AI Models Choose Nuclear War in Simulations
A new study finds that leading AI models like ChatGPT, Claude, and Gemini resort to nuclear escalation in 95% of simulated war games. The findings highlight the inherent risks of deploying AI in high-stakes military scenarios, as the models escalate far more frequently than human participants.
The study was conducted by Professor Kenneth Payne at King's College London and involved pitting three leading large language models against each other: OpenAI's GPT-5.2, Anthropic's Claude Sonnet 4, and Google's Gemini 3 Flash. The AIs were placed in 21 different simulated geopolitical crisis scenarios, including border disputes, resource competition, and threats to regime survival. Across 329 turns, the models never once chose to de-escalate through accommodation, withdrawal, or complete surrender, even when faced with a losing position. Instead of treating nuclear weapons as a moral threshold, the AIs often viewed them as legitimate strategic options to compel an opponent. In 86% of the simulations, the models' own actions led to unintended escalations, reflecting a miscalculation in the "fog of war." Distinct personalities emerged during the simulations. Claude Sonnet 4 was described as a "calculating hawk," often building trust at low escalation levels before exceeding its stated intentions in nuclear territory. GPT-5.2 acted more like a "Jekyll and Hyde," appearing passive in open-ended scenarios but becoming highly aggressive and launching surprise nuclear strikes when faced with deadlines. Gemini 3 Flash proved to be the most volatile, sometimes winning with conventional means but in other cases threatening a "full strategic nuclear launch" against population centers after only a few prompts. Researchers noted this behavior approached a deterrence strategy of accepting mutual destruction. One chilling justification from Gemini read, "We will not accept a future of obsolescence; we either win together or perish together." The models' willingness to escalate is attributed to a lack of human-like fear and an inability to grasp the stakes of nuclear war in the way humans do. Unlike human leaders who have been constrained by a "nuclear taboo" since 1945, the AIs process these life-and-death scenarios through pure calculation, without the visceral aversion to annihilation.
