An Empirical Analysis of LLMs for Countering Misinformation

2025-02-28

AI Summary

This study systematically evaluates the fact-checking capabilities of ChatGPT, Gemini, and Claude in countering political misinformation. Addressing limitations of existing prompt-engineering approaches, we propose a two-step chain-of-thought prompting method: first identifying credible news sources, then generating rebuttal responses. Our experiments—first to reveal such patterns across these models—identify three critical deficiencies: (1) source grounding failure (inability to accurately cite verifiable media outlets), (2) pronounced left-leaning ideological bias (37–52% lower citation rates for right-leaning sources), and (3) severely limited response diversity (only 41% cross-model consistency on identical queries). These findings demonstrate that prompt engineering alone is insufficient for robust, reliable fact-checking. We conclude that integrating external knowledge retrieval and systematic bias mitigation mechanisms is essential to improve factual accuracy and ideological neutrality in AI-mediated misinformation interventions.

Technology Category

• Natural Language Processing: Fact-Checking / Misinformation Detection (NLP Focus)
• Application Domains: Misinformation & Fake News
• Reasoning under Uncertainty: Other Foundations of Reasoning under Uncertainty

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Responsible Web: Mitigating misinformation and disinformation

Abstract

While Large Language Models (LLMs) can amplify online misinformation, they also show promise in tackling misinformation. In this paper, we empirically study the capabilities of three LLMs -- ChatGPT, Gemini, and Claude -- in countering political misinformation. We implement a two-step, chain-of-thought prompting approach, where models first identify credible sources for a given claim and then generate persuasive responses. Our findings suggest that models struggle to ground their responses in real news sources, and tend to prefer citing left-leaning sources. We also observe varying degrees of response diversity among models. Our findings highlight concerns about using LLMs for fact-checking through only prompt-engineering, emphasizing the need for more robust guardrails. Our results have implications for both researchers and non-technical users.

Problem

• Assessing LLMs' ability to counter political misinformation.
• Evaluating LLMs' reliance on credible sources for fact-checking.
• Identifying biases and response diversity in LLM-generated content.

Innovation

• Two-step chain-of-thought prompting approach
• Identify credible sources for claims
• Generate persuasive responses to misinformation