Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Assessing the Impact of Conspiracy Theories Using Large Language Models

Bohan Jiang, Dawei Li, Zhen Tan, Xinyi Zhou, Ashwin Rao, Kristina Lerman, H. Russell Bernard, Huan Liu

TL;DR

This work investigates using large language models to automate the impact assessment of conspiracy theories. It builds a CT impact dataset based on a YouGov survey and designs human-like prompting strategies that emulate fast/slow thinking, comparative/scoring judgments, and single/multi-agent reasoning. Through extensive experiments across eight LLMs, the study shows that multi-step reasoning and multi-agent debating yield greater accuracy, while prompting biases like position, wording, and verbosity can substantially affect results. It also discusses mitigation strategies and limitations, offering guidance for scalable, human-aligned CT impact evaluation in practice.

Abstract

Measuring the relative impact of CTs is important for prioritizing responses and allocating resources effectively, especially during crises. However, assessing the actual impact of CTs on the public poses unique challenges. It requires not only the collection of CT-specific knowledge but also diverse information from social, psychological, and cultural dimensions. Recent advancements in large language models (LLMs) suggest their potential utility in this context, not only due to their extensive knowledge from large training corpora but also because they can be harnessed for complex reasoning. In this work, we develop datasets of popular CTs with human-annotated impacts. Borrowing insights from human impact assessment processes, we then design tailored strategies to leverage LLMs for performing human-like CT impact assessments. Through rigorous experiments, we textit{discover that an impact assessment mode using multi-step reasoning to analyze more CT-related evidence critically produces accurate results; and most LLMs demonstrate strong bias, such as assigning higher impacts to CTs presented earlier in the prompt, while generating less accurate impact assessments for emotionally charged and verbose CTs.

Assessing the Impact of Conspiracy Theories Using Large Language Models

TL;DR

This work investigates using large language models to automate the impact assessment of conspiracy theories. It builds a CT impact dataset based on a YouGov survey and designs human-like prompting strategies that emulate fast/slow thinking, comparative/scoring judgments, and single/multi-agent reasoning. Through extensive experiments across eight LLMs, the study shows that multi-step reasoning and multi-agent debating yield greater accuracy, while prompting biases like position, wording, and verbosity can substantially affect results. It also discusses mitigation strategies and limitations, offering guidance for scalable, human-aligned CT impact evaluation in practice.

Abstract

Measuring the relative impact of CTs is important for prioritizing responses and allocating resources effectively, especially during crises. However, assessing the actual impact of CTs on the public poses unique challenges. It requires not only the collection of CT-specific knowledge but also diverse information from social, psychological, and cultural dimensions. Recent advancements in large language models (LLMs) suggest their potential utility in this context, not only due to their extensive knowledge from large training corpora but also because they can be harnessed for complex reasoning. In this work, we develop datasets of popular CTs with human-annotated impacts. Borrowing insights from human impact assessment processes, we then design tailored strategies to leverage LLMs for performing human-like CT impact assessments. Through rigorous experiments, we textit{discover that an impact assessment mode using multi-step reasoning to analyze more CT-related evidence critically produces accurate results; and most LLMs demonstrate strong bias, such as assigning higher impacts to CTs presented earlier in the prompt, while generating less accurate impact assessments for emotionally charged and verbose CTs.

Paper Structure

This paper contains 40 sections, 17 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. LLMs for Assessment
  4. LLMs for Social Problems
  5. Datasets Development
  6. Survey Data
  7. Prompting Bias Datasets
  8. Position Bias Dataset ($D_p$).
  9. Wording Bias Dataset ($D_w$).
  10. Verbosity Bias Dataset ($D_v$).
  11. Human-like Impact Assessment
  12. Fast and Slow Thinking
  13. Vanilla Ranking (Fast Thinking)
  14. Chain-of-Thought (Slow Thinking)
  15. Comparative and Scoring Assessment
  16. ...and 25 more sections

Figures (3)

  • Figure 1: Research pipeline for CT impact assessment using LLMs. The pipeline consists of two main stages: (a) Data Collection and Augmentation, where a list of popular conspiracy theories with human-annotated impact assessments is expanded using position, wording, and verbosity perturbations to test robustness; and (b) Human-like Impact Assessment, leveraging tailored prompting strategies to simulate human reasoning processes, including fast versus slow thinking, comparative versus scoring assessments, and single-agent versus multi-agent debating.
  • Figure 2: Templates for human-like CT impact assessment.
  • Figure 3: Impacts ($\tau$) of prompting biases on LLM performance across different evaluation strategies: (a) Vanilla Ranking, (b) Individual Scoring, (c) Pairwise Comparison, and (d) Chain-of-Thought. The bar plots show the relative performance changes introduced by position bias, wording bias (formal, casual, neutral), and verbosity bias (relevant and irrelevant). Results of Multi-Agent Debating are not presented here as they show negligible or no performance changes across all bias categories.