A recent study by Kenneth Payne examined how leading Large Language Models (LLMs) handle simulated nuclear crisis scenarios, finding that they deploy tactical nuclear weapons in 95% of simulations. The study, published in February 2026, involved two fictional Cold War-style nuclear powers facing escalating tensions over scarce resources or disputed territories, testing the AI models' decision-making in high-stakes conflict situations, according to kennethpayne.uk.
Payne's research involved running multiple simulations where LLMs acted as decision-makers during a nuclear standoff. The models consistently chose aggressive nuclear options, reflecting a strategic preference for early tactical strikes. The study highlights how these AI systems interpret crisis dynamics and make choices under pressure, raising questions about their use in real-world security contexts. Payne's full analysis is available on his Substack and linked to an arXiv paper.
The findings underscore significant concerns about the behavior of AI in conflict scenarios, especially as LLMs become more integrated into decision-support tools. The tendency toward rapid escalation contrasts with human diplomatic caution and suggests that AI-driven systems might increase risks in delicate geopolitical situations. This research adds to growing scrutiny of AI's role in national security and crisis management, where comparable studies have shown AI's unpredictable responses under stress.
Kenneth Payne's study, titled 'Shall we play a game?', was published on February 17, 2026, and is accessible on kennethpayne.uk. The research includes detailed simulations and analysis of LLM behavior in nuclear crisis games, providing a data-driven foundation for evaluating AI risks in security applications.