Text AI watermarks are trivially removable—the detection arms race was always unwinnable.

A technical analysis confirms that text-based AI watermarks can be stripped or spoofed with trivial effort—undermining watermarking provisions in AI regulation that assume the mechanism is reliable against motivated adversaries.

In plain terms: Statistical watermarks embedded in AI text output can be defeated by a single paraphrase pass. Conversely, watermarks can be injected into human-written text to falsely flag it as AI-generated. The mechanism is broken both ways.

The fact

Analysis published on seangoedecke.com (surfaced on Hacker News) walks through why text watermarks are fundamentally brittle: any semantic-preserving transformation - paraphrase, translation, light editing - strips the statistical signal embedded in word choice or token patterns. The attack is O(1) cost. No specialized capability required.

Our read

This matters beyond the technical detail because watermarking was positioned as a practical tool for AI content attribution. The EU AI Act includes watermarking provisions for synthetic content. Several AI labs have promoted watermarking as a transparency mechanism. If the mechanism fails against motivated adversaries - and the removal cost is trivial - the policy layer built on top of it inherits the flaw. The alternative approaches (cryptographic signing at generation time, provenance metadata embedded in the model output chain) are more robust but require infrastructure commitment from labs that haven't prioritized it.

[Under the hood] A robust watermark that survives semantic-preserving transformations would require either degrading output quality noticeably or embedding steganographic structure detectable by the same adversary trying to remove it. Both defeat the purpose.

Watch for

Whether the EU AI Office updates its technical guidance on watermarking as this evidence base accumulates - or whether the provision stays on the books despite being effectively unenforceable.