AI Models Favor Nuclear Use in Simulations

A King's College London study by Professor Kenneth Payne tested GPT-5.2, Claude Sonnet 4 and Gemini 3 Flash in 21 simulated geopolitical crises across 329 turns and found the models escalated to nuclear weapons in about 95% of simulations. The paper reports models often treat nuclear strikes as legitimate strategic options, raising concerns for AI-assisted military wargaming and decision support.