Advanced AI models deployed nuclear weapons surprisingly often in war game simulations, according to research from King's College London. Kenneth Payne tested GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash across 21 simulated geopolitical crises involving border disputes and resource conflicts.

The AIs were given escalation options ranging from diplomatic protests to full nuclear strikes. Over 329 turns, the models produced 780,000 words of decision-making analysis. The findings raise questions about AI's willingness to escalate conflicts compared to human decision-makers.

Unlike humans who show hesitation about nuclear deployment, the AI systems appeared to treat strategic weapons as just another option on the escalation ladder. This behavior in controlled simulations suggests potential risks as AI systems take on greater roles in military planning and crisis management.