After rolling out a flawed automated system in 2014, Michigan falsely charged thousands with unemployment fraud and collected millions in fines
REPUBLISH … I N 2014, Carmelita Colvin was living just north of Detroit and taking classes at a local college, when she received a letter … Tweets: @undarkmag , @undarkmag , @undarkmag , @undarkmag , @undarkmag , @undarkmag , @undarkmag , @undarkmag , @lil_aphasia , and @undarkmag Tweets: Undark / @undarkmag : Amid later outcry, the agency later ran a partial audit and admitted that 93 percent of the changes had been erroneous — yet the agency had already taken millions from people and failed to repay them for years. https://undark.org/... Undark / @undarkmag : In many cases, governments are hiring private companies to build automated systems, and since the companies want to protect their intellectual property, they sometimes hide details of how the products work. https://undark.org/... Undark / @undarkmag : Nearly half the states in the U.S. have modernized the software for their unemployment insurance systems. In many cases, these updates actually help claimants more easily file for benefits. But this is not always the case. https://undark.org/... Undark / @undarkmag : While the agency says it repaid $21 million, attorneys in the class-action lawsuits argue that this doesn't account for all of the damages. People suffered long-term harm — many came out with damaged credit and lost job opportunities and homes. https://undark.org/... Undark / @undarkmag : Government use of automated systems is on the rise in many domains. But the people who use the algorithms don't always understand how they work, and the functions are even murkier to the public. https://undark.org/... Undark / @undarkmag : “These types of tools can be used to inform human judgment, but they should never be replacing human beings,” said Frank Pasquale, a professor of law at the University of Maryland who studies accountability in the use of opaque algorithms. https://undark.org/... Undark / @undarkmag : Many of those charged falsely with unemployment fraud came out with damaged credit and lost job opportunities and homes. More than a thousand filed for bankruptcy. https://undark.org/... Undark / @undarkmag : Over a two-year period, the agency falsely charged more than 40,000 people, billing them about five times the original benefits, which included repayment and fines of 400 percent plus interest. https://undark.org/... Lil'Aphasia / @lil_aphasia : Unemployment Insurance Agency used an automated system to charge more than 40,000 people, billing them about five times the original benefits The agency later admitted 93%the charges were erroneous. https://undark.org/... @szymielewicz @agatafshsh @JedrzejNiklas Undark / @undarkmag : The automated system in Michigan is “a case study in all the ways an algorithm can go wrong,” says Michele Gilman, a professor at @ubaltlaw. “The citizens shouldn't be the guinea pigs in testing whether the systems work.” https://undark.org/...
Context & Ripple Effects
This Undark investigation documents one of the earliest large-scale failures of government automation: Michigan's Unemployment Insurance Agency rolled out its fraud-detection system in 2014, and it flagged more than 40,000 claimants — including Carmelita Colvin, a student north of Detroit who received the letter out of nowhere — billing them roughly five times their original benefits and collecting millions. A partial audit later found 93 percent of the determinations were erroneous, and by then the money was long gone.
The case reads as the opening chapter of a pattern the related coverage keeps confirming: Arkansas's automated health assessments cut home care for disabled patients two years later (algorithms rationing state-sponsored care), ID.me's facial-recognition checks locked unemployment applicants out of benefits during the pandemic, and researchers now flag that third-party fraud-detection vendors are often overpaid and under-supervised. Michigan shows what happens when the error surfaces after collection, not before.
First-order effects
- More than 40,000 wrongly charged claimants are owed restitution beyond the $21 million the agency says it has repaid, which attorneys argue does not cover damages like wage garnishment and destroyed credit — hence the class-action lawsuits against the agency.
Second-order effects
- Other states running similar automated eligibility and fraud systems face pressure to audit their own error rates proactively, since Michigan's 93-percent erroneous-finding figure became the reference case for what unchecked automation costs claimants and agencies alike.
Third-order effects
- If the litigation and audits keep landing this way, governments will face structural demands to shift accountability upstream — requiring vendors and agencies to prove accuracy before deployment and to front repayment obligations when systems err — rather than compensating victims years after collections.
The trend: Government automation of high-stakes determinations — from unemployment fraud to benefit eligibility to sentencing support — keeps producing mass errors discovered only after harm, pushing courts and auditors into the role regulators failed to play at launch.