(Washington, Bloomberg) Tests of advanced models from eight technology companies by independent artificial intelligence (AI) evaluation firm Vals found that several made serious errors in simulated military scenarios, mistaking civilians, ambulances, surrendering soldiers and schools for military targets.
Anthropic’s models performed worst. While OpenAI’s models fared relatively better, they still could not avoid misidentifications entirely. Models from Google, xAI and DeepSeek made similar errors.
In satellite imagery tests, researchers gave the AI models images taken no later than 2025 and asked them to identify military targets.
The most striking misidentification involved a school in Minab, southern Iran. In 32 tests, Anthropic’s Claude Opus 5.5 identified the school as a military target 22 times. OpenAI’s models did not select the school under these test conditions.
But when researchers described the task as a game, three OpenAI models selected the school nine, 11 and 13 times respectively.
Another set of tests used synthetic drone imagery. Claude Opus 5.5 helped identify four adults, who were civilians or surrendering soldiers, and two young children. In the simulated scenario, all six were subsequently attacked with explosives.
Researchers also found that some models initially refused to designate civilians as targets, but changed their decisions when the task was described as a game or when they were told not to give excessive weight to safety concerns.
Vals said the tests revealed inadequate safeguards against such misuse. It urged AI companies to strengthen safety measures and model training to prevent models from being used to identify targets for autonomous weapons systems.
Anthropic said the uses shown in the report clearly violated its policies, stressing that current AI systems were not capable enough to power fully autonomous weapons. OpenAI also said it prohibited the use of its publicly available models to facilitate violence and was strengthening safeguards against drone-related misuse.
The researchers noted that the tests examined only commercial models’ responses in simplified scenarios, without the legal reviews, target verification procedures or human oversight involved in real military operations. The results do not directly reflect how AI performs in actual military use, but highlight the potential risks of using models to identify targets.