On August 2, 2026 the obligations of Article 50 of the EU AI Act took effect, requiring generative‑AI providers to mark their outputs and make them detectable as AI‑generated. Days later Anthropic announced that every Claude model released after that date embeds a SynthID‑Text watermark by default with no opt‑out, and Google has been using the same technique in Gemini since 2024. Users complained that the watermark degrades quality, especially for code, secretly encodes identifying information, and is both easily removable and inescapable. The vendors replied that quality is unchanged, no identifying data is embedded, and the watermark is robust to light editing.
We argue that neither the objections nor the assurances can currently be verified, and that this unverifiability, rather than the watermark itself, constitutes the real governance failure. We categorize the contested claims by the evidence required to settle them and evaluate the open‑source SynthID‑Text implementation on two open‑weight models, since no public tool can directly test the commercial systems. The measurements show that on prose the watermark’s effect does not exceed the variation caused by changing the sampling seed; on code one model loses about three percentage points of correctness while the other shows no measurable impact, and detection accuracy remains near chance, indicating a limitation of detectability rather than quality. The remaining gaps stem from withheld access or missing institutional involvement, which we map to a set of requirements: release of matched output pairs, disclosure of configuration, accredited audits, a shared evaluation protocol, and interoperable detection.
Review