Pete Hegseth, America’s secretary of war, is taking a my-way-or-the-highway approach to the use of artificial intelligence on the battlefield. On February 24th he gave an ultimatum to Anthropic, maker of the Claude family of models: if it did not agree to terms set by the Pentagon on usage of its AI for military purposes, it would face severe penalties. It is not the first time this administration has picked a fight with a company that fails to follow its orders. But Anthropic has leverage.
The showdown took place during a meeting at the Pentagon between Mr Hegseth and Dario Amodei, Anthropic’s boss. Among AI labs, Anthropic was the first to do classified work for the Department of War (DoW), via a partnership with Palantir, a data firm, and Amazon Web Services, a cloud provider. But it also has clear red lines when it comes to the use of its models for national security. In negotiations with the DoW, it has insisted that Claude must be used neither for mass domestic surveillance nor for autonomous weapons.
The restrictions have put it at loggerheads with Mr Hegseth, who has stipulated that companies providing the Pentagon with AI models must give it carte blanche to do with them what it likes, provided the actions are lawful. At the latest meeting with Mr Amodei, Mr Hegseth vowed to terminate Anthropic’s contract by February 27th if the AI lab did not agree to his terms. A senior Pentagon official said that if Anthropic did not “get on board” with the DoW’s conditions, it would risk being labelled as a supply-chain risk, and that the government could invoke the Defence Production Act (DPA), which gives the president authority to oblige companies to do national-security work.
Anthropic’s main contract with the DoW is worth no more than $200m, a trifling sum for a firm that generated an annualised $14bn of revenue in February. But it cannot take the stand-off lightly. Barring all of the Pentagon’s suppliers from using Claude would have a much bigger impact on the AI lab. It is a measure usually reserved for companies linked to hostile powers. As for the DPA, it has been invoked in recent emergencies such as the covid-19 pandemic, but is rarely brandished in such an adversarial way.
That the Pentagon is threatening such drastic measures, however, indicates its reliance on Anthropic. Dangling the DPA suggests that the DoW is reluctant to rip Claude out of its systems. That is because Anthropic’s model is seen as one of the best available, according to former defence officials with ties to Silicon Valley.
Could the stand-off create an opening for rival labs with fewer qualms? OpenAI, Anthropic’s nemesis, has been slower to seize the opportunity to work with the DoW. Its models are used by Microsoft, with which it was once joined at the hip, for highly classified defence work, but OpenAI is not a party to the contract. Some contestants in a competition to build voice-activated drone-swarming technology for the Pentagon are using the lab’s models, but again its involvement is indirect. Its only formal contracts with the DoW are for unclassified work, and the use of its models for national-security purposes is considered on a case-by-case basis.
Fears of militarising AI run deep at both Anthropic and OpenAI. The pair are also alert to the risk of losing their brainy AI researchers, many of whom come from abroad and may not share the Trump administration’s ideology.
By contrast, Elon Musk, who previously warned against “killer robots”, appears to have shed his compunctions. SpaceX, his rocket and satellite company, and xAI, his model-maker (with which SpaceX is merging), are said to be competing together in the Pentagon contest to make drone-swarming technology. Grok, xAI’s model, is “on board” for use in classified settings, says a Pentagon official.
Google, having scrapped restrictions on the use of AI for defence purposes in 2024, is also taking on contracts for classified and unclassified work with the Pentagon. That is a striking reversal for the Silicon Valley giant. In 2018 it was forced to relinquish a Pentagon contract called Project Maven, which used machine learning to analyse footage from drones, after an internal revolt at the company.
Since the Project Maven days, the mood in Silicon Valley has become more pro-Pentagon. But that saga carried lessons for both sides that are worth remembering. For tech firms, it may be unrealistic to think that they can control how their technologies are used on the battlefield. They can urge caution, but it is constitutional oversight of the armed forces that ultimately determines how wars are fought.
And for the DoW’s part, demanding unfettered access to technologies with the potential for extreme lethality requires building a bedrock of trust. That can be eroded if these technologies are used for actions of dubious legality. Controversial decisions such as strikes against civilian drug-smuggling boats in the Caribbean have raised concerns in Silicon Valley about how autonomous weapons systems could be misused in the future. If Mr Hegseth is not careful, he may jeopardise his access to more than just Claude.
To track the trends shaping commerce, industry and technology, sign up to “The Bottom Line”, our weekly subscriber-only newsletter on global business.