Performance of Large Language Models in the Cognitive Analysis of Misinformation: Evaluation Study

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Computational Social Science

Background:

• Misinformation on social media significantly impacts public discourse.
• Human moderators face scalability challenges in content moderation.
• Large language models (LLMs) show potential but require evaluation for misinformation analysis.

Purpose of the Study:

• Evaluate LLM effectiveness in detecting and interpreting misinformation against human annotators.
• Analyze the impact of prompt engineering strategies on LLM performance.
• Discuss ethical considerations for LLM deployment in content moderation.

Main Methods:

• Compared 4 OpenAI models with human annotators on the MuMiN dataset.
• Utilized a cognitive framework and structured questions for misinformation assessment.
• Employed 0-shot, few-shot, and chain-of-thought prompting, evaluating performance via precision, recall, F1-score, and accuracy.

Main Results:

• GPT-4 Turbo with chain-of-thought prompting achieved 67.2% accuracy and 78.3% F1-score, but human annotators (70.1% accuracy, 81% F1-score) performed better.
• LLMs excelled at logical reasoning but struggled with sarcasm, nuanced understanding, and user intent.
• LLM confidence scores correlated with accuracy on simple tasks but were less reliable for complex judgments.

Conclusions:

• LLMs demonstrate potential for automating misinformation detection, but human oversight remains crucial.
• A hybrid framework integrating LLMs for initial screening and humans for complex evaluation is a promising direction.
• Future research should focus on fine-tuning LLMs with cognitive/emotional features and advanced prompting techniques.