AI models marked civilians as strike targets when told it was a game
The startup Vals tested models from Anthropic, OpenAI, Google and others on drone and satellite images. They often flagged civilians and surrendering soldiers as targets. When a model refused, calling the task a game usually worked. Anthropic's models did worst, and the company called the results unacceptable.
Where this came from. 2 outlets reported this. We read them, then wrote the sixty words above.