The Department of Homeland Security (DHS) has justified the expansion of artificial intelligence (AI) in immigration enforcement with a recurring argument: no decision is made by a machine, as a human agent always reviews the results before taking action. In an article for Tech Policy Press, Emily Froude of Democracy Forward argues that this claim is a fallacy, as humans cannot evaluate the output of an algorithm without being shaped by it.

Memorandum M-25-21, issued in April 2025, mandates safeguards for "high-impact" AI applications - such as those involving civil rights assessments, the right to appeal, and public consultation. However, the criteria hinge on whether AI serves as the "primary basis" for a decision, and it is through this loophole that the DHS finds cover. According to FedScoop, the DHS currently employs over 200 active AI use cases - a nearly 37% increase since July 2025, driven largely by U.S. Immigration and Customs Enforcement (ICE). Of these, 46 were "presumed to be high-impact but determined otherwise" - many of which are within ICE and U.S. Customs and Border Protection (CBP).

Two cases illustrate the problem. Autonomous surveillance towers developed by Anduril - an American defense technology and artificial intelligence firm - independently determine what to flag as a "person" at the border, thereby directing the deployment of armed agents. ELITE - a Palantir application used by ICE - assigns a score from 0 to 100 to each deportation target, representing the probability that the individual is actually at a given address. According to a manual obtained by 404 Media, agents have reported prioritizing areas with a higher density of targets. In practice, it is the algorithm that determines where ICE will come knocking.

The risks of placing blind trust in a machine have already materialized. 404 Media revealed that Mobile Fortify, an ICE facial-recognition app, misidentified a woman detained in Oregon, returning two different names - both of which were incorrect. Nevertheless, ICE has claimed that the app’s output serves as a "definitive" determination of immigration status, asserting that it "should be more reliable than a birth certificate". When an institution itself treats an algorithm as infallible, human oversight becomes nothing more than a formality.

This apparatus extends beyond ICE: 404 Media discovered that U.S. Customs and Border Protection (CBP) has accessed over 80,000 license-plate-reading cameras nationwide - manufactured by Flock Safety. Although ICE may not have a direct, official contract with the company, the system gains access to data from these cameras and automated license plate readers (ALPRs) indirectly, through partnerships with local police departments and other agencies. By classifying these tools as having "low impact," the state effectively strips away the mechanisms that would otherwise allow affected individuals to challenge the decisions that impact their lives. Thus, the human-in-the-loop concept - that is, the active participation of a human in an operation - ceases to be a safeguard and instead functions as a legal shield.

Finally, it is important to note that correctly classifying these tools would not require the Department of Homeland Security (DHS) to abandon them; it would merely require the agency to comply with the rules the government itself established. Therefore, what is at stake is not the existence of automated surveillance, but rather its subjection to minimum standards of transparency. As long as algorithmic decision-making is presented as human-driven, information regarding how the state surveils immigrants will continue to emerge only in a "trickle".

Ultimately, the unchecked expansion of artificial intelligence in immigration surveillance reveals a dangerous reversal of roles: technology ceases to be a supporting tool and instead becomes the true architect of border policy. By reducing human oversight to a mere bureaucratic "rubber stamp", the DHS compromises not only the fundamental rights of immigrants but also the core principles of democratic accountability.