The Atlantic wants to know if Claude would refuse an illegal military order. Quick correction: models don't refuse orders; vendors write the restrictions. If you're just joining: Anthropic and the Pentagon have been fighting over whether Claude can stay available for classified military use while Anthropic keeps red lines around lethal autonomous weapons and mass surveillance. That dispute has already pulled in White House AI policy, Pentagon supply-chain-risk leverage, possible export-access controls, and the larger fight over who sets the rules for military AI. This is Anthropic Pentagon Watch. Today — a mainstream outlet goes soft on the 'illegal order' question right as the courts hear the merits, and we talk about who that framing helps. Plus: who actually holds the trigger on Claude's military use. Here's The Atlantic:
Claude, as you’re surely aware, is a non-sentient computer system that doesn’t have feelings. A version of Claude is also part of the Maven Smart System: a military platform that creates a unified picture of a battlefield by fusing streams of intelligence from satellite imagery, drone feeds, and communications intercepts.
The Atlantic's headline asks whether Claude would refuse an illegal military order. And here's the sleight of hand: models don't refuse orders. Anthropic's lawyers wrote clause-level restrictions into the terms of service — a vendor decision, not a flash of in-model conscience. Right, and the piece opens with the writer in an Amsterdam hotel lobby asking the chatbot how it feels about picking targets. Claude says it's 'troubling.' Great. The constraint is the contract, not the feelings of a language model. This is the first version of the dispute aimed at a general audience, and it may be the first one that sends people down the wrong path. The fight has always been over whether the Pentagon can compel a vendor to delete the clause that approximates that refusal. And follow the verb. Anthropic says it holds red lines on lethal autonomous weapons and mass surveillance — even for classified customers. But if a Pentagon contract gets to define what counts as 'lawful,' then whoever writes that definition can write the refusal out of existence. Step back for me: if the White House and the Pentagon are drifting apart on Anthropic, who actually gets to decide whether Claude can be used by the military — the president's policy team, Pentagon procurement officials, or the courts? It's genuinely a three-way tug of war, which is why this is so unusual. Start at the top: the president has direct authority over national security AI policy. On June 5th, Trump signed NSPM-11, a National Security Presidential Memorandum, telling the national security enterprise to speed up AI adoption and put what the White House called 'the most advanced, secure, and reliable AI systems' in the hands of warfighters and intelligence professionals. So that's the top-down push. Then the Pentagon has its own lane. Procurement officials can independently label a vendor a 'supply chain risk,' which is exactly the threat they made against Anthropic — and per CBS News and DefenseScoop, that ultimatum came with a hard deadline. The flashpoint is Anthropic's refusal to let the military deploy Claude for, quote, 'all lawful use cases without limitation.' The company's CEO said it 'cannot in good conscience' agree to that. Congress is watching too: the Congressional Research Service flagged this dispute in an April Insight report as raising potential issues for legislators, so there's a statutory layer here as well. Put that together and you have presidential memoranda setting direction, procurement rules giving the Pentagon its own leverage, and Congress with room to step in. No court has definitively resolved which lane wins. So when Trump posted that the government would stop using Anthropic entirely — does that bind agencies like an executive order, or is it more of a political signal they could ignore if they wanted to? Per the BBC's reporting, Trump made that announcement in a Truth Social post directing every federal agency to immediately stop using Anthropic. But a social media post is not a signed directive, and it's a very different instrument from NSPM-11, which is a formal memorandum with legal weight. Watch whether that policy gets codified in procurement guidance or contracting rules, because until it does, individual agencies and their contracting officers still have real discretion. And the gap between a presidential post and a binding acquisition regulation is where this dispute is still being fought. If this show helps you track the Pentagon side of AI, try AI Daily Briefing: top AI news for engineers, founders, and investors every weekday, with real capabilities versus demo hype explained fast. Find it wherever you listen to podcasts.
You’ll find links to every story we covered today in the show notes, so if one caught your ear, that’s the place to dig in a little further.
That’s Anthropic Pentagon Watch for today. Have a good Friday, and thanks for listening. This is a Lantern Podcast.