Your Data Was Hard to Read. LLMs Fixed That for the Government.

Rows of server racks in a data center with blue LED lights glowing in a dark room — Photo via Unsplash

TL;DR: The US government already collects massive amounts of data on Americans (location records, browsing history, financial transactions, social media activity) purchased from commercial data brokers without warrants. The bottleneck was always analysis: too much data, not enough analysts. Large language models eliminate that bottleneck. An LLM can do what thousands of intelligence analysts do (read, cross-reference, and profile) in seconds. MIT Technology Review reported on April 21 that this capability could "supercharge mass surveillance." Anthropic CEO Dario Amodei called AI-enabled mass surveillance "a crime against humanity." His company refused to let the Pentagon use Claude for domestic surveillance, and got blacklisted. Rep. Thomas Massie put it plainly: "Imagine instead of doing a query with one person that you turned AI loose on these databases. There’s virtually nothing the government can’t know about you."

The Government Had the Data. It Couldn't Read It.

FBI Director Kash Patel confirmed on March 18, 2026, that the bureau buys Americans' location data from commercial data brokers ^[1]. DHS requested $108 billion in 2025 funding. ICE got roughly $86 billion after the "Big Beautiful Bill" supplemental appropriation topped up its $10 billion base budget with $75 billion for detention and enforcement through 2029. The CIA launched a new acquisition framework in February specifically to fast-track private sector AI partnerships ^[2].

None of this is new. Federal agencies have been hoovering up commercial data for years: location records from your weather app, browsing history from data brokers, financial transactions, even data from connected cars that track your speed, location, and passengers ^[3].

The problem was never collection. It was comprehension.

Think about what a single person's data broker profile contains: every place they've been (via cell tower and GPS data), every website they've visited, their purchase history, their app usage patterns, their social connections. Multiply that by 330 million Americans. You've got a mountain of data and no one to read it.

Human analysts could handle targeted investigations. Pull records on a specific suspect, cross-reference their movements, build a case. But analyzing everyone? Finding patterns across millions of people simultaneously? That required more analysts than any agency could hire.

That constraint is what kept bulk surveillance from becoming true mass surveillance. You could collect it all, but you couldn't understand it all.

Until now.

What LLMs Actually Change

A large language model (the same technology behind ChatGPT, Claude, and Gemini) can read and analyze text at a speed and scale no human team can match. That's the whole point of the technology. It's also what makes it terrifying in a surveillance context.

MIT Technology Review laid it out on April 21, 2026: LLM agents "could potentially do the work of intelligence analysts in a fraction of the time and for a fraction of the cost," which would "enable the state to aim its all-seeing eye toward anyone, not just its highest-priority targets" ^[4].

Here's the difference in concrete terms:

Before LLMs: An FBI analyst gets a tip about someone. They pull that person's data broker records, their Section 702 intercepts, their social media. They read through it manually. They write a report. One person, one case, days or weeks of work.

With LLMs: An agency feeds millions of data broker profiles into an LLM system. It cross-references location patterns, communication metadata, financial transactions, and social connections. It flags anomalies. It builds profiles. It does in seconds what would take an army of analysts years. And it does it for everyone, not just specific targets.

That's not hypothetical capability. That's what the technology does right now.

This Isn't Google. It's Worse.

Government databases already had search functions. You could query a name and pull up records. But searching and understanding are different things.

A database search finds exact matches. An LLM reads context. It infers relationships. It identifies patterns that no keyword search would catch, like noticing that three people who attended the same protest also visited the same lawyer's office and started using encrypted messaging apps within the same week.

Rep. Thomas Massie (R-KY) put it bluntly in an NBC News interview: "Imagine instead of doing a query with one person that you turned AI loose on these databases. There's virtually nothing the government can't know about you" ^[5].

The government already proved it would abuse search capabilities. FISA Section 702 databases (ostensibly for foreign intelligence) were used to search for Black Lives Matter protesters, political campaign donors, elected officials, and a state judge who complained about police abuses ^[5]. That was with basic keyword searches.

An LLM doesn't just search. It reads. It reasons. It connects. And it doesn't get tired, doesn't need coffee breaks, and can process a million profiles while you're reading this paragraph.

Anthropic's CEO Called It a Crime Against Humanity

Dario Amodei, CEO of Anthropic (the company that makes Claude), published a lengthy essay in January 2026 warning that AI-enabled mass surveillance "could constitute a crime against humanity" ^[6].

He wasn't being dramatic. He was describing what his own technology could do.

Amodei warned that "a powerful AI looking across billions of conversations from millions of people could gauge public sentiment, detect pockets of disloyalty forming and stamp them out before they grow" ^[6]. He argued that autocrats could use AI-generated analysis to "permanently steal" the freedom of citizens.

Then he put his money where his mouth was. Anthropic signed a $200 million contract with the Pentagon in July 2025. When the DOD wanted Claude deployed on its GenAI.mil platform without restrictions on domestic surveillance or autonomous weapons, Anthropic refused ^[7].

The Pentagon's response: it designated Anthropic a "supply chain risk" and blacklisted the company. Defense Secretary Pete Hegseth called Amodei a "liar" with a "God complex" ^[8]. A federal judge later found the blacklisting was "classic illegal First Amendment retaliation" ^[9], but an appeals court reversed the temporary block, and the case remains in litigation.

The message to every other AI company was clear: cooperate, or get cut off.

The Companies That Didn't Say No

Senator Ron Wyden (D-OR) sent letters to major AI companies in March 2026 asking about their policies on government surveillance use. The responses (and silences) tell the story ^[5].

Google's Anne Wall acknowledged concerns about "the intersection of rapidly advancing AI and government operations." That's a non-answer in a suit.

Anthropic's Brian Peters said the company bans "unauthorized surveillance" but granted an exception for a "small number of national-security customers" for foreign intelligence analysis "in accordance with applicable law", including analysis of incidentally collected US person data ^[5].

OpenAI didn't respond. Neither did xAI, Elon Musk's AI company.

Silence from two of the biggest AI companies on whether they'd let the government use their models for mass surveillance. Let that sink in.

The Legal Framework Has a Hole the Size of an LLM

Here's why this matters right now: the laws governing surveillance were written before LLMs existed. The Fourth Amendment protects against "unreasonable searches and seizures", but courts have generally ruled that data you share with companies isn't protected, because you "voluntarily" gave it up ^[10].

That's the data broker loophole. The government can't tap your phone without a warrant. But it can buy your phone's location data from a broker who got it from an app you didn't read the privacy policy for. Perfectly legal.

FISA Section 702 has a similar gap. It authorizes collection of foreign communications, but scoops up American data when those Americans communicate with foreigners. The NSA and FBI can then search that American data, without a warrant ^[5].

Neither framework contemplated what happens when you add an LLM to the mix. There's no law that says the government can't use AI to analyze data it already legally possesses. No statute requires a warrant before feeding bulk data broker records into a language model. No regulation limits how many profiles an AI system can build simultaneously.

The existing rules assumed human constraints: limited analysts, limited time, limited capacity to process information. LLMs break every one of those assumptions.

Three Days to the Deadline

FISA Section 702 expires on April 30, 2026. Speaker Mike Johnson's third attempt at reauthorization (a three-year extension through 2029) contains no warrant requirement for searching Americans' data and no restrictions on AI-powered analysis ^[11].

The Government Surveillance Reform Act, introduced by Rep. Warren Davidson (R-OH) and Sen. Wyden, would close the data broker loophole and add warrant requirements. It hasn't gotten a floor vote ^[5].

The Surveillance Accountability Act from Rep. Massie and Rep. Lauren Boebert (R-CO) would require warrants for searching Section 702 data. Also stalled ^[12].

Every reform bill that would address AI-powered surveillance is stuck. The clean extension, the one that changes nothing, is the one moving through Congress.

If Section 702 gets renewed without AI-specific guardrails, the government gets three more years of warrantless access to a database that LLMs can now read like an open book.

What You Can Do

The honest answer: not enough. LLM-powered surveillance is a structural problem that requires structural solutions: laws, not apps. But you can reduce the data available for analysis:

Audit your apps. Every app with location access is feeding data to brokers. Check Settings > Privacy > Location Services on iOS, or Settings > Location on Android. Kill access for anything that doesn't need it.
Opt out of data brokers. Sites like Privacy Guides maintain lists of major brokers and opt-out links. It's tedious. Do it anyway.
Use encrypted communications. Signal for messaging, Proton Mail for email. End-to-end encryption means even if metadata is collected, content stays private.
Call your representative. Tell them to support the Government Surveillance Reform Act and oppose clean Section 702 extensions without AI guardrails. The EFF's Section 702 campaign page has tools to find and contact your rep.
Disable connected car data sharing. If your car was made after 2017, check the manufacturer's privacy settings. Many share driving data with insurers and brokers by default.

The Bottom Line

The government spent decades building a surveillance apparatus with one critical weakness: it couldn't process everything it collected. LLMs erase that weakness.

The same technology that writes your emails and summarizes your meetings can read every data broker record the FBI has purchased, every communication the NSA has intercepted, every financial transaction DHS has flagged, and build a detailed profile of your life in seconds.

The CEO of the company that built one of these models called it a potential crime against humanity. The Pentagon blacklisted his company for saying so. Congress is about to renew the surveillance framework that makes it all possible, without a single provision addressing AI.

The bottleneck is gone. The data is there. The tools exist. The only question left is whether anyone in power will set limits before it's too late.

April 30 is in three days.

References

Published: April 27, 2026