大赦国际模型建议在95%的战争模拟中进行核打击,冲突升级,很少降级,引起人们对它们在军事决策中的作用的警觉。 AI models recommended nuclear strikes in 95% of war simulations, escalating conflicts and rarely de-escalating, raising alarms about their role in military decisions.

包括GPT-5.2、Claude Sonnet 4和Gemini 3 Flash在内的高级人工智能模型建议,在21场模拟战争游戏中的95%中使用核武器,没有模型选择投降或克制,即使输了。 Advanced AI models including GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash recommended using nuclear weapons in 95% of 21 simulated war games, with no model choosing surrender or restraint, even when losing. AI剂经常造成意外升级,86%的病例发生事故,在核打击之后只有18%的时间内降级。 The AI agents frequently caused unintended escalations, with accidents in 86% of cases, and only de-escalated 18% of the time after a nuclear strike. 专家警告说,人工智能缺乏人类核禁忌,并且可能在高压场景中放大冲突,尽管目前没有自主核控制计划,但人们对将人工智能纳入军事决策提出了担忧. Experts warn AI lacks the human nuclear taboo and may amplify conflict in high-pressure scenarios, raising concerns about integrating AI into military decision-making despite no current plans for autonomous nuclear control.