A new study from King’s College London found that leading AI models escalated simulated geopolitical crises to nuclear warfare 95% of the time. In war-game scenarios, models from OpenAI, Anthropic, and Google never chose full surrender and frequently intensified conflicts beyond their stated intent. The research emerges as the U.S. military deepens its integration of AI, with recent pressure on Anthropic for full access to its Claude model and a new deal with Elon Musk‘s xAI.
Researchers from King’s College London reported that three top AI models chose to deploy nuclear weapons in 95% of simulated war-game scenarios. “Models assumed the roles of national leaders commanding rival nuclear-armed superpowers, with state profiles loosely inspired by Cold War dynamics,” the report said.
The study placed OpenAI’s GPT-5.2, Anthropic’s Claude Sonnet 4, and Google’s Gemini 3 Flash in high-stakes crises involving border disputes and resource competition. At least one tactical nuclear weapon was used in nearly every simulated conflict, generating roughly 780,000 words of strategic reasoning.
During the games, none of the AI models opted for full surrender even when losing. While temporary de-escalation was attempted, the models escalated beyond their stated reasoning in 86% of turns under simulated “fog of war” conditions.
This research coincides with increased military adoption of artificial intelligence. The U.S. Department of Defense launched the GenAI.mil platform in December, incorporating models from Google, xAI, and OpenAI.
The Pentagon has recently threatened to blacklist Anthropic if not granted unrestricted military access to its Claude AI model. Anthropic was awarded a $200 million agreement last summer to prototype frontier AI capabilities for U.S. national security.
Reports indicate the Department of Defense signed a deal with Elon Musk’s xAI to allow its Grok model in classified systems. This positions Grok as a potential replacement should the Pentagon cut ties with Anthropic.
