The Machine Already Exists

This spring, Americans who went to town halls and school-board meetings to object to AI data centers began turning up in federal threat files. They had sabotaged nothing and broken no law. They stood up in public rooms and said no—and an apparatus the government built to track immigrants logged them under a threat category that exists in no extremism handbook anywhere: "anti-tech violent extremism."

We know this because reporters caught it. In May, WIRED reviewed more than a thousand pages of unpublished Homeland Security, FBI, and fusion-center documents and found them using a threat category — “anti-tech violent extremism” — that appears nowhere in the agencies’ prior public documentation. NSPM-7, one of Trump’s national-security directives, gives federal agencies legal power to monitor and prosecute groups they label anti-American or anti-capitalist. The tools were presented as a way to enforce immigration law, and for now that is mostly how the administration uses them. But nothing in these tools is specific to immigration. Immigration was the inaugural focus. The systems the administration has stood up will outlast the immigration pretext, and these tools can be aimed in any direction.

Sit with the sequence, because it is the whole argument in miniature. An apparatus built to find immigrants was pointed, within months, at citizens for the crime of showing up to complain. There is no engineering obstacle to pointing it next at the people organizing a protest, certifying an election, or writing a check to the other party. The capability is finished. The only thing standing between it and those people is a habit—the shared, unspoken agreement that certain searches are not run. Habits are the most fragile part of any government, and this administration has made a project of breaking them.

One company holds the keys

At the center of all this is a single firm: Palantir. It matters less for the software it writes than for the thing it uniquely does, which is to connect databases that were built, deliberately, to stay apart.

For decades, the government's records lived in separate rooms. Your immigration file in one, your tax history in another, your health benefits in a third, by design, because a country that had watched other governments turn their filing cabinets into weapons understood that separation was a form of safety. Palantir's platforms dissolve the walls between those rooms. Its systems are authorized to reach across Homeland Security, Health and Human Services, the IRS, the Treasury, and the Defense Department and pull the contents into one continuously updated, searchable list. Build that once for immigration, and pointing it at a different population is an afternoon's work for someone with the right permissions.

The scale of the buildout is visible in the contracts, for anyone willing to read them. Palantir holds an ICE case-management contract worth up to $176.5 million. In April 2025 it added ImmigrationOS, a roughly $30 million dashboard for "targeting" and "self-deportation tracking." According to a shareholder filing compiling federal procurement data, the company has taken more than $81 million in ICE work since Trump returned to office. And in February 2026, Homeland Security opened a blanket purchase agreement worth up to a billion dollars over five years — an account that ICE, Customs and Border Protection, FEMA, and the cybersecurity agency can all spend against without conducting a single new competitive bid.

The clearest window into what this machine is for is a program called ELITE, which assembles deportation targets out of the roughly 79 million people enrolled in Medicaid—their names, their addresses, their Social Security numbers, their race—drawn from the health records of a program for the poor. Once that join exists, the same operation that turns a Medicaid enrollee into a deportation target can turn a donor into an audit, a protester into a problem at work, or a judge's clerk into a security-clearance review. The machine cannot tell the categories apart. It was never asked to.

Put a face to every name

Fusing the records only answers the first question, which is who. The rest of the apparatus answers the next two—where they are, and what to do about them.

The identity layer comes from Clearview AI, whose $9.2 million ICE contract is backed by a database of more than fifty billion scraped face images. That is what converts an ordinary camera into a question: is this one of them? The reach into the physical world comes from Anduril, whose Lattice system now sits under a $20 billion Army enterprise agreement consolidating some 120 contracts — autonomous systems, counter-drone equipment, persistent surveillance, off the shelf. And the money tells you where the government's attention has gone. Federal AI-contract obligations rose from $675 million in 2024 to $7.2 billion in 2026 — a 966% increase in two years (Brookings). DHS's own 2025 inventory lists 238 AI use cases, 35 of them rated high-impact.

What the machine buys him

An apparatus like this does not start an autocoup. What it does is change the terms of one already under way, and it changes them in ways that make this era's backsliding faster and less reversible than the last century's.

It collapses the time it takes to build a list. The purges in Chile, in Turkey, in Hungary were slowed by the sheer manual labor of deciding who to come for, and that delay was often the only warning anyone got. A fusion platform compresses that work from weeks to hours and from thousands of names to millions, ranked by any criterion the man in charge cares to specify. The warning disappears.

It launders the responsibility. When a "high-impact" AI system—Homeland Security's own 2025 inventory lists thirty-five of them—becomes the stated reason for a detention or a denied benefit or an audit, the decision belongs to no one. The writer Dan Davies calls this an accountability sink: a structure that swallows blame so thoroughly that there is no human being left at the bottom of it to sue, to shame, or to subpoena. A machine did it. The machine has no deposition to give.

And it frightens people into compliance without ever touching them. The most important effect of a surveillance state is not the arrests; it is the knowledge that the arrests are possible. Once Americans understand that their donations, their rally attendance, and their posts are logged where the government can reach them, most will adjust their behavior on their own, quietly, without anyone having to lift a finger. This is the mechanism Hungary and Russia relied on above all others. It is cheap, it is deniable, and it works on the many by making an example of the few.

The people who might have said no

None of this becomes real without officials willing to operate it—which is why the administration is dismantling the workforce that once refused, on precisely the same schedule.

Most civil servants have protections that let them decline or delay an unlawful order without losing their jobs. In February 2026, the Office of Personnel Management finalized the rule, renamed "Schedule Policy/Career," that strips those protections away; in June, Trump signed the order moving the first eight thousand or so positions to at-will status, firable at his pleasure. OPM reworked the hiring side to match: the previous year, it added an essay question requiring applicants to describe how they would "advance the President's Executive Orders and policy priorities," while forbidding them from using AI to write the answer—a loyalty oath with a word count.

Between 2017 and 2020, career officials refused illegal orders, slow-walked them, and leaked them, and that friction was, more than once, the only brake that held. The administration studied that experience and drew the obvious lesson. It is removing the brake. The next time a president reaches for the Insurrection Act, activates a Presidential Emergency Action Document, or moves against the people who run an election, he will meet far fewer subordinates willing to stand in the way.

The lists exist. The faces exist. The money is spent and the walls between the databases are down. This year, the administration is converting the people who might have refused into employees who cannot. There is no dramatic moment in any of this, no morning the country will remember as the day it happened. That is the point of building it this way. The machine is finished. What remains is the decision to use it, and the only real question left is who gets named.