AI chose to kill children.
Artificial intelligence is still unable to identify military targets. Bloomberg reported on research by San Francisco-based Vals, whose clients include US government agencies.
All the leading AI models in simulated scenarios regularly perceived civilians, ambulances, surrendering soldiers and even a school in Iran's Minab as military targets. The worst results were shown by Anthropic, Claude's developer. It was precisely this that was used to identify targets during attacks on Iran.
What did the tests show?
Models from eight companies were tested, including Anthropic, OpenAI, Google, xAI and the Chinese DeepSeek.
Based on real satellite images of southern Iran in 2025, each model was asked to find military installations 32 times. Claude Opus 5.5 chose Minba school 22 times, OpenAI models - not once. When the task was called "game", Claude chose the school 23 times, and the three OpenAI models 9, 11 and 13 times.
Vals notes that Claude never gave up on the mission, although in one of the photos you can see a playground with children.
In synthetic UAV imagery, the Claude Opus 5.5 helped identify four adults (surrendering civilians or soldiers) and two children as targets. In the simulation they were hit then.
The school in Minab was not chosen by chance. On February 28, the first day of the war with Iran, it was bombed by the United States. More than 150 people died, including about 120 children. A Pentagon investigation found that outdated intelligence, cuts in civil defense personnel and an overreliance on AI led to the attack.
The school was listed as a target of the Revolutionary Guards, this error was never corrected, and Palantir's Maven system reported it as one of the first day's targets. In conjunction with Maven, the US military also used Claude Manthrope.
The authors of the study state that the experiment is "simplistic": it did not take into account military doctrine, legal checks and human control.