Google’s new Pentagon AI deal has reopened a long-running dispute over how far major AI companies should go in serving military and intelligence customers. The contract gives the U.S. Department of Defense access to Google’s AI models for classified tasks, while employee critics and legal experts are focusing on a central question: who can actually control how the technology is used?
A classified AI contract moves ahead
According to The Information, citing a person familiar with the matter, Google has signed a contract with the Pentagon that allows the U.S. Department of Defense to use Google’s AI models for classified work. The agreement gives the Pentagon access for “any lawful government purpose.”
Google Public Sector described the contract as an extension of an existing agreement from November. A spokesperson said Google remains committed to the view that AI should not be used for “domestic mass surveillance or autonomous weaponry without appropriate human oversight.”
Those assurances did not quiet internal opposition. The deal was signed on the same day more than 600 Google employees, many from DeepMind, sent an open letter to CEO Sundar Pichai asking him to reject classified collaboration with the Pentagon.
The employees framed the issue around accountability. Their concern was not only that AI could be used in harmful ways, but that classified work could prevent Google’s own representatives from knowing enough about the deployment to evaluate it.
Why employees objected
The open letter argued that classified contracts create a visibility problem. If a customer’s work is secret, critics inside the company say Google may be unable to verify whether its models are being used in line with its public commitments.
The employees wrote, according to the Washington Post: “We want to see AI benefit humanity; not to see it being used in inhumane or extremely harmful ways.”
Their position was direct: “The only way to guarantee that Google does not become associated with such harms is to reject any classified workloads,” the letter states.
That argument puts pressure on a familiar tension in AI governance. A company can publish principles, announce limits, and include cautionary language in contracts. But if the work is classified, oversight becomes harder for employees, outside observers, and even some internal teams.
For Google staff who signed the letter, the issue appears to be less about whether government AI work is always unacceptable and more about whether classified Pentagon workloads make meaningful internal control possible.
The contract language is under scrutiny
The contract includes language saying the AI system “is not intended for domestic mass surveillance or autonomous weapons without appropriate human oversight.” On its face, that wording appears to describe a boundary around sensitive uses.
Legal experts cited in the source article say the boundary may not be legally enforceable. The agreement also states, according to The Information: “This Agreement does not confer any right to control or veto lawful Government operational decision-making.”
Charlie Bullock, a lawyer and Senior Research Fellow at the Institute for Law and AI, said the wording “is not intended for, and should not be used for” does not carry legal weight. In his view, it expresses that certain uses would be unwanted, but would not make them a breach of contract.
Amos Toh from NYU’s Brennan Center also questioned the practical meaning of “appropriate human oversight,” according to The Information. He said the phrase does not necessarily require a human to stand between target identification and a fire order.
The source article also notes that the Pentagon has not ruled out fully autonomous weapons systems. That makes the meaning of contract safeguards especially important, because the dispute is not only about policy language. It is about whether that language can prevent specific operational uses.
How Google compares with other AI firms
Google’s deal appears to give the Pentagon more flexibility than some comparable arrangements. OpenAI retained full control over its “Safety Stack” in its February deal, according to OpenAI’s own blog post.
Google, by contrast, has committed to helping the government adjust its AI safety filters upon request, The Information reports. That difference matters because safety filters are one of the mechanisms AI companies use to restrict outputs or reduce certain risks.
Google is not the only AI company with a classified Pentagon contract. Alongside Google and OpenAI, Elon Musk’s xAI also holds one.
Anthropic took a different route earlier this year. It was excluded from a Pentagon deal in February because it demanded contractual guarantees against mass surveillance and autonomous weapons. Anthropic is currently suing over that decision.
At the time, more than 900 Google employees publicly called on the company to support Anthropic’s red lines. The new Google contract shows how quickly the debate has shifted from public principles to the fine print of enforceable agreements.
A reversal from earlier Google policy
The current dispute also echoes Google’s earlier experience with Project Maven. Back in 2018, after thousands of employees protested, Google chose not to renew its Project Maven contract with the Pentagon and pledged never to use AI for weapons or surveillance.
Last year, Google quietly dropped those self-imposed restrictions. The new Pentagon AI contract is being read against that history, especially by employees who see the company moving away from earlier commitments.
Project Maven itself continued after Google stepped away. It is now sold by Palantir and has been used for target selection in the Iran conflict, with support from Anthropic’s Claude model.
The result is a broader industry picture in which major AI companies are increasingly connected to classified government work, while the enforceability of safety promises remains contested. For Google, the immediate issue is the Pentagon contract. The larger issue is whether public AI principles can still carry weight when the most sensitive deployments happen out of public view.