Original title: Watermarks Without Verification: AI Text Watermarking After the EU AI Act
Alexander Nemecek, Vipin Chaudhary, Erman Ayday · September 2026
New laws require companies to mark AI text so people know a computer wrote it. Companies like Anthropic and Google use a hidden text pattern called SynthID-Texta specific watermarking method that embeds a hidden statistical pattern in generated text so software can later recognise it as AI-written.. Users worried this secret mark damages text quality, ruins coding help, tracks people, or fails to work. Meanwhile, the companies claimed it causes no harm and stays intact through light editing. The researchers pointed out a major problem: right now, the public cannot verify either side's claims. Because the real systems are closed, the researchers tested an open version on two open models. They found that on prose, the watermark changes text quality no more than picking a different random seed. On code, correctness fell by three points on one model and stayed below measurement on the other. But watermarked code was barely detectable, scoring near chance. The real issue is missing access, proof, and oversight.
Imagine a city requires all food trucks to stamp their sandwiches so inspectors can tell where each one came from. The trucks stamp them, but inspectors are only allowed to test samples the trucks hand over themselves, never the sandwiches sold to customers. Customers say the stamps ruin the taste; owners insist they don't. In this paper, the EU requires text produced by AI systems to be marked so it can be detected as machine-written, but the deployed systems are closed, so users' complaints and providers' assurances about the marking remain equally untestable.
The EU now requires AI-generated text to be detectable, and Anthropic has turned on a text watermark with no opt-out while Google has used the same method since 2024. If neither complaints nor assurances can be tested, users cannot know whether text quality changes, whether identifying information leaks, or whether the mark is removable — so the law's requirement is unenforceable in practice.
The study cannot test the actual deployed Claude or Gemini systems; it tests an open-source implementation on two open-weight models. It also does not settle whether deployed watermarks degrade quality or encode identifying information, and on code detection remains near chance, which the authors call a detectability limitation rather than a quality one.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Watermarks Without Verification: AI Text Watermarking After the EU AI Act — Alexander Nemecek, Vipin Chaudhary, Erman Ayday
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.