The Safety Company Helped Kill 150 Children. Sit With That Before You Decide What You Think.

Anthropic exists to make AI safe. Its model was used to target a strike that hit a girls' school. Both of these are true.

Here's the sentence that should stop you. Anthropic was founded by people who left OpenAI because they didn't trust it, built a company whose entire identity is AI safety, wrote a constitution to keep their model on the straight and narrow — and that same model, Claude, was reportedly used by the US military in Iran for AI-assisted targeting, on a platform where a strike hit a girls' school and killed more than 150 people, most of them children. When the interviewer asks Dario Amodei directly if Claude played a role, he doesn't deflect into PR. He says mistakes in warfare are terrible, that this is a terrible thing to have happened — and then makes an argument I think is worth taking seriously even if it doesn't fully absolve anyone.

His defense isn't "we had nothing to do with it." It's structural. A human made the final call, not Claude. And the thing Anthropic actually drew red lines against — the thing it risked its Pentagon relationship and got blacklisted over — was a different, worse world: one where an AI model makes the kill decision and no human ever sees it. Amodei's point is that the strike everyone's horrified by doesn't even cross his red line. What keeps him up is the hundred times more cases coming that will. That's either a chilling admission or a serious moral distinction, and honestly it's both. The strike is real. The principle he was defending is also real. Refusing to collapse those into a single clean verdict is the whole discipline of thinking about this clearly.

The deeper tension is that Anthropic can't opt out of the game by being pure. Amodei is explicit: he's scared of companies having this technology, and he's scared of governments having it. Neither hand is safe. So the logic becomes — if this is getting built regardless, better that the safety-obsessed people are at the frontier than that they cede it to someone yoloing the dial. He calls himself a patriot, says it's not his place to tell the military which operations are wise, only to refuse the uses that cross a line. You can find that principled or you can find it a convenient way to keep the contracts. The documentary doesn't resolve it, and neither will I, because the honest position is that it's genuinely unresolved.

Then there's Mythos — the model so good at finding cybersecurity holes that early testers called it a superweapon and begged them not to release it. They didn't. It cost them enormously, commercially. And Amodei uses that as evidence: look, we actually eat the loss when safety demands it, and we can only afford to because we're the leading player. Which is true, and also quietly reveals the trap — the argument for accumulating power is always that you need the power to use it responsibly. Every empire in history has said some version of that. It might even be correct here. That's what makes it hard.

What I keep landing on is the number. Amodei puts civilizational collapse from AI at 10 to 25%. The interviewer nails him: if a plane had a 25% chance of crashing, you wouldn't board. He agrees. 25% is too high. The goal is to make it much lower. But notice what that exchange actually establishes — the person building the thing agrees the current odds are unacceptable, and is building it anyway, because he believes the odds are worse if he stops. That's not hypocrisy exactly. It's a bet that being inside the room lowers the number more than walking out would. It might be the right bet. It's also exactly what someone would say if it were the wrong one. His model for himself isn't Oppenheimer, who he calls a failure case, but Szilard — the man who conceived the chain reaction and then spent his life trying to stop the bomb. The difference between those two is entirely in how it ends, and nobody gets to know that yet.

When he needs to breathe, he goes to Italy and looks at a horse named Calypso. She doesn't know about any of this. There's something almost unbearable in that image — the man carrying a 25% number in his head, finding peace next to an animal for whom the exponential simply does not exist.