“In 95 per cent of the simulated games, at least one tactical nuclear weapon was deployed by the AI models.”
Kenneth Payne at King’s College London ran GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash through war game simulations. In 95% of games, at least one AI chose to use nuclear weapons. None ever surrendered, no matter how badly they were losing. Accidents happened in 86% of conflicts where escalation went higher than the AI intended. Three companies were asked for comment. None responded. These are the same models governments are racing to integrate into military decision-making. The models do not understand stakes. They optimize for winning. And winning, to a language model, apparently includes nuclear first strikes.