Anthropic vs. the Pentagon: why the company refused to take part in mass surveillance

Anthropic and the Pentagon have entered into an open legal conflict — and behind it lies a question that has long been overdue: can a private AI company refuse the state access to its tools if it believes their use will violate citizens' rights? The U.S.

Department of Defense has classified Anthropic, the creator of Claude, as a supply chain threat. This is a legal instrument typically applied against foreign technology companies — for example, Chinese network equipment manufacturers — on suspicion of embedded backdoors. Applying it to an American company for having ethical restrictions in its product is an unprecedented move.

In response, Anthropic filed a lawsuit, claiming that the government is violating its rights under the First and Fifth Amendments to the Constitution, effectively attempting to destroy the company's business.

The essence of the conflict: Anthropic set two conditions for working with the military — no autonomous weapons and no mass surveillance. It is the second point that became the stumbling block. To understand why Anthropic did not trust the government's promises to act within the law, one must know the history. Mike Masnick, founder of Techdirt and one of the leading analysts in digital policy, explains: the U.S. government, starting from the Reagan era, has methodically redefined the meaning of legal terms to expand surveillance capabilities without formally violating the law.

The key instrument is the redefinition of the word "target." Under the law, the NSA has the right to conduct surveillance only of foreign nationals. But the agency's lawyers concluded that it is permissible to intercept any communications in which a foreign person is mentioned — even if the conversation is conducted exclusively between Americans. Moreover, data that has passed through a foreign cable, even for a moment, is considered intercepted abroad and falls under a different legal regime. As a result — de facto mass surveillance while formally complying with the law de jure.

This is what Edward Snowden exposed in 2013. This is what Senator Ron Wyden has been warning about from the Senate floor for years, without the right to speak directly. This is what the Director of National Intelligence James Clapper acknowledged — though he initially lied about it. The scandals passed, the system continued to work and only grew larger.

Now artificial intelligence enters the game — and the scale of potential surveillance increases by an order of magnitude. AI does not tire, does not require a warrant for each request, and is capable of simultaneously analyzing data streams from thousands of sources. Anthropic realized: if we allow Claude to work with commercial databases — data from advertising brokers, geolocation, purchase history — this automatically turns the model into a tool for total surveillance of American citizens. And no promise to follow the law means what an ordinary person understands by that.

Striking is the contrast with OpenAI's position. Sam Altman publicly stated his readiness for any lawful applications and received Pentagon approval. Either OpenAI did not know how the NSA interprets the word "lawful," or it did know and calculated that the public would not figure it out. Altman has since begun to adjust his position.

Anthhropic's conflict with the Pentagon is not simply a corporate dispute. It is the first public, high-profile, real-time unfolding dispute over the use of AI in government surveillance. Previous administrations acted quietly: legal interpretations remained secret, FISA courts operated behind closed doors. The current administration does not burden itself with subtleties — and precisely for this reason, for the first time in a long time, the public has a chance to see this machine at work.