Poll: Majority rejects AI fraud detection if eligible people mistakenly lose out

(The Center Square) – A majority of voters oppose using artificial intelligence to hunt fraud in government programs if it also risks wrongly flagging or denying benefits to people who qualify for them, a new The Center Square poll found.

The Center Square Voters’ Voice Poll found 53% of registered voters oppose using AI to detect fraud in government programs if it risked wrongly denying eligible people, against 37% who support it and 10% who weren’t sure. Federal agencies estimated $186 billion in improper payments in fiscal 2025, according to the Government Accountability Office, which has separately estimated the government loses $233 billion to $521 billion a year to fraud.

The Centers for Medicare and Medicaid Services screened 1 billion payments through the Treasury Department’s Do Not Pay system in fiscal 2024 and prevented 363,386 payments worth $2.6 billion, according to the Congressional Research Service.

Improper payments and fraud are not the same thing. An improper payment is one that should not have been made or was made in the wrong amount, and agencies report an estimate every year. According to GAO, most are overpayments, and most stem from errors or missing paperwork rather than theft. Fraud requires intent to deceive, and GAO counts a loss as fraud only when a court has determined it was fraudulent. That makes fraud harder to measure, which is why GAO’s estimate is a range.

In August, when The Center Square asked voters about using AI to detect fraud, waste and abuse in government programs such as Medicaid, without mentioning any risk to eligible beneficiaries, 60% supported it and 27% opposed. The two questions are not directly comparable. The October version dropped waste and abuse, dropped the Medicaid example and added the risk of wrongly denied benefits.

Mike Noble, founder of Noble Predictive Insights, which conducted both surveys, said the October question was written as a deliberate follow-up to test whether support held up once voters were told what could go wrong.

“Voters generally like the idea of using AI to root out government fraud, but that support isn’t a blank check,” Noble told The Center Square. “When you introduce the possibility of innocent people being wrongly denied benefits, public opinion shifts dramatically. What’s particularly interesting is that we see this pattern across nearly every demographic group, suggesting concerns about accuracy and unintended consequences extend well beyond partisan politics.”

Republicans and Republican-leaning independents were the only group where support led, 51% to 39%. Democrats and Democratic leaners opposed it 66% to 26%. True independents, voters who lean toward neither party, were the most opposed, 59% to 22%. Men split 44% support to 48% oppose, while women opposed it 58% to 30%.

Federal auditors have found the same tradeoff inside a real program. In a June report on the Department of Veterans Affairs Beneficiary Travel Program, which reimburses veterans and caregivers for travel to medical care, GAO wrote that tightening controls “could help further mitigate fraud risks but may also create additional barriers for legitimate beneficiaries by further reducing the timeliness of claims and may not be cost effective.” VA spent more than $2.3 billion on the program in fiscal 2025 and reported about $1 billion in cumulative estimated improper payments from fiscal 2018 through 2024.

VA has never set a fraud risk tolerance for the program, GAO found, leaving it unable to determine whether its controls appropriately balance fraud reduction against the program’s goals. VA concurred with all nine of GAO’s recommendations.

Dylan Hedtler-Gaudette, director of government affairs at the Project On Government Oversight, said automating fraud detection carries a cost that has nothing to do with whether the technology works.

“The more you start to shift things over to algorithms or to AI agents, the more you’re taking humans out of the loop,” he told The Center Square. “The more you’re allowing humans to sort of escape the kind of accountability we ultimately need in the event that something bad does happen.”

Hedtler-Gaudette said false positives are a reason to keep people involved, not a reason to abandon the tools. AI is not good at recognizing that a false positive is a false positive rather than just another piece of data, he said, or at understanding why the difference matters.

More from this section

Trump administration sues to block Alaska election results

Scientists discover how tumors play dead to dodge chemotherapy

American politicians talk about persecuted Christians abroad – but here’s what happens when those Christians migrate to the US

“We do have to build the systems here in a way that recognizes that there can be bad data, or problematic data, in either direction,” he said. “It can point you in the right direction, but it can also point you in the wrong direction because of human error or because of technological error.”

The poll also asked who should bear the most responsibility when fraud occurs in a taxpayer-funded program. Forty-five percent said the individuals or organizations that committed it. Twenty-two percent said the elected officials who fund or manage the program and 15% said the bureaucrats overseeing it. Nine percent said outside auditors or investigators and 9% weren’t sure. The pattern was nearly identical across party lines.

Hedtler-Gaudette said the conditions that allow fraud are set by the people who build and oversee the programs.

“The environment in which this kind of fraud happens, or can happen, is entirely driven by and created by the policymakers, the members of Congress, the personnel in various federal agencies and state agencies,” he said. “That’s who creates the systems and that’s who’s supposed to be overseeing the system as well, and keeping tabs on things to make sure there isn’t fraud.”

No system reaches zero, he said, but he landed where most voters did.

“There’s blame to go around here,” he said. “But ultimately, the individual who perpetrates the fraud is the one responsible.”

Voters were also asked how important it is that nonprofits receiving taxpayer money face public reporting and accountability requirements. Eighty-four percent said important, including 59% who said very important. Nine percent said not important. It was the only question of the three with no partisan split, drawing 88% support among Republicans and Republican leaners and 83% among Democrats and Democratic leaners.

Government grants to nonprofits roughly doubled between 2016 and 2022, to about $319 billion, and the number of nonprofits receiving them roughly doubled as well, to more than 100,000, according to an ongoing The Center Square investigation of nonprofit tax filings. The investigation found that at least a fifth of nonprofits taking government grants had executive salaries consuming a quarter of the grant money they received.

Nonprofits that receive federal money already face disclosure requirements. Under federal grant rules known as Uniform Guidance, those that expend $1 million or more in federal awards in a year, a threshold that also covers state and local governments and universities, must undergo a single audit, an independent review of how the money was spent. The results go to the Federal Audit Clearinghouse, the government’s repository of record, and auditees must make copies available for public inspection.

Most tax-exempt organizations large enough to take significant federal funding also file a Form 990 each year, reporting revenue, executive compensation and the government grants they received. The IRS exempts churches from the filing requirement, and organizations with gross receipts of $50,000 or less file a short electronic notice instead.

The White House and the Office of Management and Budget did not respond to questions about the poll’s findings. The Council of the Inspectors General on Integrity and Efficiency and the National Council of Nonprofits did not respond to requests for comment; the Council of Nonprofits also declined to make anyone available for an earlier The Center Square story on government grants to nonprofits in August.

The Center Square Voters’ Voice Poll was conducted by Noble Predictive Insights from Oct. 5-7, surveying 2,527 registered U.S. voters nationally through an opt-in online panel and text-to-web cell phone messages. Respondents included 1,136 Republicans, 1,171 Democrats and 220 true independents. The margin of error is plus or minus 1.95 percentage points, and higher for subgroups.

Voters go to the polls Nov. 3.