⁸

The state's strongest counter came in a separate federal track. Individual agency supervisors argued they were entitled to qualified immunity, the doctrine that shields officials from personal liability unless they broke a right that was clearly established at the time. In June 2023 the Sixth Circuit agreed and reversed the denial of immunity for two of them. On the fuller record that discovery had produced, the court found the process more protective than the complaint alleged: claimants received notices of determination before any money was taken, with a thirty-day appeal window and multiple levels of administrative review. And it observed that what began as a case about wrongly terminated benefits had become a case about collecting benefits the claimants had in fact been paid.⁹

The two rulings answer different questions, and both can stand. The Michigan Supreme Court decided that claimants may sue the state for damages.¹ The Sixth Circuit decided only that two named supervisors cannot be held personally liable, on a record showing some notice and appeal existed.⁹ Neither court settled the fact the affair turns on: how many findings were wrong, and who should bear the loss. The state's own two reviews are the best evidence of that, and the case-level data behind them was never published, so no outsider can check the count.