OpenAI Reworks Pentagon Deal After Surveillance Backlash, But Loopholes Remain

TL;DR:

OpenAI amended its Pentagon contract after backlash that domestic surveillance was still possible under the original deal
New language explicitly covers "commercially acquired" data (geolocation, web browsing, financial records from data brokers) not just "private information"
The contract still says AI "shall not be intentionally used" for surveillance: critics say that qualifier leaves room for "accidental" mass tracking
Publicly available data is still fair game: Anthropic demanded restrictions on this; OpenAI didn't get them
Sam Altman admits the deal "was definitely rushed" and "the optics don't look good"

The Contract Gets a Rewrite

Three days after announcing its Pentagon deal, and watching the backlash explode, OpenAI CEO Sam Altman went back to the Pentagon for a do-over.

Altman approached Undersecretary of Defense Emil Michael to rework the contract language. The original text only prohibited collecting Americans' "private information." That left a gaping hole: geolocation data, web browsing history, personal financial records, all available from data brokers, all technically not "private."

The amended language now reads: "The Department understands this limitation to prohibit deliberate tracking, surveillance, or monitoring of U.S. persons or nationals, including through the procurement or use of commercially acquired personal or identifiable information."

That's progress. But it's not what Anthropic demanded, and it's not what privacy advocates wanted.

The "Intentionally" Problem

Here's the sentence that matters: the AI system "shall not be intentionally used for domestic surveillance of U.S. persons and nationals."

That word, intentionally, does a lot of heavy lifting.

What happens if an AI system analyzes data that incidentally includes Americans? What if the Pentagon says it wasn't "intentional"? What if the surveillance happens as a side effect of legitimate intelligence gathering?

Anthropic saw this coming. The company refused to sign any deal that allowed use "for all lawful purposes," because mass surveillance of Americans is often lawful. The government buys location data legally. It purchases social media profiles legally. It accesses financial records legally.

OpenAI's contract permits use "for all lawful purposes, consistent with applicable law, operational requirements, and well-established safety and oversight protocols." That's exactly what Anthropic rejected.

The Public Data Loophole

OpenAI's deal doesn't prohibit collection of Americans' publicly available information. Full stop.

Social media posts. Public records. Protest photos. Church membership rolls. Political donations. Anything you've ever put online, or anyone's ever put online about you.

Anthropic specifically demanded restrictions on this. The company argued that AI "can supercharge the legal collection of publicly available data," turning scattered social media posts into comprehensive surveillance profiles.

OpenAI's national security chief Katrina Mulligan acknowledged the gap: "We can't protect against a government agency buying commercially available datasets."

That's the point. If the data's already for sale, the contract doesn't stop the Pentagon from buying it.

What OpenAI Claims Is Different

OpenAI says it has three layers of protection that go beyond contract language:

Cloud-only deployment: OpenAI retains "full discretion" over its safety classifiers. The Pentagon runs OpenAI's models on OpenAI's infrastructure.
Cleared personnel in the loop: OpenAI employees with security clearances are involved in system operation.
Termination rights: OpenAI claims it can terminate the contract if the government violates the terms.

The company also says it has technical systems that "classify prompts and reject those violating OpenAI's stated limits" and model fine-tuning that makes "systems less compliant with prohibited requests."

One OpenAI employee isn't buying it. Leo Gao, who works on AI alignment, accused his employer of accepting Pentagon terms allowing use "for all lawful purposes," then applying "window dressing" to create an illusion of additional restrictions.

What Anthropic Got That OpenAI Didn't

Anthropic sought two categorical prohibitions that OpenAI's contract doesn't have:

An explicit ban on "mass domestic surveillance of Americans"
An explicit ban on "fully autonomous weapons"

Instead of carving out those exceptions, OpenAI's contract anchors restrictions to existing legal frameworks. The company claims the agreement "locks in" current law standards even if those laws change, but that means the protections are only as strong as current law.

And current law allows the government to buy your location data without a warrant.

"The Optics Don't Look Good"

Altman isn't pretending this went smoothly. The deal "was definitely rushed," he admitted to employees. "The optics don't look good."

He justified the speed by arguing it could "de-escalate" tensions with the military that threatened the entire AI industry. After Anthropic got designated a "supply chain risk" (the same label applied to Huawei) Altman apparently decided quick agreement was better than no agreement.

Many OpenAI employees disagree. Before the deal was signed, rank-and-file workers had signed an open letter supporting Anthropic's refusal to sign without explicit safeguards.

The market had a response too. Claude overtook ChatGPT in App Store rankings as users boycotted OpenAI over the Pentagon contract.

What This Means

OpenAI got a contract with the Pentagon. The contract has more protections than the original draft. But it still allows:

Collection of publicly available data on Americans
Use "for all lawful purposes," which includes a lot of surveillance
Activities that aren't "intentionally" surveillance but might effectively be surveillance

The difference between OpenAI and Anthropic comes down to this: Anthropic demanded categorical restrictions that override current law. OpenAI accepted restrictions tied to current law.

Current law lets the government track Americans without warrants through data broker purchases. Current law allows collection of social media posts, public records, and geolocation data. Current law permits the kind of AI-enabled surveillance that privacy advocates have been warning about for years.

OpenAI's contract doesn't change any of that. It just references it.

Sources

Published: March 3, 2026