Evaluating the Effectiveness of Large Language Models for Targeted Political Messaging
(A2025-126399)

AUTHORS

Aseem Behl, University of Tübingen; Adrian Zarbock, University of Tübingen

ABSTRACT

Recent advancements in AI, particularly with models like GPT-4o, have raised concerns about potential misuse in political campaigns. To address this, our research investigates the persuasive power of targeted political messages generated by large language models (LLMs). Specifically, our study explores whether targeted political messages created by generative language models are perceived as more persuasive than non-targeted messages. Building on prior research, we conducted an online experiment to compare the persuasiveness of targeted and non-targeted political messages. Unlike earlier work, this study exclusively measures persuasion, distinguishing it from confirmation bias. It innovates by using implicit data collection methods for targeting, avoiding the reliance on self-reported or sensitive information. Additionally, the study employs a multidimensional political spectrum model, offering greater nuance compared to traditional one-dimensional approaches.