:focal(3499x2741:3500x2742))
Accelerating Solutions:The most wicked problems
How Daniel Ho is nudging government into the 21st century.
)
When Stanford’s Regulation, Evaluation, and Governance Lab started working with the IRS in 2019, the goal was straightforward if a bit overwhelming: Help close the staggering $500 billion gap between taxes owed and taxes paid in this country. That shortfall saps the government’s ability to pay for vital programs and services, from Social Security to food stamps to national defense.
If “taxes are what we pay for civilized society,” as Supreme Court Justice Oliver Wendell Holmes once put it, civilized society has been getting stiffed.
With an array of new machine learning techniques, RegLab set about assessing how the Internal Revenue Service identifies tax evaders—an audit of audits, in effect. Naturally the formula for selecting audits is a closely held secret within the agency.
“They treat their selection systems the way that Coca-Cola treats the recipe for Coke,” jokes Daniel E. Ho, RegLab’s director and the William Benjamin Scott and Luna M. Scott Professor of Law at Stanford Law School, professor of political science and computer science (by courtesy), and senior fellow at the Stanford Institute for Human-Centered Artificial Intelligence (HAI) and the Stanford Institute for Economic Policy Research. “Only three people are allowed to know, and they’re not allowed to fly on a plane together.”
But peering into opaque bureaucracies is nothing new for the team. RegLab regularly partners with government agencies on a pro bono basis to conduct high-impact demonstration projects; using machine learning and data science, the ultimate goal is to modernize how these agencies do business.
“We can figure out how to do this better,” says Daniel Ho. “That is what motivates this core group of individuals who work on these kinds of problems within the RegLab.”
The work of RegLab is now advancing with support from Stanford Impact Labs, which awarded a Stage 3 investment over four years to support the lab’s solutions-oriented, partnership-based work. Stage 3 funding is designed to supercharge teams that have a demonstrated track record of successful partnership-based R&D cycles and are ready to take their work to the next level.
When Ho and his research team started working with the IRS, they knew they had to build in safeguards alongside any machine learning they introduced to the tax administration process. But before that could happen, they needed a detailed understanding of the existing auditing system. Going over reams of microdata—they reviewed roughly 148 million tax returns and 780,000 audits—the team discovered something stunning: Black taxpayers are three to five times more likely to be audited than non-Black taxpayers.
The discovery, Ho says, was “jaw-dropping.” The IRS has long been suspected of uneven tax enforcement across racial groups. Indeed, “scholars like Dorothy Brown had raised these questions for years,” Ho adds, citing a professor at Emory University known for her work on the racial implications of federal tax policy. RegLab’s study provided the most direct evidence to date of how ostensibly neutral processes can nevertheless perpetuate inequality.
In January 2023, the team released a paper, “Measuring and Mitigating Racial Disparities in Tax Audits.” A number of prominent publications covered the report, from the New York Times to USA Today, and soon lawmakers were demanding answers from the IRS. “This latest working paper should put to bed any question about the existence of a problem with racial inequities in audits,” Senator Elizabeth Warren wrote to the heads of the IRS and the Treasury Department.
)
From the desk of Elizabeth Warren
)
On April 13, 2023, Senator Elizabeth Warren wrote to the heads of the IRS and the Treasury Department, highlighting RegLab’s findings. Read the full letter in PDF format.
As the study noted, the auditing disparity wasn’t deliberate; the data the agency uses for predicting tax evasion is race-blind. But by disproportionately focusing on mistakes in tax breaks like the earned income tax credit—which African-Americans claim at higher rates—the system had effectively ingested racial disparities into its audit selection process.
“There are people who live paycheck to paycheck, whose refunds are delayed by manual data entry of returns filed on paper. It’s a system that’s ripe for modernization,” Ho says.
In response to the findings, the IRS announced in September 2023 that it was making significant changes to its audit selection system. Its approach to evaluating the returns of lower-income Americans would be revamped—particularly how it audits recipients of the earned income tax credit.
“We are making broad efforts to overhaul compliance efforts in a manner that robustly advances our commitment to fair, equitable, and effective tax administration,” IRS commissioner Daniel I. Werfel wrote in a letter to Senator Ron Wyden, chairman of the Committee on Finance, describing the agency’s response to the RegLab study.
For Ho, the discovery didn’t just right a historic wrong. It flipped a truism about machine learning on its head.
“Often there’s this narrative of algorithmic bias. Here, it was actually the turn toward machine learning that led to us discovering these disparities in existing legacy systems,” he says. In the end, machine learning is what led to these dramatic findings.
RegLab’s work with the IRS continues, driven by the team’s itch for measurable impact.
“We have these incredible seven schools of peopleat the top of their fields, and a kind of generational shift of people who are wanting to do more engaged work. They’re not satisfied just like, ‘Oh, I’m going to publish this paper.’ This is a generation of people who want to use that skill set to solve some of the most wicked problems. It’s that passion for impact.”Daniel Ho
:focal(2990x4112:2991x4113))
Some of those wicked problems are local. In the summer of 2020, walloped by COVID-19, Santa Clara County found itself struggling to serve its nearly 2 million residents. RegLab partnered with the county’s Public Health Department to make a variety of improvements in how vital public health resources were delivered. Where partnerships with academia can prove cumbersome—think long, elaborate studies—the team understood the urgency of the moment. Zeroing in on a host of ways the county’s pandemic response could be modernized, RegLab helped usher in improvements everywhere from contact tracing to testing to vaccine delivery. The partnerships led to several published studies and an innovation award from the National Association of County and City Health Officials—as well as greater preparedness for future emergencies.
Making a difference in the real world is hardly a novel goal within the academy. But in practice, Ho says, substantive change is elusive.
The secret to RegLab’s success? A true cross-discipline team, with zero interest in business as usual.
“None of us could do this alone,” Ho says. “It’s really the fact that you have the hydrologist working together with the machine learner, working together with the social scientist and statistician, that makes this kind of stuff work.”
Other RegLab efforts underway
Unemployment insurance
Through a partnership with the U.S. Department of Labor, RegLab is prototyping methods to modernize the administration of unemployment insurance. Such modernization is acutely needed. While 97 percent of unemployment insurance payments were paid on a timely basis before COVID-19, that number dropped to just above 50 percent, right when people needed benefits the most. Some 46 million Americans relied on the system during the pandemic. The partnership aims to build a path so that the system will not collapse during a future economic crisis.
Environmental sustainability
In 2020, 60-75 percent of major facilities permitted under the Clean Water Act self-reported being in noncompliance. As a result, tens of millions of Americans are exposed to pollution hot spots and unsafe drinking water. RegLab crafted partnerships with federal and state Environmental Protection Agencies to develop data-driven compliance systems, including the use of remote sensing and computer vision to develop an environmental monitoring system of the future.
Accelerating Solutions: The impact
Why it matters
RegLab engages directly with public agencies that make decisions affecting billions of dollars and millions of people. From public health to the environment to inequality, the impact of RegLab’s work is both broad and deep.
The opportunity
On-campus accelerators for learning, health, and social impact will positively impact the widest populations with the greatest needs. The accelerator model opens new research and development pipelines, simultaneously unlocking pathways for Stanford faculty and students to create impact and cultivating external partnerships to design and scale solutions faster.
)
:focal(790x561:791x562))
:focal(1918x2231:1919x2232))