Helping vets. Finding tax cheats and illegal rhino horns. How AI could transform government

WASHINGTON ? What if artificial intelligence could prepare the nation for disastrous weather events, root out fraud and tax cheats, speed up benefit determinations, enforce workplace safety rules ? and even find illegal rhino horns?

This cutting-edge technology also can undermine privacy, embed discrimination into decision making, erode trust and create public safety risks.

But now, the federal government has become a proving ground for whether rapidly advancing artificial intelligence – which President Joe Biden has called “the most consequential technology of our time” ? will be, or even can be, embraced by an increasingly wary public.

Biden administration officials, who want the federal government to lead by example in the responsible use of AI, say they’re aware how important it is to earn public confidence.

“AI can’t build trust. But bad AI can certainly break trust between VA and veterans,” Veterans Affairs Secretary Denis McDonough said in September. “So, we need to get these emerging AI technologies right. If we do, trustworthy AI can help VA scale our impact, improve our outcomes and speed the delivery of our care.”

Biden's new AI executive order

In October, Biden announced new actions, including proposed rules for how federal agencies can use the technology, requiring each agency put someone in charge of overseeing AI, and launching a talent search to recruit AI experts to work for the government.

“Getting technical talent into the federal workforce is the single biggest obstacle for effective regulation,” said Daniel Ho, a law professor at Stanford University who serves on the national committee advising the White House on AI policy. “Government cannot govern AI if it does not understand AI.”

Ho made that comment at one of the many congressional hearings this year on the uses of AI across the public and private sectors. Lawmakers are working on legislation to regulate the technology beyond what the administration can do through executive action.

California Rep. Jay Obernolte, the only member of Congress with an advanced degree in artificial intelligence, said in March his colleagues needed time to understand the risks of a technology with which few are familiar.

The intense focus comes as a growing share of Americans are expressing concern about the role AI is playing in daily life.

Only 1 in 10 Americans surveyed this summer by the Pew Research Center said they are more excited than concerned about the increased use of artificial intelligence. About half (52%) were more concerned than excited, a 14 percentage point increase from eight months earlier.

'Deep accountability challenges'

Researchers at Stanford University and New York University concluded in a 2020 report that AI promises to transform how government agencies function. But little was known at the time about how the federal government was using AI beyond a few headline grabbing examples or surface-level descriptions.

And while use was widespread, it was not sophisticated and posed “deep accountability challenges,” researchers warned.

For example, although the law generally requires the government to give a reason for denying a benefit, such as disability assistance, decisions made by many of the more advanced AI tools are not, by their structure, fully explainable.

“Even the engineers who design them do not always understand how they reach the conclusions they reach,” Michigan Sen. Gary Peters, the chairman of the Senate Homeland Security and Governmental Affairs Committee, said at a March hearing he conducted.

Slow public disclosure

Months after the 2020 report came out, then-President Donald Trump issued an executive order on artificial intelligence, part of which required agencies to list all non-classified or non-sensitive uses of AI.

But two years later, about half the large agencies that were known to have used some AI hadn’t published the required inventory, according to a December 2022 study by the Stanford Institute for Human-Centered Artificial Intelligence.

And some of the published inventories lacked important information. For example, Customs and Border Protection did not report it was using facial recognition technology to track who is entering and exiting the country.

“The initiative to start cataloging those use cases was an important one, and it’s very much work in progress,” Arati Prabhakar, director of the White House Office of Science and Technology Policy, said at a September congressional hearing when asked about the disclosure rate.

Meanwhile, federal agencies are being inundated with sales pitches from AI companies promising the next big thing.

“We're hearing from our federal procurement officers that they're basically being bombarded by companies wanting to demonstrate the promise of their products,” Peters said at a September hearing.

Hundreds of AI uses

The government’s growing public list of current and planned AI uses includes more than 700 examples. It does not cover sensitive areas like intelligence gathering and the military. The Defense Department alone has more than 685 unclassified AI projects, according to the nonpartisan Congressional Research Service.

Disclosures from other agencies show AI is being used to document suspected war crimes in Ukraine, test whether coughing into a smartphone can detect COVID-19 in asymptomatic people, stop fentanyl smugglers from crossing the southern border, rescue children being sexually abused and find illegal rhino horns in airplane luggage – among many other things.

In 2021, the VA became one of the first federal agencies to release an AI strategy. The agency says it’s uniquely suited to advance the use of AI because of its vast data sets of administrative, financial and medical records. And because many doctors get at least some of their medical training at the VA, the agency can train them in AI as well.

“There's no reason that the VA shouldn't be the world leader in adopting AI into health care practices,” Sen. Joe Manchin, D-W.Va., said at a November hearing on the agency’s research.

How the VA is using AI

Early in the coronavirus pandemic, the VA developed a machine learning model to predict how sick a patient would get from COVID-19.

Another tool uses real-time data to anticipate episodes of post-traumatic stress disorder or the risk of suicide.

For those who’ve reached out to the Veterans Crisis Line, a natural language processing engine could more quickly identify and help veterans in crisis.

While the administration touts the benefits of AI, the sweeping executive order that Biden announced in October tasks federal agencies with roughly 150 action items with urgent deadlines, including proposed new rules to address issues like privacy and bias.

Well-documented bias in AI

Bias in AI is well-documented, Fei-Fei Li, co-director of the Stanford Institute of Human Centered Artificial Intelligence, told Congress at a September hearing. For example, predictive tools used to approve or reject loans are less accurate for low-income minority groups because there’s not as much data in their credit histories.

And a Stanford study found that the IRS audits Black taxpayers at least three times more often than non-Black taxpayers because of problems in the computer algorithms used to spot potential tax cheats.

Under the draft policy Biden announced that is still being finalized, safeguards include real-world testing, independent evaluations and public notification.

If an agency wants to use AI to help determine who qualifies for a social safety net program, officials “would be required to address algorithmic discrimination and build in an avenue for appeal,” Prabhakar, the director of the White House Office of Science and Technology Policy, said recently. “This is how we get to more responsible use of AI by government.”

Role for Congress regulating artificial intelligence

Rob Weissman, president of the consumer advocacy organization Public Citizen, said Biden’s executive order is impressive in the breadth and relatively depth of issues it covers.

“However, it’s just a first step in that it instructs agencies to take action across a diverse set of areas,” he added. “How impactful the (executive order) is ultimately will depend on those agencies’ actions.”

It’s also, he said, no substitute for Congress putting rules and restrictions into law.

Rep. Nancy Mace, R-S.C., one of the lawmakers working on AI legislation, noted that the administration’s recent guidance to federal agencies arrived two years after the deadline Congress previously set. That makes her skeptical agencies will meet their new deadlines, "because their track record is pretty useless," she said at a hearing she chaired last week.

Rumman Chowdhury, a data and social scientist who has built AI for the past decade, told Mace that AI being used by the public sector "needs to work for 100% of the people from Day One."

"This is not an Uber for puppies," Chowdhury said. "These are things that critically matter to individuals. So, we have to be very careful in how we roll things out so that they’re equitable for all.”

Don't get fooled How to spot artificial intelligence deepfakes

This article originally appeared on USA TODAY: Biden's AI order comes as use grows: From veterans to rhino horns