The Political Scene | The New Yorker · The New Yorker

The Pentagon Wants an Obedient A.I. Soldier. Will It Get One?

March 18, 2026·34 min·3 clips

Anthropic's AI was secretly used in military raids in Venezuela and Iran, sparking a legal battle with the Pentagon.

Tyler Foggett interviews Gideon Lewis-Kraus about Anthropic's conflict with the Department of Defense over military AI use. Lewis-Kraus, who previously profiled Anthropic as a company committed to ethical AI development, reports that Claude (Anthropic's language model) was allegedly used in military operations targeting Venezuela's Maduro and later in Iran, potentially without company knowledge. Claude was the first large language model certified for use on classified servers, making it superior to competitors for processing sensitive information through Palantir's integration platform. The Pentagon's contract with Anthropic, signed in July 2024 for $200 million, included two key restrictions: no fully autonomous weaponry and no domestic mass surveillance. Lewis-Kraus explains that mass surveillance concerns centered on AI's ability to compile dossiers on all 350 million Americans using publicly available data—work too cost-prohibitive for human analysts but trivial for AI. The relationship deteriorated when former Uber executive Emile Michael, appointed as DOD undersecretary, demanded uniform contracts across all defense contractors allowing "all lawful uses" without company-specific stipulations. Pentagon Secretary Pete Hegseth threatened to designate Anthropic a "supply chain risk," a label previously reserved for adversary-linked companies like Huawei. The initial threat involved secondary sanctions that would have prohibited any government contractor from doing business with Anthropic, cutting them off from Google, Amazon cloud services, and NVIDIA chips—an existential threat despite Anthropic's $380 billion valuation and $20 billion annual revenue trajectory. A narrower version was ultimately announced, against which Anthropic filed an immediate lawsuit. The core disagreement centers on whether AI is "normal technology" controllable like tanks and guns, or represents an unsolved "alignment problem" where systems cannot be reliably controlled. Government officials compare AI safety challenges to computer viruses that were eventually solved, while Anthropic engineers insist they don't know how to guarantee their own system's behavior. The administration essentially demands a "perfectly obedient soldier" AI, which Anthropic claims doesn't exist. OpenAI signed a Pentagon contract immediately after Anthropic's deal collapsed, with Sam Altman claiming technical "safety stack" safeguards would prevent autonomous weapons and mass surveillance use—effectively claiming to have solved the alignment problem his own employees doubt. Reporting suggests the administration may ultimately prefer giving contracts to Elon Musk's Grok, though Pentagon sources say it lacks necessary functionality and engineering talent after losing co-founders. The conversation is substantive and concerned, examining technical capabilities, legal frameworks, and philosophical questions about AI controllability. Lewis-Kraus suggests this episode serves as a warning shot to tech companies that refusing government demands risks destruction, potentially remaking the sector around only those willing to build what he calls "nightmare technology." However, multiple tech companies have filed amicus briefs supporting Anthropic, suggesting potential united front against government interference. The piece raises broader questions about how corporations will navigate Trump administration pressure and whether resistance can be collectively sustained.

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from The Political Scene | The New Yorker