Adversarially Enhanced Financial Misinformation: A Comparative Analysis of LLM- vs. GAN-Generated Content Exposing AI Moderation Vulnerabilities
Christopher Santorelli, Victor Ginart Belmonte, Ryan Mastropaolo · 2025
As Large Language Models (LLMs) become more pervasive, their capability to generate convincing financial news poses an escalating threat to investor decision-making and market stability. However, contemporary content moderation and AIbased verification systems exhibit notable vulnerabilities when confronted with the subtle linguistic manipulations introduced by advanced prompt engineering techniques and adversarial training. This study investigated the comparative credibility, influence, and detectability of AI-generated financial headlines produced via Zero-Shot, Few-Shot (8-Shot), and Chain-of-Thought (CoT) prompting, with CoT outputs further used to train a GAN for adversarially enhanced text generation. We compiled a combined dataset of NASDAQ-listed securities and web-scraped, human authored news, generated additional AI-driven headlines under three prompting paradigms, and conducted a survey of randomly sampled headlines ($\mathbf{n} \boldsymbol{=} \mathbf{3 0 0}$) to assess the credibility, market perception impact, investment influence, and AI detectability. The analysis revealed that headlines generated through Chain-of-Thought prompting consistently scored higher in perceived authenticity, influenced investment sentiment more profoundly, and were harder for participants to classify as AI-written. The findings underscore the urgent need for adversarially robust content moderation and verification mechanisms, capable of adapting to the rapidly evolving landscape of AI-generated financial misinformation, particularly when Chain-of-Thought reasoning is leveraged to enhance GAN-generated content.