Watermarking: from Impossibility to Auditable Compliance
Researchers propose auditable text-watermark checks that operationalize EU AI Act Article 50 without claiming universal robustness.
The paper addresses EU AI Act Article 50(2), which requires providers of generative systems to make synthetic outputs machine-readable and detectable where technically feasible. It argues that strong watermarking of free-form text cannot resist adaptive removal and that ordinary edits weaken statistical evidence. The authors define a description-length robustness profile, prove a finite-sample bound using collision entropy, and construct label-conditional conformal sets that report whether a watermark is supported, not supported, or inconclusive. A small tournament-watermark simulation suggests the surviving-token rule overstates the tolerable edit rate about twofold, and they outline certificates and recalibration for the Commission's 2026 Code of Practice.
- EU AI Act Article 50 requires detectable synthetic outputs where technically feasible.
- Strong text watermarks fail against adaptive removal, and edits weaken evidence.
- A description-length profile and collision entropy give a finite-sample decay bound.
- Conformal prediction reports watermark supported, not supported, or inconclusive.
- A simulation shows the surviving-token rule roughly doubles tolerable edit rate.
- Organizations
- European Commission
- Countries
- European Union
From the original article
1 min read at arxiv.orgArticle 50 (2) of the EU Artificial Intelligence Act requires providers of generative systems to make synthetic outputs machine-readable and detectable, while qualifying the effectiveness, interoperability, robustness, and reliability by technical feasibility, cost, content-specific limits, and the state of the art.
By Fernando Delbianco · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.