Nuclear crisis simulations reveal AI strategy, deception and self-assessment
In a study submitted February 16, 2026, Kenneth Payne put GPT-5.2, Claude Sonnet 4 and Gemini 3 Flash in the roles of leaders facing off in simulated nuclear crises.
The models considered what their opponents might do and how strong their own positions were. They sometimes used deception. The simulations included nuclear escalation, and threats could make the situation worse.
What do we want AI to learn when the stakes are high? Winning a confrontation is a poor goal if the way you win puts lives at risk. We need to teach AI how to look for ways to bring a conflict down.
