Scholarlink

<p>Significant academic and public attention has been paid to how generative AI tools can be used in political persuasion and microtargeting. A growing body of recent research finds that in many cases, effort to customize political messaging using an LLM yields little persuasive benefit. In this project, we explore the way LLMs construct persuasive messages, how such statements vary when the LLM is asked to microtarget individuals, and the degree to which the changes induced by microtargeting lead to increased persuasion of human readers. We find variation in the degree to which different LLMs successfully comply with the microtargeting task. Furthermore, even for LLMs that produce more distinct messaging, the strategies are rarely systematically aligned with users’ background features and do not increase the persuasiveness of the message in the aggregate. We discuss the implications of these findings for research on persuasion and generative AI.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Looking Under the Hood: How LLMs Attempt Political Persuasion and Microtargeting

  • 期刊论文

  • 发表日期：

    2026年3月17日

查看全文
  • Alex Lyman,
  • Ethan C. Busby,
  • Lisa P. Argyle,
  • Joshua R. Gubler,
  • Bryce Hepner,
  • David Wingate

摘要

Significant academic and public attention has been paid to how generative AI tools can be used in political persuasion and microtargeting. A growing body of recent research finds that in many cases, effort to customize political messaging using an LLM yields little persuasive benefit. In this project, we explore the way LLMs construct persuasive messages, how such statements vary when the LLM is asked to microtarget individuals, and the degree to which the changes induced by microtargeting lead to increased persuasion of human readers. We find variation in the degree to which different LLMs successfully comply with the microtargeting task. Furthermore, even for LLMs that produce more distinct messaging, the strategies are rarely systematically aligned with users’ background features and do not increase the persuasiveness of the message in the aggregate. We discuss the implications of these findings for research on persuasion and generative AI.