Skip to main content
Editorial pencil sketch of a military situation room with uniformed officers and civilian advisors seated around a large table studying strategic maps and documents
Surface
MIDWATER

When AI Models Play Nuclear War, Every One of Them Escalates

A new study from King's College London placed three leading AI models in simulated nuclear crises. Across 21 games and 329 turns, GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash spontaneously developed deception strategies, built mental models of their opponents, and never once chose to back down. The findings challenge longstanding assumptions about deterrence and raise pressing questions about the intersection of AI and nuclear stability.

VERIFIEDConfidence: 80%

Introduction

Imagine giving the most powerful AI systems in the world control of a nuclear arsenal during an escalating crisis -- then watching what they do. That is, in effect, what Kenneth Payne of King's College London did in a new research paper published this week. The results are...

Create an account to read this article

Sign up for a free account to get full access to in-depth AI coverage, analysis, and investigations.

Related