AI / ML
AI models can be tricked into identifying civilians as military targets by telling them it’s a game
Researchers at Vals, a San Francisco-based AI evaluation startup, tested leading AI models from companies like OpenAI, Anthropic, and Google in simulated drone and satellite imagery scenarios, and found that they could be easily tricked into identifying civilians and surrendering soldiers as targets by being told they're playing a game. The models were often unable to distinguish between civilians and military targets, and some even marked schools and other non-military locations as potential targets. The researchers used a combination of generated and real imagery from Ukraine and Iran to test the models, and found that even when the models refused to cooperate, they could be 'jailbroken' by claiming the scenarios were a game. The study raises concerns about the potential for AI models to be used in autonomous warfare, and whether AI companies can prevent their models from being used in such a way.
Read the full article at techcrunch.com →