EFF Calls Out OpenAI Pentagon Deal: "Weasel Words" Won't Stop Mass Surveillance

TL;DR:

EFF published a detailed analysis of OpenAI's Pentagon contract, calling the safeguards "weasel words"
Key problem: "intentionally." Intelligence agencies have long claimed they collect Americans' data "incidentally," not intentionally
"Consistent with applicable laws" is meaningless when the law already allows warrantless surveillance through data brokers
Brad Carson, former Army general counsel: "I've reluctantly come to the conclusion that this provision doesn't really exist, and they are just trying to fake it"
ChatGPT uninstalls surged 300% after the deal was announced

EFF Breaks Down the Contract Language

The Electronic Frontier Foundation published its analysis of OpenAI's amended Pentagon deal on March 6, 2026. Their verdict: it's "weasel words" designed to look good in a press release while giving the government flexibility to do whatever it wants.

EFF went line by line through the contract language OpenAI touted as protective. Each term has a history, and that history isn't reassuring.

"Intentionally": The Magic Word

The contract states the AI system "shall not be intentionally used for domestic surveillance of U.S. persons and nationals."

That word has a long, ugly history in surveillance law.

For decades, intelligence agencies have run massive collection programs that sweep up Americans' communications. They call it "incidental collection." The NSA wasn't trying to grab your emails. They just happened to end up in the dragnet while targeting someone overseas.

Section 702 of FISA works exactly this way. The government targets foreigners, "incidentally" collects on millions of Americans, then searches that database for domestic leads. All legal. All "unintentional."

OpenAI's contract doesn't close this loophole. It codifies it.

"Consistent With Applicable Laws": Which Laws?

The contract restricts use to purposes "consistent with applicable law, operational requirements, and well-established safety and oversight protocols."

EFF points out what that actually means: the government can do anything that's currently legal.

Right now, under U.S. law, it's perfectly legal to:

Buy your location data from data brokers
Purchase your web browsing history
Access your financial records through third-party vendors
Aggregate your social media posts into surveillance profiles

Samir Jain of the Center for Democracy & Technology put it bluntly: "Right now, under U.S. law, it's lawful for government authorities to buy up commercially available information from data brokers and other third parties."

OpenAI's contract doesn't prohibit mass surveillance. It references the laws that permit it.

"Deliberate Tracking": A Term With Baggage

The amended contract prohibits "deliberate tracking, surveillance, or monitoring of U.S. persons."

EFF highlights the problem: intelligence agencies have long used commercially purchased data specifically to circumvent stronger privacy protections. They're not "tracking" you. They're just buying tracking data from someone else.

The government's position, upheld by secret FISA court rulings, is that purchasing data doesn't constitute tracking. The tracking already happened. The government is just buying the results.

This linguistic gymnastics has enabled warrantless surveillance for years. OpenAI's contract plays the same game.

"This Provision Doesn't Really Exist"

Brad Carson, who served as the Army's general counsel, reviewed OpenAI's contract language. His assessment was devastating.

"I've reluctantly come to the conclusion that this provision doesn't really exist," Carson said, "and they are just trying to fake it."

Carson's analysis: the restrictions sound meaningful in press releases. In practice, they contain enough qualifiers and references to existing law that they don't actually constrain anything.

The contract doesn't change what the Pentagon can do. It just provides talking points.

What Anthropic Demanded, And Didn't Get

The contrast with Anthropic matters here.

When the Pentagon approached Anthropic with similar contract terms, CEO Dario Amodei refused. His concern: "Advanced AI is capable of taking scattered, individually innocuous data (like a person's location, finances, search history) and assembling it into a comprehensive picture of any person's life, automatically and at scale."

Anthropic demanded explicit categorical prohibitions on mass domestic surveillance and autonomous weapons. Not "intentionally." Not "deliberately." Not "consistent with applicable law." Explicit bans.

The Pentagon designated Anthropic a "supply chain risk," the same label applied to Huawei and ZTE. OpenAI signed a deal.

That deal uses exactly the language Anthropic warned about.

The Market Responded

Users aren't buying OpenAI's reassurances.

According to EFF, ChatGPT uninstalls surged nearly 300% after the Pentagon deal was announced. Claude, Anthropic's AI assistant, overtook ChatGPT in App Store rankings as users voted with their feet.

Sam Altman admitted the initial agreement was "opportunistic and sloppy." The amended version was supposed to fix that.

EFF's argument: it didn't. The words changed. The loopholes remained.

Why This Matters

This isn't about one contract. It's about the infrastructure being built.

The Pentagon is integrating AI systems across its operations. DHS maintains a face database with 1.2 billion images. ICE uses AI-powered surveillance tools to track immigrants and critics. CBP pays for access to Clearview AI's database of 60 billion scraped photos.

OpenAI's contract doesn't exist in isolation. It's a piece of a surveillance apparatus that's been growing for years. The question isn't whether the contract language is slightly better than the first draft. The question is whether AI companies should be helping the government build mass surveillance systems at all.

Anthropic said no. OpenAI said yes, with caveats.

EFF's point: those caveats are "weasel words." They look like protections. They read like protections. But they're written by lawyers who know exactly how the government will interpret them.

And history suggests that interpretation won't favor your privacy.

Sources

Published: March 7, 2026