The Role of Sentiment Shift: Measuring and Explaining Performance of Fake News Detection After LLM Laundering

Abstract

With their advanced capabilities, Large Language Models (LLMs) can generate highly convincing and contextually relevant fake news, which can contribute to disseminating misinformation. Though there is much research on fake news detection for human-written text, the field of detecting LLM-generated fake news is still under-explored. This paper augments existing datasets to measure the efficacy of detectors in identifying LLM paraphrased fake news. By investigating which models excel at which tasks (detection, paraphrasing to evade detection, and paraphrasing for semantic similarity), we found detectors struggled to detect LLM-paraphrased fake news more than human-written text. Further, upon inspecting LIME explanations, we observed a possible sentiment shift and digging deeper revealed a worrisome trend for paraphrase quality measurement: many samples exhibit sentiment shift despite a high BERTSCORE.