At the Center for Strategic and International Studies (CSIS), a Washington, D. C. -based think tank, the Futures Lab is developing projects to use artificial intelligence (AI) to revolutionize diplomacy. Funded by the Pentagon's Chief Digital and Artificial Intelligence Office, the lab experiments with AIs like ChatGPT and DeepSeek to address issues of war and peace. While AI tools have recently assisted foreign ministries worldwide with routine tasks such as speechwriting, they are now being explored for high-stakes decision-making roles. Researchers are testing AI’s ability to draft peace agreements, prevent nuclear war, and monitor ceasefire compliance. The U. S. Defense and State Departments, along with other countries like the U. K. and even Iran, are also venturing into AI to reshape diplomatic practices, including negotiation planning. Futures Lab Director Benjamin Jensen notes that, although the idea of AI aiding foreign policy has existed for some time, its practical implementation remains nascent. In one study, researchers tested eight AI models with thousands of questions on deterrence and crisis escalation scenarios. Results showed that models such as OpenAI’s GPT-4o and Anthropic’s Claude leaned "distinctly pacifist, " choosing force in under 17% of cases. Conversely, Meta’s Llama, Alibaba Cloud’s Qwen2, and Google’s Gemini showed more aggressive tendencies, favoring escalation up to 45% of the time. Moreover, AI responses varied by country perspective, recommending more aggressive policies for U. S. , U. K. , or France diplomats, and advocating de-escalation for Russia or China, highlighting the need to tailor models to institutional doctrines. Retired U. S. Army Special Forces officer and AI strategist Russ Berkoff attributes this variability to human biases embedded by the developers, emphasizing that differences arise from those who build the algorithms, not the AI itself. This unpredictability poses a “black box” challenge, Jensen explains, as AI systems do not hold values or judgments but generate outputs through complex mathematical processes.

CSIS has also launched an interactive program called "Strategic Headwinds" to help negotiate peace in Ukraine by training an AI on hundreds of peace treaties and news articles to identify agreement areas that could lead to a ceasefire. Mark Freeman, Executive Director at Spain’s Institute for Integrated Transitions (IFIT), supports AI’s potential in conflict resolution, favoring faster “framework agreements” and limited ceasefires over protracted peace talks, which historically have been less effective. He believes AI could accelerate these fast-track negotiations. Similarly, Andrew Moore, an adjunct senior fellow at the Center for a New American Security, envisions a future where AI may initiate negotiations, simulate leaders like Putin or Xi Jinping for crisis response testing, assist ceasefire monitoring, analyze satellite imagery, and enforce sanctions, automating tasks that once required large teams. However, Jensen acknowledges challenges, including amusingly unexpected AI outputs. For example, when asked about "deterrence in the Arctic, " the AI misinterpreted "deterrence" as law enforcement and "Arctic" as snowy terrain, leading to bizarre scenarios about arresting Indigenous peoples for snowball throwing. Such errors demonstrate the necessity of training AI with specialized diplomatic and policy data rather than general internet content, which is often dominated by irrelevant material. Stefan Heumann, co-director of Berlin’s Stiftung Neue Verantwortung, cautions that AI cannot replace essential human elements like personal relationships between leaders, which significantly influence negotiations. He also notes AI’s struggle to assess long-term consequences of short-term actions, citing the appeasement at Munich in 1938 as an example where simplicity in terms like "de-escalate" fails to capture complex realities. Furthermore, Heumann points out that AI thrives in open environments but is limited when analyzing closed societies like North Korea or Russia. Andrew Reddie, founder of the Berkeley Risk and Security Lab, echoes these concerns, observing that adversaries gain an edge as open democracies like the U. S. publish vast information readily accessible for training enemy AI, whereas authoritarian states do not. He also warns that AI tools are less helpful for “black swan” geopolitical challenges that fall outside known patterns. Despite these criticisms, Jensen believes many concerns are surmountable but stresses practical issues. He envisages two futures for AI in American diplomacy: one where AI, trained on diplomatic tasks and documents like cables, produces actionable insights to resolve pressing issues effectively, and another less optimistic path not detailed here. The potential for AI to transform diplomacy is significant, but careful development and context-aware application remain essential.