"State of play: The two sides are heading into the meeting on two totally different pages. An Anthropic spokesperson said: "We are having productive conversations, in good faith." Defense officials say negotiations have shown no progress and are on the verge of breaking down."
"Anthropic is willing to loosen its existing usage restrictions, but wants to wall off two areas: the mass surveillance of Americans, and the development of weapons that fire without human involvement. The company "is committed to using frontier AI in support of US national security," the spokesperson said. The Pentagonsays it's unduly restrictive to have to clear individual uses with the company, and has demanded that all AI labs make their models available for "all lawful uses.""
"Setting the scene: Amodei has been very vocal about the risks of AI-gone-wrong, and has positioned his company as the safety-first AI leader. Officials have described a culture clash between Hegseth's brash Pentagon and the Silicon Valley firm. The senior Pentagon official said: "The problem with Dario is, with him, it's ideological. We know who we're dealing with." Reality check: Beyond the personalities that will sit across from each other on Tuesday, there are deeper questions about the role AI can and should play in national security."
Anthropic and Pentagon negotiators are misaligned over permitted uses of Anthropic's Claude model. Anthropic proposes loosening some restrictions while explicitly barring mass surveillance of Americans and autonomous weapon development. The Pentagon rejects case-by-case clearance and insists all AI labs make their models available for all lawful uses. Pentagon officials have threatened to deem Anthropic a supply-chain risk, which could void contracts and force partners to certify they do not use Claude. Offboarding Anthropic would be complex because the company is deeply integrated and currently offers superior capabilities, and the dispute exposes broader legal and policy gaps about AI in national security.
Read at Axios
Unable to calculate read time
Collection
[
|
...
]