Google’s deal to let the Pentagon use its Gemini AI models on classified networks has triggered major concerns inside the company, where employees fear the agreement gives the U.S. military too much freedom to use the technology while leaving Google with little power to enforce its own safeguards.
The agreement, first reported by technology news outlet The Information, allows the Department of Defense to use Google’s AI models for classified work and for “any lawful government purpose,” according to a person familiar with the matter.
Pentagon wins access to Gemini on classified networks
The deal gives the Pentagon access to Google’s most advanced AI systems inside secure government environments, expanding the Defense Department’s use of commercial AI tools as Washington accelerates efforts to bring frontier models into national-security operations.
The agreement appears to mirror terms accepted earlier this year by OpenAI and xAI, which also allowed their models to be used for broad lawful government purposes. For Google, however, the deal is more sensitive, given the company’s history of employee opposition to military AI work.
Researchers warned against a broad deal
The classified AI deal followed months of pressure from Google and DeepMind researchers who urged senior leaders not to accept Pentagon terms they viewed as too broad.
More than 100 employees had pressed Jeff Dean, Google’s chief scientist, to reject any military arrangement that did not clearly rule out domestic mass surveillance and autonomous weapons. Earlier this week, more than 600 Google workers signed a letter to CEO Sundar Pichai opposing the company’s negotiations with the Defense Department.
Some employees were caught off guard after the deal became public, with reports of the agreement circulating internally before workers received any clear companywide explanation.
Contract terms raise red flags
The main concern centers on the strength of the deal’s guardrails.
The contract says Google’s AI models are not intended for domestic mass surveillance or autonomous weapons, including target selection, without appropriate human oversight and control, but critics say the clause reads more like guidance than a real restriction.
The agreement also states that Google does not have the right to control or veto lawful government operational decisions. That clause has alarmed some AI-safety researchers, who argue it could leave the company unable to block Pentagon uses that employees have repeatedly warned against, provided those uses remain within the law.
The distinction is critical because saying a system “should not” be used for certain purposes may set an expectation, but it is weaker than language that clearly bars those uses, and for some Google employees, that gap is the heart of the dispute, since Google may have little power to enforce those red lines once Gemini enters classified military systems.
Anthropic’s fight shaped the debate
Google’s decision follows an earlier clash between the Pentagon and Anthropic over similar terms.
Earlier this year, Anthropic refused to accept Pentagon terms that would have allowed its Claude models to be used for all lawful government purposes, after weeks of talks over whether the company could keep limits on military, surveillance and weapons-related uses.
The dispute later led the Defense Department to designate Anthropic as a supply-chain risk, turning the company into a reference point for AI workers pushing for tougher limits on government access to frontier models.
At Google, employees had urged leadership to follow Anthropic’s example and insist on stronger safeguards before signing any defense agreement. Instead, Google signed a deal that could expand its federal AI business while reviving an internal debate over how far the company should go in military technology.
AI chip push sends Alphabet shares up nearly 10 percent
The AI controversy comes as Alphabet shares have rallied on Google’s plans to sell its custom Tensor Processing Units to select customers, giving the company a new way to monetize demand for AI computing power.
Google has said it expects only a small amount of TPU-related revenue later this year, with most revenue from the agreements expected in 2027, as the chip agreements point to longer-term cloud demand.