Google's AI Models Head to Classified Pentagon Networks The tech giant joins OpenAI and xAI in providing artificial intelligence for defense operations, as hundreds of employees petition against the deal.
Google has agreed to supply its AI models for classified Pentagon networks, according to reporting from The Information today. The contract requires Google to modify its safety filters at the government's request and permits use for any lawful government purpose, including weapons targeting with human oversight.
This marks a major shift for Google, which faced employee revolt over its Project Maven drone imaging contract in 2018. Now the company faces the same regulatory fault line that recently led to Anthropic's blacklisting by the Department of Defense. Where Anthropic maintained strict ethical guardrails that prevented military applications, Google's contract threads a narrower needle: prohibiting fully autonomous weapons and domestic mass surveillance, but otherwise deferring to military operational decisions.
The deal arrives as the Pentagon speeds up its AI adoption across classified systems. OpenAI and Elon Musk's xAI have already signed similar agreements, creating what amounts to a new defense contractor tier among foundation model providers.
According to multiple outlets, over 560 Google employees signed a petition urging the company to refuse the contract, citing ethical concerns about AI in military applications. The company has not publicly responded to the petition.
The contract's specific provisions show careful negotiation around AI safety concerns. Google must adjust its models' safety filters when defense officials ask, but retains some boundaries: the AI cannot be used for domestic mass surveillance or for autonomous target selection without appropriate human oversight, according to Reuters' reporting.
Get the latest model rankings, product launches, and evaluation insights delivered to your inbox.
That phrase carries considerable weight. It suggests weapons systems where AI identifies and ranks targets, but a human makes the final decision to fire. This mirrors existing drone strike protocols, where operators review algorithmic recommendations before authorizing lethal action.
The timing follows the Pentagon's recent falling out with Anthropic, whose Claude models were reportedly blacklisted from defense networks due to the company's unwillingness to modify its strict usage policies. Anthropic held hard lines; Google appears to have negotiated conditional boundaries.
Neither Google nor the Department of Defense provided comment to the outlets reporting this story.
The contract exposes a growing divide in Silicon Valley over military AI applications. While some companies pursue defense contracts as a growth market, others maintain that current AI systems, prone to hallucinations and unpredictable failures, shouldn't power weapons systems at all.
Google employees' petition echoes the 2018 Project Maven protests that led the company to withdraw from that drone imaging contract. The deal permits weapons targeting applications as long as humans retain decision authority. Anthropic's stricter ethical stance resulted in DoD blacklisting, creating an opening for competitors. The contract explicitly prohibits domestic mass surveillance and fully autonomous weapons. Safety filter modifications must be made at government request, removing Google's veto power over specific uses.
The Pentagon's push to integrate foundation models into classified systems appears to be accelerating despite ongoing debates about AI reliability in high-stakes applications. With three major AI providers now under contract, the focus shifts from whether military AI deployment will happen to how human oversight mechanisms will function when algorithmic recommendations arrive at the speed of modern warfare.