It is a tough and very principled confrontation. They have been seemingly irreconcilably opposed to each other for weeks: the US Department of Defense on one side and Anthropic on the other, one of the most innovative companies in artificial intelligence and an important strategic supplier to the Pentagon.
It’s a defining moment for the AI industry, wrote The Wall Street Journal recently. But there is also a lot at stake in this showdown for the Pentagon – and for war and peace in the world in general. The central question is whether American AI companies can ban a certain use of their products if they believe that this would be irresponsible and even dangerous, given the state of the art.
Anthropic does not want the US military to use its AI model Claude for mass surveillance of the US population, nor for the deployment of autonomous weapons that can kill without human intervention. This is stated in the terms of use of the $200 million contract the company signed with the Pentagon last summer.
Anthropic was founded in 2021 precisely to put safety first in the development of AI. At other AI companies such as OpenAI, where Anthropic CEO Dario Amodei initially worked, economic interests would too often override security interests.
Donald Trump’s government wants to get rid of the restrictive conditions that are so important to Anthropic. After a tough conversation with Amodei, Defense Minister Pete Hegseth gave Anthropic an ultimatum on Tuesday: if the AI company does not drop the restrictive conditions before one minute past five in Washington on Friday afternoon (one past eleven o’clock in the evening in the Netherlands), exceptionally severe punitive measures will follow, which could be fatal to the company.
Like other AI companies, Anthropic would have to settle for allowing its products to be used by the Pentagon for anything that doesn’t violate the law. Furthermore, it is up to the ministry, Hegseth believes, to determine what it does with the systems it purchases.
Visit to Elon Musk
During a visit to Elon Musk, whose company xAI is a competitor of Anthropic, the minister already fired a shot across the bow: “The Pentagon will not use AI models that prevent soldiers from waging wars.”
But Anthropic does not shy away from military work. It has successfully competed for various assignments. Claude is used extensively within the War Department, as the Defense Department calls itself under Hegseth (Congress has not yet approved the name change). Claude is the only AI model that meets the strict requirements for use in secret operations, and it is already being used for this purpose. For example, Claude played an important role in the kidnapping of Venezuelan President Maduro.
Anthropic has gradually relaxed its usage rules in recent months. For example, the Pentagon may now use Claude for the development of lethal autonomous weapons, but not for their deployment.
Soldiers who ignore orders
CEO Amodei explained earlier in a podcast with columnist Ross Douthat from The New York Times what he is most afraid of. “The Constitution offers certain protection, but whether this is complied with depends in our armed forces on the idea that there are people who – hopefully – are prepared to defy illegal orders. With fully autonomous weapons we can no longer fall back on that.”
About mass surveillance, he said that this is not against the law and that you also have no right to privacy on the street. But with the many cameras in public spaces and the ability to record almost every conversation, a government can “connect and analyze all that kind of information with the help of AI. And then it can say: this person belongs to the opposition. And it can then map that out for a hundred million people.”
It is unacceptable for the Pentagon that a private company would decide which resources can be used militarily and in what way. Restrictions on this should come from Congress or the government, Hegseth and the military believe, not from the business community.
The harsh threats with which Hegseth tries to persuade Anthropic are not without risk for the Pentagon. If Amodei does not improve, Hegseth said on Tuesday, it could cost the company the $200 million contract. That would be a serious, but not fatal, setback for Anthropic. It is also problematic for the Pentagon, because it will lose its most important and advanced supplier of artificial intelligence.
Golden opportunity for competitors
Hegseth told Amodei that he could also label his company as a “supply chain risk.” This means that all companies that do business with the Pentagon must from now on demonstrate that they do not use Anthropic services. Such a measure is usually used against foreign companies from hostile countries, for example, suspected of espionage. For Anthropic it would mean a potentially fatal loss of turnover – and a golden opportunity for its competition.
As a third possible punitive measure, the Pentagon can take a completely different direction, Hegseth made clear: by not damning Anthropic but rather binding it and obliging it to cooperate obediently on the basis of a law from the Cold War.
The last two measures taken by the Pentagon against Anthropic in particular could discourage innovative projects from AI companies. Fear that even experts who share Hegseth’s view that the military should be able to use AI models as they wish.
