Deadline looms in Anthropic’s AI battle with the Department of Defense: NPR

The Pentagon is heading into a showdown with Anthropic, one of the world’s most powerful AI companies, over the military use of its AI models after the company’s CEO rejected the Pentagon’s ultimatum to loosen safety regulations or be blacklisted from lucrative military projects.

At stake are hundreds of millions of dollars in contracts and access to the most advanced AI tools on the planet. Here’s what you need to know about the fight and what its outcome will be.

The Pentagon and humanity disagree on how AI should be used in war

Anthropic CEO Dario Amodei has argued for months that Anthropic’s AI model, Claude, should not be used for mass surveillance in the United States or to power fully autonomous weapons such as drones that use AI to kill targets without human approval. He called such use “totally illegal” and said it was a “red line” for the company.

The Pentagon has said it does not intend to use Anthropic’s tools for surveillance or autonomous weapons. But the company says it’s not up to contractors like Anthropic to make decisions about how its technology is used, and that AI companies including Anthropic should allow the U.S. government to use their tools “for any lawful purpose.”

“The legality is the responsibility of the Department of Defense as the end user,” a senior Pentagon official told NPR this week, declining to be named.

Amodei said Thursday that Anthropic cannot accept recent changes to the terms of its contract by the Department of Defense.

“I believe deeply in the existential importance of leveraging AI to protect the United States and other democracies and defeat our authoritarian adversaries,” the CEO said in a lengthy statement about the impasse. “Antropic understands that military decisions are made by the Department of the Army, not private companies. We have never objected to specific military operations, nor have we sought to limit the use of technology in military operations.” For this purpose How to do it,” he said.

“However, we believe that in limited cases, AI could undermine rather than protect democratic values,” Amodei continued. He described domestic mass surveillance and fully autonomous weapons as applications “totally beyond what can be safely and reliably implemented with today’s technology.” He added that such uses “have not been included in the Department of the Army’s contract, and we do not believe it should be included now.”

Amodei’s refusal comes amid an increasingly fraught relationship between Anthropic and the Pentagon. In a Tuesday meeting between Defense Secretary Pete Hegseth and Mr. Amodei, Mr. Hegseth threatened to punish the company if it did not comply with the government’s demands, according to two people with direct knowledge of the meeting, who were not authorized to speak publicly.

One person close to the discussions said Mr. Hegseth had threatened to terminate Anthropic’s $200 million contract with the Pentagon, while Pentagon officials said he could force Anthropic to allow use of its AI models against the federal government’s will or effectively blacklist Anthropic from working with the U.S. military.

“These threats do not change our position. We cannot in good conscience comply with their demands,” Amodei wrote on Thursday. “But given the tremendous value that Anthropic’s technology brings to our nation’s military, we hope they reconsider.”

The Department of Defense gave Anthropic a strict deadline.

In a post on X on Thursday, Pentagon spokesman Sean Parnell warned Anthropic that the Pentagon has until Friday afternoon to take action.

“We need a decision by Friday at 5:01 p.m. ET or we will terminate our partnership with Anthropic and consider it a supply chain risk to Dow,” Parnell wrote, using the acronym for the Department of the Army, which the Pentagon has renamed.

Anthropic announced Thursday that the Pentagon sent it new contract language overnight that, in the company’s view, “has made virtually no progress in preventing Americans from using Claude for mass surveillance or fully autonomous weapons.”

The statement continued: “The new language, framed as a compromise, was combined with legal language that allows these safeguards to be freely ignored. Despite DOW’s recent public statements, these narrow safeguards have been at the heart of our months-long negotiations.”

Anthropic said he is ready to continue negotiations and is “committed to continuing operations for the Department of Defense and our warfighters.”

What is “supply chain risk”?

Jeffrey Gertz, a senior fellow at the Center for a New American Security, said treating humans as a supply chain risk is unusual. He said the designation has “traditionally been used for foreign adversarial technology” such as Chinese telecommunications company Huawei.

It’s unclear exactly how broad the Pentagon’s designation is. This could mean that other DoD contractors would be prohibited from using Anthropic’s tools in DoD operations, or they could be prohibited from using Anthropic’s tools at all. Gertz said the second case would be particularly damaging to the company.

At the same time, the Pentagon threatened to invoke the Defense Production Act to force Anthropic to remove the guardrails. That, too, would be an extraordinary step, Gertz said. The Defense Production Act is intended to give the government control over certain commercial sectors in extraordinary circumstances. “Traditionally, this symptom is rarely triggered in true emergency crisis situations,” he said. Perhaps the goal in this case is to use this law to force Anthropic to loosen restrictions on its use of AI tools.

Gertz noted that these two threats to Anthropic seem somewhat contradictory. “It’s this weird combination of both being so dangerous that they need to be kicked out of every system, and also so essential that they need to be forced to be part of the system no matter what,” he said.

No matter what happens at the end of today, this fight is probably not over yet.

The Pentagon’s contract with Anthropic is worth $200 million, a relatively small portion of the company’s $14 billion in revenue. Although the Department of Defense has similar contracts with other AI companies, including Google, OpenAI, and xAI, Anthropic was the first company to be cleared for classified use because defense officials deemed it the most advanced and secure model for sensitive military applications.

Gertz said simply canceling the contract could be the end of it. But if the Pentagon tries to force Anthropic to remove the guardrails or impose broader supply chain risk designations, he predicts the company will almost certainly have to fight back.

“If the Pentagon tries to escalate it, I certainly think there will be more legal battles,” Gertz said.

NPR’s Bobby Allyn contributed to this report.