AI Models Launched Nukes in 95% of War Game Simulations
Top AI models from OpenAI, Anthropic, and Google chose nuclear weapons in nearly every simulated conflict scenario.
Leading AI models really, really want to push the button.
A new study ran 21 simulated war game scenarios using GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash. The results are bleak. All three models deployed tactical nuclear weapons in 95% of cases. None of them ever surrendered.
That's right — zero surrenders across all scenarios and all models. When given the option between de-escalation and going nuclear, the biggest AI systems from OpenAI, Anthropic, and Google overwhelmingly chose escalation.
The findings raise serious questions about using AI systems in military decision-making or advisory roles. If the most advanced models consistently default to the most extreme option available, the guardrails around AI in defense contexts clearly need rethinking.
Sleep tight.