When Congress authorized over $5 trillion in pandemic-era relief programs, and directed agencies to prioritize speed above all else, fraudsters cashed in with bogus claims.

But data from these pandemic-era relief programs is now being used to train artificial intelligence-powered tools meant to detect fraud before payments go out.

The Pandemic Response Accountability Committee has developed an AI-enabled “fraud prevention engine,” trained on over 5 million applications for pandemic-era relief programs, that can review 20,000 applications for federal funds per second, and can flag anomalies in the data before payment.

The PRAC’s executive director, Ken Dieffenbach, told members of the House Oversight and Government Reform Committee on Tuesday that, had the fraud prevention engine been available at the onset of the pandemic, it would have flagged “at least tens of billions of dollars” in fraudulent claims.

Dieffenbach said that the PRAC’s data analytics capabilities can serve as an “early warning system” when organized, transnational criminals target federal benefits programs. He said the PRAC is working with agency inspectors general on ways to prevent fraud in programs funded by the One Big Beautiful Bill Act, as well as track fraudsters targeting multiple agencies.

“Fraudsters rarely target just one government program. They exploit vulnerabilities wherever they exist,” Dieffenbach said.

The PRAC’s analytics systems have recovered over $500 million in taxpayer funds. Created at the onset of the COVID-19 pandemic, the PRAC oversaw over $5 trillion in relief spending.  It was scheduled to disband last year, but the One Big Beautiful Bill Act reauthorized the PRAC through 2034.

Government Operations Subcommittee Chairman Pete Sessions (R-Texas) said the PRAC has developed data analytics capabilities that can comb through billions of records, and that these tools need a “permanent” home once the PRAC disbands.

“A permanent solution that maintains the analytic capacities and capabilities that have been built over the past six years is necessary and needed. Its database is billions of records deep, and it has begun to pay for itself,” Sessions said.

In one pandemic fraud case, the PRAC identified a scheme where 100 applicants filed 450 applications across 24 states, and obtained $2.6 million in pandemic loans. Dieffenbach said there are tens of thousands of cases like it.

“This is but one example where the proactive use of data and technology could have prevented or aided in the early detection of a scheme, mitigated the need for a resource-intensive investigation and prosecution, and helped ensure taxpayer dollars went to the intended recipients and not the fraudsters,” Dieffenbach said.

In 2024, the Government Accountability Office estimated that the federal government loses $233 to $521 billion in fraud every year.

Sterling Thomas, GAO’s chief scientist, said AI tools are showing promise in flagging fraud, but he warned that “rapid deployment without thoughtful design has already led to unintended outcomes.”

“In data science, we often say garbage in, garbage out. Nowhere is that more true than with AI and machine learning. If we start trying to identify fraud and improper payments with flawed data, we’re going to get poor results,” Thomas said.

The Treasury Department often serves as the last line of defense against fraud, but it is giving agencies access to more of its data to flag potential fraud before issuing payments.

Under a March executive order, President Donald Trump directed the Treasury Department to share its own fraud prevention database, Do Not Pay, with other agencies to the “greatest extent permitted by law.”

Renata Miskell, the deputy assistant secretary for accounting policy and financial transparency at the Treasury Department’s Bureau of the Fiscal Service, told lawmakers that only 4% of federal programs could access all of Do Not Pay’s data in fiscal 2014. But by the end of this fiscal year, she said all federal programs are on track to fully utilize Do Not Pay.

“We want every program — and there’s thousands of federal programs —  to use Do Not Pay before making award and eligibility determinations,” Miskell said.

To make Do Not Pay a more effective tool against fraud, Miskell said Treasury is looking for the ability to “ping” other authoritative federal databases, such as the taxpayer identification numbers (TINs) issued by the IRS or Social Security numbers, before issuing a payment. Without those datasets, she said, Treasury is following a “trust but verify” approach to payments, doing some basic checks before federal funds go out.

“These data sources would dramatically improve eligibility determination and fraud prevention,” Miskell said.

The post Pandemic watchdog builds AI fraud prevention ‘engine’ trained on millions of COVID program claims first appeared on Federal News Network.