October 8, 2018; The Intercept
Can a computer system have bias?
Of course it can; like any system conceived of by humans, it replicates the assumptions of its creators. Studies have shown that facial recognition algorithms struggle to recognize people of color as human, replicating the bias of their white creators. The algorithm created by the US Department of Agriculture to weed out fraud among SNAP recipients is no different, and the bureaucratic system within which the algorithm operates replicates, rather than repairs, the problem.
What the SNAP alert system struggles to recognize is patterns of fraud, or instances in which people trade food stamps for cash. Since SNAP started using a debit system instead of a paper one, the Food and Nutrition Service (FNS) can use an algorithm, rather than undercover investigators, to identify fraud. The algorithm does this by recognizing patterns: Very large single purchases, lots of purchases for identical amounts, and multiple visits to the same vendor in a day are among the activities that cause the algorithm to flag a vendor.
Unfortunately, flags are sometimes raised by normal buying patterns in urban communities with high poverty rates. It might be unusual for a suburban family to visit a grocery store multiple times in a day, but it’s not uncommon to go to the corner store for milk in the morning and a snack in the afternoon. People with a steady cash flow might be able to pay for their groceries when they’re needed, but SNAP cards are only loaded once per month. Some grocers, like Porfirio Mejia, who runs P&L Deli Grocery in Manhattan, allow customers to buy things on credit, then charge a week or two of groceries at once when the new balance arrives.
Mejia, like over 1,000 other retailers, was disqualified from participating in SNAP, resulting in the loss of at least 35 percent of his business. His appeal was denied because he didn’t know he would need itemized receipts showing that the large balances were in fact eligible food items.
Unlike some other situations we’ve experienced recently, in cases of SNAP fraud, there is no presumption of innocence; the burden of proof is on the accused. The Standard of Proof set out in the case of Mini Convenience Store and Deli, which also faced losing its SNAP eligibility, reads, “In an appeal of an adverse action, the Appellant bears the burden of proving, by a preponderance of the evidence, that the administrative action should be reversed.” In other words, the small-business proprietors have to prove the algorithm wrong by providing enough evidence to override the conclusion of an “unbiased” system. Unsurprisingly, success rates are low; less than two percent of claims are reversed or modified.
The thing is, like voter fraud, SNAP fraud is a comparatively minuscule problem. The USDA estimates that fraud rates are about 1.5 percent. You might not know that from the way it is talked about; the federal government has, under several presidents of both parties, spent millions of dollars to educate the public about it. Simon Constable, a fellow at the Johns Hopkins Institute for Applied Economics, Global Health, and the Study of Business Enterprise, wrote for Forbes that though fraud had “ballooned” from 2012 to 2016, it still represented less than one percent of total spending. (Forbes made the highly objectionable observation that “there are, of course, other reasons to cast doubt on the SNAP program, such as it fostering a dependency on government aid,” which is a demeaning poverty myth that NPQ and others have repeatedly debunked.) The nation’s capital ran a public anti-SNAP fraud campaign in the spring of 2018, funded by the USDA, but city residents objected to it so strongly that the Department of Human Services not only ended the campaign but issued an apology letter.
Ariel Kagan, a senior researcher at the Food Institute, said, “If we really want to make SNAP fraud and abuse go even further down—and it’s pretty much at zero—thinking about policies and programs and campaigns that actually get to the issues that are happening and not just villainizing SNAP users [would be] a good idea.”
Urban convenience stores are at the center of the SNAP fraud issue. A USDA report says, “Trafficking was much more likely among retailers located in higher poverty neighborhoods than those in areas with less poverty. Trafficking rates are also highest in the most urban areas.” Corner stores represent 45 percent of authorized SNAP retailers, 16 percent of redemptions, and over 90 percent of the fraud cases. Is this a concentration of government fraud, or a system that isn’t designed to recognized buying patterns among the urban poor?
Stewart Fried, an attorney who has represented store owners flagged by the algorithm, says, “It’s a jerry-rigged system against small retailers unlike anything I’ve ever seen before.”—Erin Rubin