AI GOVERNANCEImmigration as a Test Case for Executive AI Governance
DHS deploys AI in immigration enforcement largely untested, undisclosed, and unreviewed, a preview of how AI could reshape executive power. DHS deploys AI in immigration enforcement largely untested, undisclosed, and unreviewed, a preview of how AI could reshape executive power. These systems are being developed and normalized first on noncitizens before expanding to all. DHS is now deploying artificial intelligence (AI) systems that already exhibit many of the governance problems AI scholars warn could emerge as more capable systems enter government.
In the wake of the Sept. 11, 2001, attacks, the PATRIOT Act and the Homeland Security Act were enacted as urgent responses to reorganize and bolster the U.S. counterterrorism apparatus. Some of these authorities were intended to be temporary emergency measures. Subsequent reauthorizations and agency practice, however, transformed key elements of those laws into enduring infrastructure through which domestic surveillance expanded, became normalized, and ultimately extended well beyond the terrorism-related purposes that initially justified them.
Much of that infrastructure was first deployed to target noncitizen communities, among whom constitutional protections are weakest, judicial review is most deferential, and political accountability is limited. The watchlisting systems, biometric databases, and information-sharing networks developed in immigration enforcement did not remain confined there. Today, they have become part of the broader surveillance architecture that affects everyone.
Once again, these systems are being developed and normalized first on noncitizens before expanding to all. The Department of Homeland Security (DHS) is now deploying artificial intelligence (AI) systems that already exhibit many of the governance problems AI scholars warn could emerge as more capable systems enter government. These systems operate at a speed that can outpace meaningful oversight. They make accountability more difficult when decisions go wrong. And they do so with remarkably little public visibility into how they function or how they are used.
Records produced through recent Freedom of Information Act (FOIA) litigation by Just Futures Law, Pangea Legal Services, and Mijente show that Immigration and Customs Enforcement (ICE) operated roughly 42 AI tools internally as of May 2024, but only 23 tools appeared in DHS’s public AI Use Case Inventory at the time. Customs and Border Protection (CBP) operated approximately 75 AI systems in the same FOIA-era records, 62 of them active, a figure that has since converged with CBP’s public inventory page, which lists roughly 60 active use cases.
Taken together, these gaps are a live example of how AI is already reshaping the exercise of executive power.
In a recent article, Cullen O’Keefe, Alan Rozenshtein, and Christoph Winter introduced the concept of “ExecAI:” the use of advanced AI systems within the executive branch and the ways those systems could erode the separation of powers. They ask what happens when presidential directives can be executed without the institutional friction that human decision-makers often provide: opportunities for dissent, whistleblowing, meaningful attribution when errors occur, and enough time for courts to intervene.
