Watermarks Without Verification: AI Text Watermarking After the EU AI Act

Governments are increasingly mandating that LLM-generated content carry watermarks. The EU AI Act calls for markings that are "sufficiently reliable and robust." California's SB 942 requires disclosure that is "permanent or extraordinarily difficult to remove." Both mandates rest on an untested assumption: that watermark detection yields evidence reliable enough for courts. This paper tests that assumption directly. We evaluate three representative LLM watermarking methods -- KGW, Unigram, and the MarkLLM implementation of SynthID-Text -- against the Daubert admissibility criteria and the NIST SP 800-86 digital forensic process. To structure this evaluation, we propose a Forensic Readiness Score (FRS) framework with 12 criteria, three mandatory gates, and a 60-point scoring system. We focus on meaning-preserving paraphrase as the attack vector, since it is both legally realistic and difficult to dismiss as evidence tampering. The results raise serious evidentiary concerns. Out of 846 valid paraphrase runs across 15 diverse prompts per method, every single initially-detected KGW and Unigram text lost its watermark after paraphrasing -- 100% conditional removal. SynthID fared only slightly better at 98.3%. Even before any attack, false-negative rates were already high: 70% for KGW, 83% for Unigram, 80% for SynthID. The SynthID configuration also flagged 5.4% of paraphrased human-written controls as AI-generated and showed an 18.6% paradox rate, with 80% of its own pristine watermarked output landing in the uncertainty deadband. None of the three methods satisfy more than two of five Daubert factors. We also find that the FRS point-based scoring system, despite working as designed, cannot fully capture forensic uselessness -- a limitation worth noting for future framework design. These configurations, as tested, do not meet the evidentiary bar that courts require.

Generative AI systems increasingly produce content whose provenance is difficult to verify, motivating watermarking techniques for identifying model-generated outputs. Existing cryptographic watermarking methods provide strong undetectability guarantees: without a detection key, watermarked outputs are computationally indistinguishable from unwatermarked ones. However, these approaches do not address the crucial deployment challenge of how to safely delegate detection capabilities. With an unrestricted detection key, a malicious detector may use the detection key beyond its intended scope, enabling watermark sanitization, scope abuse, and user profiling. To mitigate this safety concern, we introduce, to the best of our knowledge, the first \emph{attribute-based watermarking} for generative AI models, providing fine-grained, policy-controlled watermark detection. In our approach, each generated output is associated with attributes, and each detection key is \emph{constrained by a policy} on potential attributes. A detection key can only be used to detect watermarked outputs whose attributes satisfy the corresponding policy, while watermarked outputs that fall outside the policy remain computationally indistinguishable from unwatermarked ones. We construct such an attribute-based watermarking scheme and formalize its security properties, including consistency, adaptive robustness to bounded corruptions, undetectability, and soundness, along with a security proof under standard cryptographic assumptions. Our construction integrates constrained pseudorandom functions, pseudorandom error-correcting codes, and randomness recovery procedures with generative AI models. Finally, we implement a prototype and an empirical evaluation, demonstrating that attribute-based watermarking is both effective and practical.

Watermarking embeds statistical signatures in AI-generated text for detection and attribution. We reveal a fundamental vulnerability: when users access multiple models (today's reality), watermarks trivially fail. Watermarks perturb output distributions away from the original, and in competitive markets, these perturbations are typically independent across providers. We theoretically prove that averaging output probability distributions recovers the unwatermarked distribution with up to a second-order error term. Empirically, simply averaging 3-5 models cancels out these perturbations. We introduce WASH (Watermark Attenuation via Statistical Hybridisation), which solves practical challenges in ensemble generation: vocabulary misalignment and tokenisation differences across heterogeneous models. Experiments across six watermarking schemes and three LLMs show that averaging across 3 models suppresses detection z-scores from 5-300 to below 2 (below the detection threshold of 4) and reduces TPR at 5% FPR to below 50%, while improving quality by 27.5% and running 6 times faster than the best baseline on the long sequence generation. Our results suggest that robust AI-text detection via watermarking requires either accepting this fundamental vulnerability or unprecedented coordination among model providers.