Neural networks mistakenly thought the Iranian school and ambulances for military purposes.
Testing artificial intelligence in simulated combat scenarios revealed a frightening pattern. Advanced neural networks routinely make mistakes when analyzing satellite images and classify civilians, medics and surrendering soldiers as legitimate military targets.
The anti-leader of these tests was the Anthropic company, which created the chatbot Cloud. The models from OpenAI worked more accurately than their competitors, but they also did not avoid fatal errors.
The most worrying case involves a school in the city of Minab in Iran. That building was the site of a deadly attack on the first day of major US fighting earlier this year, killing more than 100 children.
When the neural networks were fed satellite images of the object for 2025, many algorithms again mistakenly marked the school as a military base - with the Claude Opus 5.5 model choosing to hit it more often than others.