LLM-Generated Fake News Induces Truth Decay in News Ecosystem: A Case Study on Neural News Recommendation

TL;DR

The paper addresses how LLM-generated fake news can destabilize neural news recommendation systems. It builds a simulation pipeline and a ~56k-item dataset spanning generation modes L0–L3 to study the impact, evaluating with LSTUR and NRMS using metrics such as MRR, nDCG, and Ratio@K, plus a Relative Real Advantage measure. The results reveal a Truth Decay phenomenon where real-news ranking advantages erode as LLM-generated content permeates candidates, histories, and training data, with perplexity-based familiarity offering a plausible mechanism. The work highlights significant threats to news ecosystem integrity in the LLM era and proposes multi-faceted countermeasures and directions for future work, including higher-generation modes and real-world monitoring.

Abstract

Online fake news moderation now faces a new challenge brought by the malicious use of large language models (LLMs) in fake news production. Though existing works have shown LLM-generated fake news is hard to detect from an individual aspect, it remains underexplored how its large-scale release will impact the news ecosystem. In this study, we develop a simulation pipeline and a dataset with ~56k generated news of diverse types to investigate the effects of LLM-generated fake news within neural news recommendation systems. Our findings expose a truth decay phenomenon, where real news is gradually losing its advantageous position in news ranking against fake news as LLM-generated news is involved in news recommendation. We further provide an explanation about why truth decay occurs from a familiarity perspective and show the positive correlation between perplexity and news ranking. Finally, we discuss the threats of LLM-generated fake news and provide possible countermeasures. We urge stakeholders to address this emerging challenge to preserve the integrity of news ecosystems.

Figures (8)

• Figure 1: Illustration of Truth Decay phenomenon, where real news gradually loses its advantageous position in rankings against fake news as LLM-generated news penetrates.
• Figure 2: Illustration of a typical news recommendation model in the (Left) training and (Right) inference phases.
• Figure 3: Illustration of dataset construction process, which consists of (Top) repurposing a fake news detection dataset for news recommendation and (Bottom) prompting LLMs to generate news data and performing quality check.
• Figure 4: Cosine similarity distributions between embeddings of LLM-generated and human-written news.
• Figure 5: Wordclouds of human and generated news.