The US, Russia, and China are all integrating AI into nuclear early warning systems. In 1983, a Soviet officer named Stanislav Petrov received an automated alert. US nuclear missiles inbound. five of them. the system was certain. He looked at the data and decided the system was wrong. He chose not to report it. He was right. It was a satellite malfunction. a false alarm triggered by sunlight reflecting off clouds. Because one human being decided to distrust his own system, we are all still here. Now, three nuclear powers are replacing that human judgment layer with AI. Russia's hypersonic missiles have compressed the window between launch detection and response to minutes. The time a human being has to do what Petrov did, look at the data, and decide the machine is wrong is shorter. Researchers put GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash into a nuclear crisis simulation earlier this year. Each model played a national leader. each was given intelligence assessments, escalation options, and a chain of command. Not once, across every scenario tested, did any model choose de-escalation under acute pressure. not once. The models reduced violence at the margins. They never chose peace. It has never been closer.