Connecting state and local government leaders
By flagging unintended consequences of changes to tax policies, AI researchers hope to shut down tax shelters used by sophisticated tax dodgers.
The United States stands to lose about $7 trillion in tax revenue because of the tax gap—the difference between taxes owed and collected, according to the Treasury Department. A team of researchers at Johns Hopkins University is looking at using artificial intelligence to close tax loopholes before they ever open.
Their solution is called Shelter Check—as in tax shelter, or a way to minimize taxes. It’s very much a work in progress, the researchers are quick to say, but the idea is to build a system that could read proposed changes to tax law, whether it’s federal, state or local, and return feedback on potentially negative and unintended consequences.
“Ideally, we would want some sort of system where people, especially people who draft legislation … could consult some kind of computer [and tell it], ‘OK, I want to enact the following legislation. This is what it will contain,’” explained Nils Holzenberger, associate professor in computer science at Institut Polytechnique de Paris, who’s working on Shelter Check. “The computer should be able to come up with ways that this legislation would be used” and the expected consequences, he said. But then Shelter Check would also be able to find unexpected, negative consequences, like, creating a tax shelter if it were combined with other parts of tax law.
The team turned to AI for two main reasons, Holzenberger said. One is the maturity of natural language processing. “NLP is now starting to get to the level where it’s able to process legal language And in fact, there’s been a real surge in interest in processing legal language,” he said.
The second reason is that even if Shelter Check is not as good as a human law clerk, tax lawyer or judge, it can still work faster and process more data than humans could in a lifetime, he said. “The federal legislation or even state legislation has a large corpus of text, and so even if we have very clever people, they’d never be able to process all of it at once,” Holzenberger said.
He and Benjamin Van Durme, an associate professor at Johns Hopkins’ Department of Computer Science, and Andrew Blair-Stanek, a tax attorney turned law professor at the University of Maryland, began looking into how AI can help with tax minimization in 2020. In a paper published in December 2022, they said that accepted opinion is that the best approach to using AI to identify tax minimization is from the bottom up—by feeding large quantities of data into machine learning models to find patterns.
“But bottom-up attempts rarely make any use of the actual text of tax law authorities like the [Internal Revenue Code], Treasury regulations, IRS revenue rulings and case law,” the paper stated. “When tax law authorities are used, they are often simplified and hand-coded by humans into the models, which would be prohibitively expensive to do for all tax law authorities.”
That’s why the researchers are flipping the bottom-up approach and aiming to attack from the top down with Shelter Check. Models would ingest the raw text of all available tax law authorities and proposed new ones to extract their meaning. Then other models “would try different combinations of facts and tax law authorities to identify tax minimization strategies.”
OpenAI’s recent release of ChatGPT-4, an AI-powered chatbot, holds some of the greatest promise so far for making Shelter Check a reality, Blair-Stanek said.
“GPT-4 is just a very fantastic technology that is much better at reasoning with legal statutes than GPT-3 was,” he said. “There really is so much legal text out there…. The people who work for state and local revenue authorities and state legislatures don’t have the time to go and think through, ‘Hey, is there any way that somebody could come up with a structure taking advantage of an aspect of our tax laws and how it interacts with [another state or federal] tax law?’”
Besides the volume and complexity of tax law, another challenge to building Shelter Check is the need for human-computer interaction (HCI) to constantly improve the system, Van Durme added.
“There’s a lot of interest now in trying to steer the AI to particularly interesting challenge problems and then collect a lot of feedback in order to further adapt that model for that task,” he said. “As we move into more and more realistic versions of Shelter Check from what was just more like a science project … it’s going to require a good engagement with real professionals. So, part of our work would be not just straight, traditional artificial intelligence research, but also what’s known as HCI, human computer interaction.”
Although this type of AI-based system can be used for other types of law, Blair-Stanek said tax law is a great candidate for it because everything boils down to a single number. Other legal cases that involve, say, someone disputing a citation or the suspension of a business license, are better suited for human brains.
With taxes, the goal is to make sure people don't take advantage of the complexity of the tax code to pay less than they owe, he said. “Computers have long been better at that than at dealing with squishy concepts like, ‘Well, should we get this business license revoked or suspended?’”
Stephanie Kanowitz is a freelance writer based in northern Virginia.
NEXT STORY: EVs Rev Up Cybersecurity Challenges