Watermarking has become a key tool for identifying text produced by large language models, with current schemes embedding statistical signals into the generation process to enable later detection. A new paper on arXiv presents what it calls a unified attack framework against these statistical generative watermarks.
The framework is designed to target the common structure of such watermarking methods, rather than any single implementation. By treating the watermark as a statistical signal, the proposed approach aims to weaken or remove it while preserving the quality and coherence of the generated text.
The paper does not claim to break all watermarks, but it underscores a growing arms race between detection techniques and adversarial methods. As watermarking becomes more widely adopted, understanding its limitations will be critical for building more robust attribution systems.