On September 26, 2025, Professor Pang Xuan from the School of International Studies at Peking University and Director of the PKU Laboratory for Global Risk Politics delivered an academic lecture titled “Hawkish Bias in Geopolitical Decision-Making: Comparative Experiments with Human and LLM Agents” at the Faculty of Social Sciences, University of Macau, upon invitation. The event was chaired by Assistant Professor Yin Weiwen from the Faculty of Social Sciences at the University of Macau. Attendees included Zhu Yibo, Associate Dean of the School of International Studies at Peking University; Pan Wei, Chair Professor at the Faculty of Social Sciences, University of Macau; and Wang Yizhou, Professor at the School of Computer Science at Peking University.
As artificial intelligence (AI) technology increasingly permeates critical decision-making domains such as international relations and foreign policy, systematically evaluating the impacts and risks of AI applications in decision-making has become a major concern for both academia and practitioners. At the beginning of the lecture, Professor Pang Xuan pointed out that AI technology has transitioned from a “future issue” to a “current reality,” with government departments in multiple countries, including the U.S. Department of Defense, already deploying AI systems in practice. Therefore, systematically assessing the potential risks and biases in AI-assisted or AI-driven foreign policy decision-making has become an urgent academic and practical task.
Subsequently, Professor Pang Xuan systematically introduced a recent study led by her interdisciplinary team. The study used a randomized controlled experiment designed to identify the “hawkish bias” in foreign policy decision-making as a benchmark. It constructed 3,973 large language model (LLM)-based agents using demographic variables of human subjects from the experiment, replicating individual and group decision-making scenarios from the original experiment through single-agent decisions and multi-agent systems. The study observed whether the agents replicated, amplified, mitigated, or complicated three types of hawkish biases observed in human decision-makers: “loss framing effects,” “intent bias,” and “reactive devaluation.” The findings revealed that, like human decision-makers, LLM agents exhibit heuristic biases in information-scarce and unstructured decision-making contexts. However, the patterns of these biases differ significantly from human cognitive biases in both mechanisms and manifestations, making them difficult for humans to understand and predict. Professor Pang Xuan further analyzed that while human biases often stem from psychological shortcuts and emotional impulses, LLM agents' biases are more likely to arise from embedded stereotypes and oversensitivity to linguistic cues in experimental designs. The study suggests that the current application of AI in political decision-making carries significant risks, potentially complicating geopolitics and increasing uncertainty.
During the discussion session, Wang Yizhou and Yin Weiwen provided comments on the study, both highly praising its innovation, rigor, and practical significance. They noted that the research offers insightful empirical evidence for understanding AI behavior patterns in high-risk decision-making and highlights current challenges in AI development. In her response, Professor Pang Xuan expanded the discussion to include how to interpret AI behavior patterns in political decision-making, further assess AI decision-making risks, and construct responsible AI governance frameworks. Faculty and students actively engaged in the Q&A session, exchanging in-depth views with Professor Pang Xuan on experimental design, model variability, and policy implications.
The lecture concluded in a vibrant atmosphere of academic exchange. The research by Professor Pang Xuan's team profoundly demonstrates that, given the limited information and open-ended nature of geopolitical decision-making, the introduction of pattern recognition-based AI systems for decision support requires extreme caution.
Written by: Wang Huiyi
Photo by:Participating Faculty and Students
Edited by: Zhu Yibo