How generative AI can help fight fraud and waste in the public sector

In March 2023, the White House announced a plan to request $1.6 billion in funding from Congress to tackle the historic levels of fraud linked to COVID-19 relief programs. It’s a high price, but the program aims to chase after an even more mind-boggling $5 billion in fraudulently claimed payments.

During the pandemic, the federal government disbursed trillions in aid to businesses struggling to stay open, assistance for local governments of all sizes, unemployment support, and direct payments to individuals. The largest payment in US history, aimed at keeping the economy afloat during an unprecedented emergency, quickly overwhelmed any watchdog agencies tasked with oversight.

Striking the right balance between providing much-needed relief and preventing grift and waste will always be a challenge. Over the past 20 or so years, the normal level of improper payment rates for employment insurance has varied from 9% to 13%. According to a report from the Department of Labor, that rate doubled to 21% during the pandemic. Times of uncertainty, unfortunately, will always attract opportunists.

So far, 1,500 people have already been charged with defrauding pandemic aid programs, and more than 450 have been convicted. That is only a dent for the 50 agents in the small business administration office sifting through two million flagged loan applications.

Where AI Fits in Fraud Detection

In 2010, less than 1% of government fraud was uncovered through tech-based oversight, with governments relying on whistleblowers, agency spot checks, and audits and reviews to launch investigations into misuse, fraud, and waste. A lot has changed since then.

Advanced algorithms and machine learning can quickly spot abnormalities, anomalies, or suspicious transactions in a mountain of data and refer it to human investigators for deeper investigation. AI-based tools are already a useful asset in fighting fraud in the public sector, banking, and healthcare.

But using these tools require training. To employ generative AI in fraud detection, you first need to train the model using both legal and illegal transactions so that it can learn the patterns behind normal and suspicious activities. However, fraud frequently involves sensitive data, and the rules around using sensitive data to train models can be unclear and even bump against ethics. Additionally, datasets need enough suspicious data to train the model. Many machine learning tools applied to fraud detection rely on a strong fraud signal, which in real-world scenarios would typically be lower than 0.5% of the data. In an ideal exercise, the data used to train any AI model would contain a 50/50 mix of fraud/non-fraud samples.

How Generative AI Can Prevent Fraud

This is where generative AI can help provide a solution.

• General Adversarial Networks (GANs) contain two main components: a generator and a discriminator. The generator ‘generates’ data and content, while the discriminator evaluates the quality of data by comparing it to real examples. GANs can be used in fraud detection by creating realistic synthetic data to train generative models and learn from both legal and illegal transactions.
  
  This approach can help detect fraudulent activities more accurately and quickly, as GANs can learn patterns from a high volume of transactions that traditional models may miss. Additionally, synthetic data can help address bias in fraud detection. By creating balanced datasets that represent theoretical citizen profiles and transaction patterns, fraud detection systems can control for other factors and avoid using unrelated patterns, improving accuracy and fairness.
  
  
• Natural Language Processing, which combines computational linguistics with statistical models, is another helpful generative tool. The Innovation Lab at the Government Accountability Office (GAO) is using Natural Language Processing and topic modeling to sift through large volumes of public comments, social media posts, regulatory exemption applications, and other sources, and organizing this information into relevant fraud schemes and taxonomy. By seeking out patterns in how fraudulent claims are written, NLP can help root out those claims faster.

Prevention is Better Than Cure

Preventing fraud is cheaper and more efficient than the hours and resources required to investigate and prosecute after the fact. The Department of the Treasury’s Financial Crimes Enforcement Network AI - examines financial data to find abnormalities and suspicious behavior that could signify money laundering, turning over suspicious cases to human investigators. The Internal Revenue Service and Securities and Exchanges Commission both have similar frontline AI detection tools that flag suspicious activity. The Centers for Medicare and Medicaid estimate that their Fraud Prevention Service algorithm has, over four years, helped to prevent or uncover nearly $1.5 billion in potentially fraudulent payments.

In the early parts of the pandemic, Paycheck Protection Program loans were not cross-checked against “Do Not Pay” databases, which would have likely caught 57,000 loans worth $3.6 billion by August 2020. Some of the fraud that occurred during the pandemic could have been caught or prevented by using the protections already in place.

It’s easy to point fingers after the fact. But it’s important to remember that during the unprecedented crisis of the pandemic, government agencies needed to prioritize getting loans and payments out the door to people in need as quickly as possible, using the tools available. Many state systems designed to fight fraud were not prepared or designed to facilitate the level of data analysis necessary. The best we can do now is learn and prepare better for next time.

“Criminal syndicates are going to look for weak links at moments of crisis to attack us,” said Gene Sperling, the White House coordinator for pandemic aid. “The right time to start building a stronger system to prevent identity theft is now, not in the middle of the next serious crisis.”

If you’re considering adding AI to your toolbox, you don’t have to do it alone. TTEC Digital has over 40 years of experience designing and implementing systems that will help you increase operational efficiency, improve citizen experience, and ensure you remain compliant with regulations.