LLM-based detection and analysis¶
Integration of large language models (LLMs) into misinformation detection pipelines, leveraging their language understanding and generation capabilities for both detection and explanation of false information.
Approaches¶
Direct classification:
Using LLMs as classifiers to assess veracity of news articles or claims through few-shot or zero-shot prompting.
Explanation generation:
Generating explanations for why content is misinformation through LLM-generated reasoning or evidence synthesis.
Synthetic data and proxy tasks:
Using LLMs to generate synthetic user reactions, comments, or explanations that enrich detection models through multi-task learning.
Expert ensemble:
Employing multiple task-specific LLM experts and merging their predictions with confidence scoring for robust detection.
Retrieval-augmented approaches:
Combining LLMs with external knowledge bases or Wikipedia to ground verification in factual sources.
Key papers in this wiki¶
- DELL: Generating Reactions and Explanations for LLM-Based Misinformation Detection (2024) — Proposes DELL, which generates synthetic user reactions from diverse demographic perspectives, creates six explainable proxy tasks with LLM explanations, and uses expert ensemble to merge task-specific predictions; achieves up to 16.8% improvement in macro F1-score on seven datasets across fake news, framing, and propaganda detection.
- Combating Misinformation in the Age of LLMs: Opportunities and Challenges (2023) — Comprehensive survey of opportunities and challenges for using large language models in misinformation detection, intervention, and attribution.
Related concepts¶
- Misinformation and fake news detection — broader detection methods
- AI and machine learning — computational foundations
- Propaganda detection — detection of coordinated influence campaigns