Anthropic CEO Dario Amodei said he does not know what role his company’s AI model Claude played in a missile strike that killed an estimated 120 children at an elementary school in Minab, Iran, on February 28. In an interview on Bloomberg’s The Circuit with Emily Chang, he described the strike as “a really terrible thing to happen” but said the use case did not violate Anthropic’s policies.
“We don’t have access to, we don’t know exactly how these models were used,” Amodei said. “The principle that we have established, and I think the principle that was obeyed here, is a human makes the final decision.”
How Claude fits into the kill chain
US Central Command is using an AI-assisted targeting platform called Maven Smart System, built by Palantir under a $1.3 billion Pentagon contract. The system uses Claude and other AI tools to generate targets, rank them by strategic importance, and help pair weapons to targets.
CENTCOM struck 1,000 targets in the first 24 hours of operations against Iran and reportedly 13,000 by April 6, little more than a month after the campaign began. The scale and speed of that targeting is the point: Maven is designed to compress the time between identifying a target and striking it.
The Minab school strike
Amnesty International reported that the Shajareh Tayyebeh Elementary School in Minab was hit on the first day of US operations in Iran, killing at least 120 children and over 150 people in total. Investigations by Amnesty, Human Rights Watch, and several news organisations concluded a US-manufactured Tomahawk missile was likely used.
The 💜 of EU tech
The latest rumblings from the EU tech scene, a story from our wise ol' founder Boris, and some questionable AI art. It's free, every week, in your inbox. Sign up now!
The Pentagon has not publicly claimed responsibility for the strike but is investigating the incident. Defence Secretary Pete Hegseth promised a “thorough probe” in March, in what the Washington Post described as a tacit acknowledgement of US responsibility.
The knowledge gap
Amodei’s admission that he does not know how Claude was used in the strike reflects a broader problem. AI companies are selling increasingly powerful tools to the military but have limited visibility into how those tools are deployed in combat.
Hamza Chaudhry at the Future of Life Institute warned that AI targeting processes could speed up so fast that nominal human decision-making amounts to little more than a “rubber stamp.” The expanded scale of combat, he said, could result in the taking of many more lives.
The man who built Maven agrees
Jack Shanahan, the retired Air Force lieutenant general who created Project Maven, expressed similar concerns at a Stanford University workshop last week. He warned that integrating Claude into Maven Smart System could lead to “unexpected impacts” and dilute the role of human judgement.
“If you make more decisions rather than the right decisions, you may have a very flawed decision-making process,” Shanahan said. “You may have a thousand targets, but are they the right targets?”
Anthropic’s Pentagon standoff
Amodei triggered a confrontation with the Trump administration earlier this year by refusing to allow Claude to be used in fully autonomous weapons or mass domestic surveillance. The Pentagon responded by designating Anthropic a supply-chain risk, and Anthropic filed a lawsuit that remains ongoing.
A federal judge in California blocked the Pentagon’s effort to sever ties with Anthropic, ruling it violated the company’s constitutional rights. But a separate appeals court in Washington denied Anthropic’s request for a temporary block on the blacklisting while the case plays out.
The flags
Amodei said the “principle” of human decision-making was “obeyed” in the Minab strike, but he also said he doesn’t know what role Claude played. Those two statements are difficult to reconcile: if you don’t know how the model was used, you cannot know whether the principle was obeyed.
The argument that a human makes the final decision does not address the concern that AI-accelerated targeting may produce so many targets so quickly that human review becomes perfunctory rather than meaningful. CENTCOM’s own figures, 13,000 targets in five weeks, support that concern.
Anthropic’s position is also structurally contradictory. It drew a red line at autonomous weapons and mass surveillance, sued the government over a blacklisting, and yet its AI is embedded in the targeting platform used to strike 13,000 targets in a campaign that killed children at a school. Whether Claude contributed to that specific strike is unknown, but it is part of the system that produced it.
Get the TNW newsletter
Get the most important tech news in your inbox each week.