AI Text Watermarks After the EU AI Act: The Verification Gap

Summary

This paper examines the governance of AI-generated text watermarking after Article 50 of the EU AI Act took effect on August 2, 2026. The provision requires generative AI providers to mark produced content and make it detectable as AI-generated. Anthropic later said that every Claude model released after that date embeds a SynthID-Text-based watermark by default, without a user opt-out, while Google has used SynthID-Text in Gemini since 2024. Users have disputed whether watermarking harms output quality, secretly encodes identifying information, and can be removed; the vendors have offered assurances about unchanged quality, lack of identifying data, and resistance to light editing. The authors argue that neither the objections nor the assurances can currently be independently verified, making unverifiability the central governance failure. Because no public tool can test the deployed systems, they evaluate the open-source SynthID-Text implementation on two open-weight models. For prose, the measured quality effect is no greater than changing the sampling seed. For code, watermarking reduces correctness by three points on one model and by less than the measurement threshold on the other, while detection remains near chance. The authors characterize this as a detectability limitation rather than evidence of a major quality cost. They attribute the remaining verification gaps to withheld access and missing institutions, and propose matched output releases, configuration disclosure, accredited audits, a shared evaluation protocol, and interoperable detection tools.