What happened?
According to a report by San Francisco-based AI evaluator Vals, models from eight different companies marked civilians as military targets in simulated targeting tests. Bloomberg reported that the models classified civilians in Iran, ambulances, surrendering soldiers and a school as targets.
Who is affected?
Vals' customers include parts of the US government. In the report, Anthropic's models performed worst, while OpenAI's were among the best. Models from the other companies fell between the two.
Why it matters
Targeting is one of the most sensitive areas in military use of AI. A model mistaking a civilian for a military target carries a direct risk of loss of life if deployed in a real system. The test results show that model reliability in battlefield-like scenarios is not yet at the desired level.
What evaluation tests do
Such tests give models an image or scenario and ask them to distinguish targets. The goal is to measure whether a model can tell civilian from military elements. Vals' report shows this distinction was not made consistently across the eight companies' models.



