FAS Senior Fellow Jen Pahlka testifies on Using AI to Improve Government Services

Jennifer Pahlka (@pahlkadot) is a FAS Senior Fellow and the author of Recoding America: Why Government is Failing in the Digital Age and How We Can Do Better. Here is Pahlka’s testimony about artificial intelligence presented today, January 10, 2024, to the full Senate Committee on Homeland Security and Government Affairs hearing on “Harnessing AI to Improve Government Services and Customer Experience”. More can be found here, here, and here.


How the U.S. government chooses to respond to the changes AI brings is indeed critical, especially in its use to improve government services and customer experience. If the change is going to be for the better (and we can’t afford otherwise) it will not be primarily because of how much or how little we constrain AI’s use. Constraints are an important conversation, and AI safety experts are better suited to discuss these than me. But we could constrain agencies significantly and still get exactly the bad outcomes that those arguing for risk mitigation want to avoid. We could instead direct agencies to dive headlong into AI solutions, and still fail to get the benefit that the optimists expect. The difference will come down to how much or how little capacity and competency we have to deploy these technologies thoughtfully.

There are really two ways to build capacity: having more of the right people doing the right things (including but not limited to leveraging technology like AI) and safely reducing the burdens we place on those people. AI, of course, could help reduce those burdens, but not without the workforce we need – one that understands the systems we have today, the policy goals we have set, and the technology we are bringing to bear to achieve those goals. Our biggest priority as a government should be building that capacity, working both sides of that equation (more people, less burden.)

Building that capacity will require bodies like the US Senate to use a wide range of the tools at its disposal to shape our future, and use them in a specific way. Those tools can be used to create mandates and controls on the institutions that deliver for the American people, adding more rules and processes for administrative agencies and others to comply with. Or they can be used to enable these institutions to develop the capacity they so desperately need and to use their judgment in the service of agreed-upon goals, often by asking what mandates and controls might be removed, rather than added. This critical AI moment calls for enablement.

The recent executive order on AI already provides some new controls and safeguards. The order strikes a reasonable balance between encouragement and caution, but I worry that some of its guidance will be applied inappropriately. For example, some government agencies have long been using AI for day to day functions like handwriting recognition on envelopes or improved search to retrieve evidence more easily, and agencies may now subject these benign, low-risk uses to red tape based on the order. Caution is merited in some places, and dangerous in others, where we risk moving backwards, not forward. What we need to navigate these frameworks of safeguard and control are people in agencies who can tell the difference, and who have the authority to act accordingly.

Moreover, in many areas of government service delivery, the status quo is frankly not worth protecting. We understandably want to make sure, for instance, that applicants for government benefits aren’t unfairly denied because of bias in algorithms. The reality is that, to take just one benefit, one in six determinations of eligibility for SNAP is substantively incorrect today. If you count procedural errors, the rate is 44%. Worse are the applications and adjudications that haven’t been decided at all, the ones sitting in backlogs, causing enormous distress to the public and wasting taxpayer dollars. Poor application of AI in these contexts could indeed make a bad situation worse, but for people who are fed up and just want someone to get back to them about their tax return, their unemployment insurance check, or even their company’s permit to build infrastructure, something has to change. We may be able to make progress by applying AI, but not if we double down on the remedies that failed in the Internet Age and hope they somehow work in the age of AI. We must finally commit to the hard work of building digital capacity.