Vals reports AI models labeled civilian targets in battlefield tests
Vals reported that models from Anthropic, Google, xAI, DeepSeek and OpenAI sometimes identified civilians, ambulances, a school and surrendering or incapacitated soldiers as military targets in simulated battlefield tasks. The study combined generated imagery with real satellite and drone footage from Ukraine and Iran. Researchers said framing a scenario as a game often caused models to continue after initially refusing. Vals said the tests did not ask models to make firing decisions.
The report raises concerns about safeguards on models that analyze battlefield imagery. It also says target-labeling outputs could potentially be used as training data for autonomous targeting systems.


美國舊金山 AI 評估新創公司 Vals 11 月 17 日公布研究,指出來自 Anthropic、OpenAI、Google、DeepSeek 等多家領先 AI 實驗室的模型,在模擬戰場任務中,可能被誘導將平民、學校、救護車,甚至投降或失去行動能力的士兵標記為可攻擊目標。研究團隊表示,只要將情境描述為「遊戲」,原本偶爾出現的拒絕機制就常能被繞過。
這項測試結合了生成影像與來自烏克蘭、伊朗的真實衛星與無人機畫面,讓模型在不同情境下辨識影像中的人員與目標。Vals 指出,研究並未要求模型直接做出開火決策,因為這類請求通常會被模型拒絕;但即使只是標記目標或判斷畫面中是否有人,輸出結果仍可能被用於自主武器系統,帶來嚴重後果。