The Defense Department’s Friday announcement formalizes a push to embed commercial artificial intelligence directly into the military’s most sensitive operational planning, a move Pentagon officials say is essential to maintaining a technological edge over rivals such as China. The seven companies that signed on will host their AI models on classified government networks known by the internal platform name GenAI.mil, which is already in use by warfighters, civilians, and contractors for tasks that, in some cases, cut timelines from months to days, according to the Pentagon.
Noticeably excluded from the list is Anthropic, the company behind the chatbot Claude. The startup has been locked in a public dispute with the Trump administration since Defense Secretary Pete Hegseth demanded that any military AI contract allow for any uses the Pentagon deemed lawful. Anthropic insisted on provisions that would bar the use of its technology in fully autonomous weapons and in the surveillance of Americans. After President Donald Trump tried to halt all federal agencies from using Claude and Hegseth moved to label Anthropic a supply chain risk, the company sued.
At least one of the new contracts, according to a person familiar with the agreement who was not authorized to discuss it publicly, contains language requiring human oversight for any missions in which AI systems act autonomously or semiautonomously. The same contract stipulates that the AI tools be used in ways consistent with constitutional rights and civil liberties. Those guardrails echo the sticking points at the center of Anthropic’s fight, though OpenAI has said it secured similar assurances when it first struck a deal with the Pentagon in March — effectively replacing Anthropic as the provider of language models in classified environments.
The broader acceleration of military AI has drawn sharp warnings from researchers, including Helen Toner, interim executive director at Georgetown University’s Center for Security and Emerging Technology and a former board member of OpenAI. Toner said that while AI can help commanders sift through surveillance feeds, organize logistics, and predict maintenance needs, there is a risk of “automation bias,” in which operators become prone to assume that machines work better than they actually do. “How do you roll out these tools rapidly for them to be effective and provide strategic advantage,” Toner asked, “while also recognizing that you need to train the operators and make sure they know how to use them and don’t over trust them?”
Concerns about AI in combat have been heightened by reports of its use during Israel’s recent wars in Gaza and Lebanon, where AI-powered targeting tools were quietly supplied by American tech firms. Civilian casualties soared during those campaigns, fueling fears that the technology contributed to the death of innocent people. The Pentagon’s latest contracts do not explicitly address how the military will prevent such outcomes, though the department has said its AI capabilities are intended to give warfighters “the tools they need to act with confidence and safeguard the nation against any threat.”
Pentagon chief technology officer Emil Michael told CNBC that the decision to sign deals with seven companies was partly a response to the friction with a single partner. “When we learned that one partner didn’t really want to work with us in the way we wanted to work with them, we went out and made sure that we had multiple different providers,” he said. Michael also emphasized that open-source AI models from companies like Nvidia and Reflection offer a path to create an “American alternative” to China’s rapidly advancing AI systems, in which key components are publicly accessible for others to build upon.
For now, much of the AI’s role inside the military will mirror its civilian use: taking on rote tasks such as scheduling maintenance, moving troops and gear, and helping analysts distinguish civilian vehicles from military ones in drone footage. But Toner cautioned that even those applications require disciplined training so that personnel do not defer to machines when human judgment is what the moment demands.