States across the nation are deploying artificial intelligence to blunt the impact of the Trump administration’s overhaul of the country’s food stamp program, which serves more than 36 million Americans.
Michigan, Nevada, and West Virginia are among the states collectively pouring millions of dollars into novel AI-powered technology that analyzes case files and flags potential errors in the Supplemental Nutrition Assistance Program.
More than 5 million people have already lost access to nutrition benefits since President Donald Trump’s One Big Beautiful Bill Act narrowed eligibility for SNAP and added work requirements. Starting next year, the legislation also requires states to cover more of their SNAP costs if their error rates exceed 6%.
The bill penalizes states if they don’t reduce their mistaken payments, despite a lack of evidence that SNAP fraud is a problem. The nonpartisan Center on Budget and Policy Priorities predicts it will ultimately burden states with hundreds of millions — or even billions — of dollars in costs and could force some to cut the program.
In an effort to keep error rates low, Michigan launched an AI tool earlier this year called the Michigan Case Accuracy Review Assistant, or MiCARA. The tool identifies potential errors and inconsistencies — such as if a household doesn’t report a temporary increase in income — using both AI and rules-based indicators.
“This is a massive improvement,” David Knezek, chief operating officer of the Michigan Department of Health and Human Services, told a state Senate subcommittee in March, as he described how the state’s new AI tool had completed 17,500 pre-certification reviews in three months, identifying $150,000 in potential payment errors before applications were approved.
Michigan officials have not said how often the tool incorrectly identifies errors. In a statement to Proof News, Erin Stover, a spokesperson for the state Department of Health and Human Services, emphasized that the tool does not approve or deny SNAP benefits.
“The eligibility specialist, not the technology, remains responsible for every eligibility decision,” Stover said.
The state’s assurances have not assuaged the ACLU of Michigan.
“We continue to have concerns about how the use of this tool could disparately impact vulnerable populations who receive SNAP benefits, especially groups such as the elderly, people with disabilities, and people whose native language is not English,” Phil Mayor, the ACLU of Michigan’s deputy legal director, said in a statement to Proof.
Even with a human hand in decision making, Mayor pointed out, “the technology can cause automation bias that affects downstream human decisions.”
Past automation failures
Critics of AI use in public benefits point to past automation failures. In 2013, Michigan’s unemployment agency replaced a 30-year-old fraud detection system with a new, automated system that promised to improve customer service, increase data accuracy, and reduce operating costs.
But the system was a disaster: During its first two years, the fraud detection tool — the Michigan Integrated Data Automated System or MiDAS — wrongfully accused approximately 40,000 people of unemployment fraud, resulting in thousands of improper denials and people’s wages being automatically garnished for fraud that the system alleged occurred years earlier. A 2016 state review of 22,000 MiDAS fraud determinations revealed that 93% did not actually involve fraud. After being falsely accused, an estimated 1,100 Michiganders filed for bankruptcy.
“If I dwell on it, I could go crazy. It’s like being robbed,” Jason Doss, a Detroit-area worker, told the newsroom Bridge Michigan after he was wrongly accused of fraud.
Following lawsuits, the state overturned many of the fraud determinations and repaid residents millions of dollars. In 2017, Wanda Stokes, the former director of the now-abolished state Talent Investment Agency, apologized for the crisis, telling The Associated Press: “I really feel horrible about this whole situation ... It shouldn’t have happened.”
Tools can ‘go wrong in really significant ways’
Other state leaders are betting on AI. Nevada has spent about $1 million on SAS Payment Integrity for Food Assistance, a tool that helps case managers identify potential SNAP payment errors, according to Kristle Muessle, a spokesperson for the state’s department of human services.
Implemented in March, Nevada’s tool uses AI to look at the entire SNAP caseload and determine which cases are most likely to have errors in them, assigning the potential errors a risk score, according to Kelly Cantrelle, deputy administrator for Nevada’s Division of Social Services.
“Before, we were searching for it, looking for it, hoping to find it, but now we have a report that simply says, ‘start here,’” Cantrelle said in a recent promotional video for SAS, explaining how the tool assists Nevada caseworkers. “We’re able to start at the highest risk score cases and work our way down from there.”
In June, SAS reported that the tool helped Nevada caseworkers identify $130,000 in monthly overpayments in its first two months of deployment. The findings have not resulted in any cases being erroneously denied or benefits being delayed, Muessle told Proof.
In West Virginia, state leaders reportedly spent nearly $900,000 on the same tech.
West Virginia's Department of Human Services received approval to make an emergency purchase of the software in March, according to documents obtained by West Virginia Watch. The department is implementing the tool, according to human services department spokesperson Angelica Hightower, and it is expected to be fully operational by early 2027.
Spokespeople for both Nevada and West Virginia emphasized that the SAS tool is designed to support caseworkers, who will continue to make all eligibility and benefit decisions.
“By helping identify cases that may need attention and prioritizing reviews, it can reduce the burden of manually reviewing cases,” Hightower said in a statement to Proof. “This also supports greater accuracy and faster service for clients.”
In August, members of the Florida Democratic congressional delegation sent a letter to Gov. Ron DeSantis and Kathryn Williams, deputy secretary of the state Department of Children and Families, demanding more information about the “alarming mandate” in the state budget to hire a vendor capable of using “machine learning” to help reduce Florida’s SNAP error rate.
“We understand that the state must act urgently to address the impending budgetary crisis created by the [One Big Beautiful Bill Act],” wrote US Rep. Maxwell Alejandro Frost and six other members of Congress. “However, we are concerned with one of the state’s solutions: The appropriation of $4 million in state funds for the new and relatively unproven technology of AI to identify SNAP errors.”
Florida officials did not respond to Proof’s request for comment.
Kevin De Liban, founder of TechTonic Justice, a nonprofit organization advocating for low-income communities fighting AI harm, cautioned that government technology projects often “go wrong in really significant ways,” harming residents who rely on public assistance.
Michele Gilman, a University of Baltimore law professor and an expert in the automation of public benefits, said the problem is not just the automation but the fact that it is nearly impossible for a recipient to argue against an AI-made decision.
“Fighting adverse results is very challenging because of the black-box nature of these systems,” she said. “It’s very hard to cross-examine an algorithm.”
