Artificially Intelligent Political Agents

The rise of Large Language Models (LLMs) like ChatGPT has significantly expanded the capabilities of natural language processing, enabling diverse applications from text summarization to programming assistance. This thesis explores the possibility of configuring LLMs to impersonate historical political figures, thus creating "Artificially Intelligent Political Agents."By assigning distinct personalities to LLMs and letting them go head-to-head in a political debate, this research investigates their responses to a wide range of political issues, which were not directly addressed during the historical figures’ lifetimes. The research questions focus on the effectiveness of LLMs in replicating the linguistic styles and character traits of historical figures, the plausibility of their stances on modern issues, and the logical coherence of their arguments.Historical figures selected for this study are Winston Churchill, Karl Marx, and Niccolò Machiavelli, representing different political ideologies. This thesis applies techniques inprompt engineering and Retrieval Augmented Generation (RAG) to achieve impersonationand proposes a multi-agent discussion framework, simulating a political debate.The evaluation methodology combines qualitative feedback from five human participants,evaluating the simulated discussion, with quantitative analysis using psychometric inventories,obtained through the Big Five Personality Test. This mixed-methods approach provides a comprehensive assessment of the LLMs’ impersonation capabilities as well as the performance of the multi-agent disucssion framework, offering insights into their potential applications in educational, historical, and entertainment contexts.The findings reveal that LLMs can convincingly impersonate historical personas and succesfully engage in a political discourse. Limitations include a lack of emotion in the impersonation and the overall sound of the simulated debate, feeling more like a written exchange than a live debate. Overall, the feedback from the participants about this thesis was very positive.