Simulating U.S. Presidents for a Friendly Chat: Applying Generative AI to Study Political History

Abstract

Advancements in generative artificial intelligence have created new opportunities to develop large language model (LLM)-based simulation models. By designing distinct personas and training them with relevant information, modelers can simulate a wide range of agents, representing diverse personalities, socio-economic backgrounds, and demographics. The potential of these simulation models, often referred to as generative agents, extends beyond creating average representations of groups; they can also be tailored to simulate specific individuals, predicting their responses or opinions under various scenarios. In this study, we take on the challenge of simulating 60 U.S. presidents to demonstrate how this approach can contribute to the study of political history. We simulate 60 generative agents using an LLM (GPT o1) primed on the inaugural addresses of presidents from 1789 to 2025. We then ask each simulated president the question, “what factors influence the economy?” We validate the simulated responses with other LLMs tasked with predicting which president is most likely to have given each response. We then use a causal loop diagram generation tool called SD Bot to extract variables and relationships from the text responses and depict mental models. Finally, we quantify and visualize presidents’ relative similarities to each other as a network.