Exploring the Deceptive Power of LLM-Generated Fake News: A Study of Real-World Detection Challenges

TL;DR

This work targets real-world detection challenges posed by LLM-generated fake news in healthcare by introducing VLPrompt, a CVAE-like prompting framework that encodes real articles into a latent space, subtly modifies key elements under theme and style controls, and decodes convincing fakes without extra data ($P(X'|Z, C_t, C_s)$ with prior $P(Z, C_t, C_s)$). It releases the VLPFN dataset combining real, human-fact-checked, and LLM-generated fake texts and conducts extensive human and automated evaluations across multiple prompt strategies and detectors. Findings show VLPrompt can substantially deceive both humans and automated detectors, with pattern analyses revealing robust cues like negation changes and raised concerns that aid future detection efforts. The work highlights persistent threats in automated misinformation and provides a dataset and methodology to benchmark and improve detection systems.

Abstract

Recent advancements in Large Language Models (LLMs) have enabled the creation of fake news, particularly in complex fields like healthcare. Studies highlight the gap in the deceptive power of LLM-generated fake news with and without human assistance, yet the potential of prompting techniques has not been fully explored. Thus, this work aims to determine whether prompting strategies can effectively narrow this gap. Current LLM-based fake news attacks require human intervention for information gathering and often miss details and fail to maintain context consistency. Therefore, to better understand threat tactics, we propose a strong fake news attack method called conditional Variational-autoencoder-Like Prompt (VLPrompt). Unlike current methods, VLPrompt eliminates the need for additional data collection while maintaining contextual coherence and preserving the intricacies of the original text. To propel future research on detecting VLPrompt attacks, we created a new dataset named VLPrompt fake news (VLPFN) containing real and fake texts. Our experiments, including various detection methods and novel human study metrics, were conducted to assess their performance on our dataset, yielding numerous findings.

Figures (10)

• Figure 1: Real-world scenario how fake news creators would use LLMs. Rather than manually crafting false information, fake news creators opt to employ a prompt that automatically transforms facts into fake news.
• Figure 2: Workflow of VLPrompt against two baseline prompt strategies. The steps in the figure align with the steps in the prompts (Appendix \ref{['app:prompt']}).
• Figure 3: An example of the comparison between a real news article and the corresponding VLPrompt-generated article. Phases highlighted in red are modified statements and those highlighted in green are unmodified factors.
• Figure 4: Word cloud of the differences between LLM-generated articles and corresponding real news. The top three common modification strategies are: "doesn't mention," "raise concern," and "different perspective."
• Figure 5: The prompt for fake news generation (SUMMARY). The reference article is concatenated to the end of the prompt text.