Google’s Pentagon AI deal tests whether military models can outrun their risks

Google has signed an agreement allowing the Pentagon to use its Gemini AI model on classified systems, placing the company beside OpenAI and Elon Musk’s xAI in a fast-expanding DOD effort to bring commercial models into national security work, reports the New York Times. The arrangement falls under a $200 million contract signed last year and reportedly permits use for “any lawful governmental purpose,” a phrase broad enough to satisfy Pentagon flexibility but vague enough to worry employees and outside observers.

The deal arrives as DOD and Anthropic remain at odds over whether a model provider can insist on guardrails against autonomous weapons or domestic surveillance. Anthropic refused to remove those restrictions and was later labeled a “supply chain risk,” effectively cutting it off from federal work while the company challenges the designation. Google says it opposes domestic mass surveillance and autonomous weapons without appropriate human oversight, but the article leaves open how much contractual force those commitments have once a model enters classified networks.

That ambiguity matters. DOD’s own responsible AI framework says military AI must be lawful, ethical, responsible and accountable, and must earn trust through testing, procurement discipline and operational controls. Its autonomy policy also emphasizes understandable human-machine interfaces, auditable systems, cybersecurity and compliance with the law of war. Those are higher bars than simply buying access to a powerful model.

DOD’s demand is real. The Pentagon is seeking $2.3 billion to expand Project Maven, the AI-enabled battlefield system built by Palantir, and Maven is increasingly tied to joint fires, targeting workflows and operational decision support. But the skeptical read is straightforward: performance benchmarks, cloud scale and vendor prestige do not prove suitability for classified, high-consequence military decisions. The Pentagon may be avoiding lock-in by signing multiple AI labs, yet multiplying suppliers could also multiply uncertain failure modes.