The weapon of mass deportation isn't advanced AI: it's merging every database

By American Community Media · September 17, 2026.

Between December 2024 and January 2026, the U.S. Department of Homeland Security's (DHS) immigration-related artificial intelligence use cases grew by 36%, according to an analysis by the American Immigration Council. The figure is the technical backdrop to what lawyers, researchers and civil rights organizations described at a September 10 panel in San Francisco, organized by American Community Media: an apparatus that cross-references data from multiple sources —federal, state, local and commercial— to locate, detain and deport people, sometimes before even knowing who they are.

The most frequently cited case was Operation Black Rose, an October 2025 DHS raid on a van in Woodburn (Oregon), a town with one of the largest Latino day-laborer communities in the state. According to Stephen Manning, the attorney representing the victims in the class action arising from the case, agents surrounded the vehicle without knowing who was inside —they flagged it as a "target-rich location"— and used Palantir's deportation management platform ImmigrationOS, together with Mobile Fortify, a mobile facial recognition tool, to "locate in real time" the occupants. Manning sums up the sequence as "detain first, justify later": the government argues in court that none of the occupants, handcuffed on the sidewalk, was "under arrest" at that moment, which lowers the required legal standard from probable cause to mere reasonable suspicion. All were processed for deportation except the lead plaintiff, identified by her initials MJMA, intercepted by Manning's team at an ICE facility before being transferred out of Oregon; a recent court order compelling the handover of evidence will now make it possible to investigate how that surveillance net was built. Manning warns that the infrastructure will outlive this administration: "Once you build it, you reuse it. It doesn't go away."

For Steven Hubbard, a data scientist at the American Immigration Council, the power of this system "doesn't necessarily come from especially sophisticated AI... but from the enormous volume of data the agencies can already search, connect and analyze." By his classification, taken from DHS's own AI inventory, the agency deploys three types of tools: machine learning to detect fraud and score national security risk; language AI to translate, summarize and classify documents; and computer vision, which includes facial recognition and license plate readers. Hubbard compares it to the "skip tracing" debt collection agencies use to locate a person's addresses and assets, but now on a scale that human verification can no longer keep up with.

A March 2025 executive order by Donald Trump, presented on the argument of eliminating "information silos," gave immigration agencies access to Internal Revenue Service (IRS) records, Medicaid, the SNAP food assistance program and the Department of Housing's housing verification system, as well as state voter rolls. Added to that is the data purchased from commercial brokers such as Thomson Reuters and LexisNexis, which —according to Hubbard— rely on a privacy law loophole to sell information that would otherwise require a court order.

One of the least scrutinized channels is the commercial driver's license database of the American Association of Motor Vehicle Administrators (AAMVA). The issue gained prominence after a fatal crash in Florida involving a Sikh truck driver, Harjinder Singh, whom DHS described as undocumented, a version the California governor's office rejected, arguing that it was the Trump Administration itself that approved his commercial license. Florida then sued California and Washington for issuing licenses to people without papers and the Supreme Court declined to hear the case; last summer, DHS and the Federal Motor Carrier Safety Administration demanded that AAMVA hand over the complete records of its 17 million drivers, and 22 states, among them California, sued in August to block the transfer, obtaining a temporary restraining order.

Surveillance also reaches speech. Saira Hussain, an attorney at the Electronic Frontier Foundation (EFF), describes the lawsuit United Auto Workers v. Department of State, filed by EFF together with Muslim Advocates and the Yale Law School clinic, against a program that requires certain visa holders to keep their social media profiles public; in her view, the aim is to identify and intimidate the "disfavored speech" the government itself defines, such as support for Palestinian rights or criticism of U.S. policy. DHS, for its part, has sent hundreds of subpoenas to Meta, Google and Reddit to identify anonymous ICE critics, at times without notifying the affected users, as The New York Times reported.

An ACLU report co-authored by researcher Denise Bell, based on more than 1,200 immigration enforcement incidents across eight states during 2025, documents 624 actions in everyday spaces —sidewalks, parking lots, bus stops, parks—, 252 at workplaces and 107 at courthouses and immigration appointments. The same report records 432 incidents of misconduct, including use of force; 119 acts of retaliation against people filming the operations; and 437 episodes of alleged racial profiling. "It was not incidental, marginal bias. It was the core of how mass surveillance was operated in 2025," says Bell, who links the rise in misconduct to the 287(g) agreements —which authorize local police to enforce immigration law— and to agents operating masked and in unmarked vehicles. Bell sums up her reading of the program, in her own words, as "a blueprint for authoritarianism."

What this record reveals is not an AI to be feared for its intelligence, but for its administrative function: merging what was previously scattered. Neither Mobile Fortify nor ImmigrationOS are frontier systems comparable to the large language models we usually analyze in this section; they are, in essence, database correlation engines with a layer of facial recognition and relatively conventional machine learning. The leap is not technical, it is institutional —an order that erases the silos that for decades separated tax, health, housing and motor vehicle records— and that is what makes it more dangerous: no AI breakthrough is needed to deploy an apparatus of this reach, it is enough to decide politically that the data will be shared.

This fits with something we have been arguing at Zendoric: AI's short-term risk rarely comes from a distant superintelligence, but from the mundane application of already mature tools —facial recognition, risk scoring, data mining— at state scale and without the controls that existed when those searches required a court order or an investigator's manual work. The Black Rose case, with its "detain first, justify later" logic, and the ACLU's finding that racial profiling was "core" and not incidental, show how technical efficiency can erode basic procedural safeguards while responsibility is diluted among agencies, contractors such as Palantir and data brokers operating in the gray zone of privacy law.

In the long run we remain convinced that the ability to process information on a large scale is, in itself, neutral: the same technology that here catalogs day laborers' vans could, under a different political mandate, coordinate the distribution of medical treatments or detect healthcare fraud that today costs lives, on the way toward the abundance we defend as a horizon. But that horizon does not arrive on its own: it depends on society deciding, with the same legal tools now litigating against AAMVA or against DHS, who controls those data flows and under what safeguards. Manning's warning —"once you build it, you reuse it; it doesn't go away"— matters beyond this administration: the surveillance infrastructure erected today on the pretext of immigration will be available tomorrow for any other purpose, whoever is in power. This report is the first installment of a two-part series by American Community Media.