Fighting Lies with Intelligence: Using Large Language Models and Chain of Thoughts Technique to Combat Fake News

Abstract

The proliferation of fake news in the digital age presents a substantial challenge, outpacing the capabilities of conventional fact-checking methods. To address this, we introduce a pioneering strategy that utilizes fine-tuned Large Language Models (LLMs) for discerning fake news through the generation of logical reasoning that validates or critiques news headlines. This strategy seamlessly merges the predictive prowess of LLMs with the requisite for coherent explanations, facilitating not only the detection of fake news but also offering transparent, reasoned justifications for each classification. Leveraging the inherent “Chain of Thought” (CoT) reasoning and model distillation processes of pre-trained LLMs, our approach enhances detection accuracy while rendering the models’ complex decisions accessible to human understanding. This research signifies a groundbreaking contribution, extending beyond mere methodological progress by presenting an open-source dataset fortified with CoT annotations, establishing a new benchmark for fake news detection. This dataset, consisting of a diverse mixture of human-annotated news and those generated under human-guided contexts using the OpenAI GPT 3.5 model, promises to be a valuable resource for future scholarly endeavours in the field. By optimizing two distinct LLMs (FLAN-T5 and Llama-2), our methodology demonstrates unprecedented efficacy, surpassing the existing state-of-the-art results by 11.9% and elevating the overall performance of LLMs in fake news detection.