A new AI model trained on COVID-19 loan applications could have flagged potentially tens of billions of dollars in fraud before payouts, according to the US government’s Pandemic Response Accountability Committee (PRAC). The model, which draws insights from nearly five million Economic Injury Disaster Loan applications, was developed after the pandemic began, meaning its preventative potential was untapped during the peak of the COVID crisis.

PRAC’s executive director, Ken Dieffenbach, emphasized to the House Oversight Committee the importance of applying lessons learned from the pandemic to combat fraud in government programs. He stated that the AI, known as the Fraud Prevention Engine, is capable of processing 20,000 applications per second, potentially flagging large volumes of fraudulent activities.

The Engine employs a combination of unsupervised and supervised machine learning models to detect patterns associated with fraud. Simple anomalies, such as shared bank account numbers among applicants, are leveraged to uncover potential deceit. Since its development, the Fraud Prevention Engine has supported the recovery of over $500 million, a fraction of what the AI could potentially flag.

Looking forward, PRAC aims to expand the AI’s use beyond pandemic-related oversight. The committee’s mandate has been extended to 2034, with $88 million in funding to support its enhanced oversight of programs facilitated by recent budget reconciliations.

Pete Sessions, a member of the Oversight Committee, stressed the need for a permanent home for the AI system to secure its benefits beyond PRAC’s current timeline, recognizing its extensive database of fraud-related records as a significant asset.