Go to ACL ARR 2025 May homepageA Paraphrase Attack Safe and Distortion Free Watermarking Technique for AI-Generated Text
Abstract: A report from the European Union Law Enforcement Agency forecasts that by 2026, up to 90% of online content may be synthetically generated. This surge raises significant concerns among policymakers, who warn that “Generative AI could act as a force multiplier for political disinformation. The combined effect of generative text, images, videos, and audio may surpass the influence of any single modality.” In response, California’s Bill AB 3211 mandates the watermarking of all AI-generated content. However, existing watermarking techniques remain vulnerable to tampering and can potentially be circumvented by malicious actors. With the widespread adoption of Large Language Models (LLMs) across various applications, there is an urgent need for robust text watermarking solutions. Early watermarking models for LLMs proposed by Kirchenbauer et al. faced criticism after studies demonstrated that paraphrasing could effectively remove these watermarks. In this paper, we introduce PECCAVI, the first text watermarking technique that is both resistant to paraphrase attacks and distortion-free, surpassing all existing methods in performance.
Paper Type: Long
Research Area: Ethics, Bias, and Fairness
Research Area Keywords: AI-generated content,Text watermarking,Large Language Models ,Paraphrase attack resistance, disinformation
Contribution Types: NLP engineering experiment
Languages Studied: English
Submission Number: 7624
Loading