In his latest budget proposal, President Joe Biden is set to unveil a set of new tax increases on wealthy Americans. Rather than raising taxes, maybe he should just focus on collecting what some of them already owe.
Each year, the IRS takes in $600 billion less than it should. By one estimate, half of that is because of underreporting by those among the super-wealthy who hide income by setting up sophisticated partnerships or other entities. If you’re anxious about the amount of US debt, those numbers should grab your attention.
The IRS just doesn’t have the resources to chase them down. Following years of budget cuts and understaffing, the IRS has mainly targeted poor families with audits because doing so is easier and cheaper than pursuing the complicated tax matters of wealthy filers. But artificial intelligence could change that balance of power, helping the archaic, beleaguered agency do a better job of going after the real money.
The Inflation Reduction Act allocates almost $80 billion to the IRS over the next decade. Once IRS commissioner nominee Daniel Werfel is confirmed, one of his first orders of business should be dedicating some of the funding to tap AI to help revamp the entire audit process.
Take businesses structured as partnerships, where audit rates have dropped to 0.05% and the average tax rate is just 16%. (The top federal income tax rate is 37%.) According to a recent paper led by economists at Stanford University, about 15% of partnerships are complicated — meaning they may build LLC upon LLC upon LLC, and so forth, and have overlapping partners.
Some efforts are underway, but it’s still very difficult for the IRS to determine if those complicated partnerships are reporting the right amount of income. And many of the agency experts specializing in this area have retired or will be retiring soon.
But looking at more than 7 million partnership entities from 2013 to 2015, the researchers found that machine learning was successful in helping to predict which entities were noncompliant — in other words, didn’t pay all that they owed in taxes. This research shows AI has the potential to peel away the layers more easily and efficiently, flagging noncompliant partnerships to human agents who could follow up.
The IRS is pretty tight-lipped about any AI or machine learning it’s currently using on the enforcement front, but during a webcast in 2018, the agency revealed that technology was helping it root out certain noncompliance in minutes. That used to take humans weeks or months.
Honest taxpayers should rejoice at this prospect. Today, too many compliant payers are burdened with unnecessary audits. It would benefit both taxpayers and the agency to stop wasting time on this painful process when it isn’t necessary. AI could recognize patterns and guide examiners to audits that pay off.
Still, there are some caveats. We’re not headed toward a future where the IRS is run by green-visor wearing robots. AI can only augment the capacity of IRS examiners, not displace them. As Janet Holtzblatt, a senior fellow at the Urban-Brookings Tax Policy Center put it: “Humans still need to be the teachers and graders.”
The Netherlands provides a good example of how relying solely on AI can introduce new problems. In 2013, the Dutch tax authorities started using a self-learning machine algorithm to check that child-care subsidies were going to the correct recipients. The algorithm suffered from an ingrained racial bias and innocent families were forced to give their credits back without appeal. (The prime minister and his entire cabinet resigned in 2021 following the scandal.)
In the US, new research shows that the IRS’s current algorithms can also be discriminatory. A new working paper shows that Black taxpayers are more likely to be audited than other taxpayers. In the most egregious example, a single Black man with dependents who claims the earned income tax credit is almost 20 times as likely to be audited as a non-Black claimant who is married and filing jointly. Yikes.
But that doesn’t mean we should give up on using technology to improve tax compliance. Daniel E. Ho, an economist at Stanford who worked on this paper as well as the one on complex partnerships, told me, “There’s this anxiety about machine learning, but it can also lead to the discovery of disparities in incumbent legacy systems.” Basically, the machine learning helped to reveal the inequity and now it’s up to the humans to fix it.
If AI is applied the right way and has proper oversight, it could go a long way toward making the IRS’s auditing fairer, better targeted and more profitable for the US government. There’s nothing Orwellian about that. It’s progress.
Alexis Leondis is a Bloomberg Opinion columnist covering personal finance. Previously, she oversaw tax coverage for Bloomberg News.
