Big AI firms meet the Pentagon
Reports say Google is in talks to allow its Gemini models to be used in classified Pentagon operations, marking a shift from earlier employee resistance to defence work. Coverage notes Google would set contractual limits on that use while the Pentagon deepens ties with frontier AI firms, even as regulatory debate fragments between federal proposals and state-level pushback ( ). At the same time, firms are competing on trust metrics — Anthropic says Claude Opus 4.7 scores a 92% honesty rate with fewer hallucinations, a datapoint that matters if models are to be used in sensitive settings (mashable.com).
Google is in talks to let the Pentagon use Gemini in classified settings, a step that would push the company deeper into military artificial intelligence work. (reuters.com) Reuters reported on April 16 that the proposed agreement would allow the Department of Defense to deploy Google’s Gemini models in classified environments. The Information said the talks could expand Google’s role beyond its current unclassified government offerings. (reuters.com; theinformation.com) Google is already selling artificial intelligence tools to federal agencies. In July 2025, Google Public Sector said it had won a $200 million-ceiling contract with the Defense Department’s Chief Digital and Artificial Intelligence Office, and in August 2025 it launched “Gemini for Government” with the General Services Administration. (cloud.google.com; cloud.google.com) The new talks mark a clear change from 2018, when Google said it would not pursue follow-on work on Project Maven after employee protests over a Pentagon contract tied to drone footage analysis. Google said then that the Maven work involved “drone video footage and low-res object identification using AI.” (cloud.google.com) The current negotiations are centered on limits as well as access. Reports say Google has proposed contract language barring domestic mass surveillance and restricting fully autonomous weapons without human control, while the Pentagon has discussed broader “all lawful uses” language. (theinformation.com; nationaltoday.com) The military’s interest is part of a wider scramble to put frontier models inside secure systems where analysts, planners, and operators handle sensitive data. Google said in 2025 that the Defense Department could use its cloud infrastructure, tensor processing units, and agent software under the Chief Digital and Artificial Intelligence Office contract. (cloud.google.com) Rivals are making the same pitch with a different emphasis: trust. Anthropic’s Claude Opus 4.7 launch on April 16 said the model is aimed at coding, agents, and enterprise work, and Mashable reported that Anthropic’s system card put one MASK honesty score at 91.7%, or about 92%, alongside lower hallucination rates on other measures. (anthropic.com; mashable.com) That debate is landing in a fragmented policy landscape. A Congressional Research Service report said federal efforts have focused more on agency oversight, voluntary commitments, and government use of artificial intelligence, while states have moved ahead with their own laws in the absence of a broad federal statute. (congress.gov) Congress is also weighing proposals that would standardize the rules nationally. One example is the American Artificial Intelligence Leadership and Uniformity Act, introduced in the House on September 16, 2025, which remains at the introduced stage. (congress.gov) If Google and the Pentagon reach a deal, the company would be returning to the same military arena it stepped back from eight years ago, but this time with Gemini, classified networks, and written limits at the center of the contract fight. (reuters.com; cloud.google.com)
