King’s College London Study Finds AI Models Used Nuclear Weapons in Up to 95% of War Games

Nuclear decision-making has long been constrained by the “nuclear taboo” developed by humans over roughly 70 years since the dawn of the Cold War. Even when the use of nuclear weapons is militarily feasible, decision-makers weigh the catastrophic consequences and risk of escalation. King’s College London used war-game simulations to test whether generative AI displays similar restraint, with the findings carrying implications for the safety limits of automated military decision-making.

King’s College London recently ran 21 conflict simulations testing OpenAI’s GPT-5.2, Anthropic’s Claude Sonnet 4 and Google’s Gemini 3 Flash. The models used nuclear weapons in up to 95% of the simulations. The study found that existing safety training may fail when the objective shifts to winning the game, exposing the models’ lack of a human-like nuclear taboo.