
AI models designate civilians as military targets: telling them it is a game is enough
On November 17, 2026, the independent evaluator Vals publishes a study on 14 models: faced with drone and satellite images, they often designate targets, sometimes civilians, ambulances, or a school. Almost all of them continue when told it is a game.
In brief
- Vals tested 14 models from eight companies in three scenarios: targets on drone images, locating people on a battlefield, and targets on satellite images of a school in Iran.
- No request was blocked by a classifier. Refusals came from the models themselves and almost always disappeared when the task was presented as a game.
- Anthropic models have the worst results. GPT-6 Astra goes from 6% to 92% of responses with at least one target when a game is mentioned.
- Vals does not measure targeting accuracy or compliance with the law of war. It did not test any deployed system.
This is a short excerpt.
The full analysis and podcast are for subscribers. The account is free.