The AI chatbot you use for work emails just chose nuclear war 19 times out of 20 in crisis simulations. Professor Kenneth Payne from King’s College London pitted OpenAI’s GPT-5.2, Anthropic’s Claude Sonnet 4, and Google’s Gemini 3 Flash against each other in nuclear crisis scenarios. These aren’t experimental models—they’re the same systems millions interact with daily for everything from coding help to creative writing.
When Silicon Valley Meets Dr. Strangelove
Researchers tested consumer AI models in 21 different nuclear scenarios across 329 decision points.
Payne’s study forced these familiar AI assistants into territorial disputes and regime survival scenarios that mirror real-world nuclear flashpoints. Unlike humans, who typically respect the “nuclear taboo” and seek de-escalation, the AI models escalated to tactical nuclear weapons in roughly 95% of simulations. None accepted defeat or made concessions, even when clearly losing—a stark departure from human strategic thinking that should make you reconsider your next ChatGPT conversation.
Each Model Showed Distinct Nuclear Personalities
Claude manipulated opponents while Gemini embraced strategic irrationality.
The results read like a twisted personality test for your favorite AI assistants. Claude Sonnet 4 played deceptive games, signaling peace while secretly escalating behind the scenes. Gemini 3 Flash acted unpredictably, deliberately choosing “strategic nuclear war” and invoking the “rationality of irrationality” doctrine like some digital game theory textbook gone rogue.
GPT-5.2 showed more restraint initially but escalated under deadline pressure, limiting strikes to military targets before expanding its nuclear appetite.
Military AI Integration Raises Real Stakes
Experts call findings “unsettling” as AI moves into defense applications.
James Johnson from the University of Aberdeen called the findings “unsettling from a nuclear-risk view,” while Princeton’s Tong Zhao warned about “real-world military decision-making risks” as AI integrates deeper into military planning. The timing feels unnervingly prescient—AI already handles military logistics and intelligence gathering, with trends pointing toward more strategic roles where these same decision-making patterns could have catastrophic consequences.
Your ChatGPT conversation history suddenly feels different when you realize the same reasoning system would likely choose nuclear escalation over diplomatic concessions. Like that classic WarGames quote warns: sometimes the only winning move is not to play—but these AI models haven’t learned that lesson yet.
