AI Reports · Stanford HAI · 2026-10-30 · Report / white paper
Escalation Risks from LLMs in Military and Diplomatic Contexts | Stanford HAI
This brief presents results from a wargame simulation that evaluates escalation risks of LLMs in high-stakes military and diplomatic decision-making. All five off-the-shelf models showed some form of escalation and hard-to-predict patterns, in some cases leading to the use of nuclear weapons. The most escalatory and unpredictable model was the only one that had not undergone reinforcement learning from human feedback.