AI Watermark Evidence Fails Forensic Readiness: An Empirical Evaluation (arxiv.org)

Recent research has revealed significant shortcomings in the forensic readiness of watermarking methods designed for AI-generated content. With regulations like the EU AI Act and California's SB 942 advocating for reliable watermarks, the study critically evaluates three watermarking techniques—KGW, Unigram, and SynthID-Text—against established legal standards for admissibility in court. Utilizing a new Forensic Readiness Score (FRS) framework, the testing demonstrated alarming results: all KGW and Unigram watermarks were completely erased through paraphrasing, while SynthID performed marginally better with a 98.3% failure rate. High false-negative results further undermine the credibility of these methods. The findings suggest that current watermarking technologies fall short of the evidentiary standards required for legal disputes, raising urgent questions for developers and regulators. The inability of these methods to withstand simple attack vectors like meaning-preserving paraphrasing indicates a crucial gap in the application of AI watermarks, which could complicate their use in legal scenarios. This study signals a need for enhanced watermarking strategies that can reliably withstand legal scrutiny, highlighting the technical challenges that remain in maintaining the integrity of AI-generated content.