Scholarlink

This chapter explores the intersection of Artificial Intelligence (AI) and elections, focusing on the critical challenge of ensuring safety in the age of Large Language Models (LLMs) including AI-generated misinformation and its impact on electoral integrity. This chapter is divided into two complementary parts. The first part describes Do-Not-Answer, a framework featuring a three-level hierarchical taxonomy of LLM risks including hallucination, bias, toxic language, and misinformation. Do-Not-Answer has been used to create datasets in multiple languages that serve to evaluate LLMs with respect to mitigation strategies, content filtering, and model alignment. The second part discusses the ElectAI taxonomy and dataset. ElectAI has been created to aid claim understanding with respect to election processes, equipment, and claims of fraud in both AI- and human-generated social media posts. The two parts combined present the reader with a comprehensive overview of both general and election-related AI safety issues along with strategies to address them.

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

Creating Frameworks and Datasets at the Intersection of AI Safety and Elections

  • 章节

  • 发表日期：

    2025年

查看全文
  • Alphaeus Dmonte,
  • Yuxia Wang,
  • Preslav Nakov,
  • Haonan Li,
  • Xudong Han,
  • Marcos Zampieri,
  • Kevin Lybarger,
  • Massimiliano Albanese,
  • Timothy Baldwin

摘要

This chapter explores the intersection of Artificial Intelligence (AI) and elections, focusing on the critical challenge of ensuring safety in the age of Large Language Models (LLMs) including AI-generated misinformation and its impact on electoral integrity. This chapter is divided into two complementary parts. The first part describes Do-Not-Answer, a framework featuring a three-level hierarchical taxonomy of LLM risks including hallucination, bias, toxic language, and misinformation. Do-Not-Answer has been used to create datasets in multiple languages that serve to evaluate LLMs with respect to mitigation strategies, content filtering, and model alignment. The second part discusses the ElectAI taxonomy and dataset. ElectAI has been created to aid claim understanding with respect to election processes, equipment, and claims of fraud in both AI- and human-generated social media posts. The two parts combined present the reader with a comprehensive overview of both general and election-related AI safety issues along with strategies to address them.