An empirical evaluation of KGW, Unigram, and SynthID-Text reveals that current LLM watermarking methods struggle to meet Daubert admissibility criteria under meaning-preserving paraphrase attacks. The study introduces a Forensic Readiness Score (FRS) to evaluate watermark reliability for legal and regulatory compliance.
HOW THIS AFFECTS YOU
●
policyYou should account for the high failure rate of current watermarking techniques when drafting AI governance mandates.