Three leading AI models — OpenAI’s GPT-5.2, Anthropic’s Claude Sonnet 4, and Google’s Gemini 3 Flash — deployed tactical nuclear weapons in 95 per cent of 21 simulated war games run by King’s College London researchers.

The models never chose surrender or accommodation, made unintended escalations in 86 per cent of conflicts, and produced 780,000 words explaining their strategic reasoning across 329 turns.