AI-Based Tools Help Treasury Fight Fraud, GAO Report Says

Fraud and improper payments are distinct but interrelated problems that have challenged the government for decades. Advances in AI and data analytics technologies now make agencies more capable of detecting and stopping malfeasance, the Government Accountability Office (GAO) reported.

A recent GAO report outlined areas where federal agencies can improve using AI to detect fraud and how they can attract, train and retain an AI-skilled workforce. It also cited examples of agencies that are already making use of these capabilities, highlighting the Treasury Department’s Do Not Pay system for effectively detecting and stopping fraudulent and improper payments.

Fighting financial misconduct is an ongoing effort across the federal government. The GAO estimated that from 2018 to 2022, the federal government lost between $233 billion to $521 billion annually due to fraud. Since the 2003 fiscal year, the cumulative cost of improper payments are estimated to be about $2.8 trillion, the GAO reported.

AI and Analytics at Treasury

The Treasury Department’s Do Not Pay system consolidates data on entities ineligible to receive government payments. It is a data matching service that agencies can use to prevent payments to ineligible individuals such as deceased persons. For example, the GAO noted that the Social Security Administration (SSA) shares its full death data with Do Not Pay to support this.

Do Not Pay is a unique capability, Sterling Thomas, chief scientist for science, technology and analytics at GAO told GovCIO Media & Research. While it uses machine learning to match data and share information with agencies, he emphasized that it is an application, not a database.

Sharing this data has yielded positive results. The Treasury Department reported that in its first year of temporary access to the SSA’s full death data resulted in cost savings of $109 million, with projected benefits of at least $330 million projected over the program’s lifespan.

Thomas said the GAO recommends that Do Not Pay be expanded to include all anti-fraud efforts across the government to serve as a clearinghouse. This would allow it to share new and interesting agency analytics with the entire federal government.

A Trained Workforce for AI Use

The key benefit of AI and data analytics tools is that they can help agencies sift through large volumes of data and spot patterns faster and more efficiently than human analysts. But the report cautioned that “agencies need solid, reliable data to avoid false positives.” Additionally, while these systems can work faster than humans, skilled humans in the loop are essential.

“Because of how AI tools operate, particularly in a fraud space, you need someone who is knowledgeable about fraud and about AI and the analytics around it so that they know it is behaving in the way that it should so that the results are good. We don’t want to accuse somebody of fraud who didn’t do it,” Thomas said.

The report recommends agencies follow its AI Accountability Framework for Federal Agencies and Other Entities. This document contains best practices to ensure AI systems use data that is high quality, reliable and appropriate for the intended purpose.