Anthropic has announced a watermark detection API that allows third-party developers to verify the origin of text generated by Claude. The system utilizes a variant of the SynthID-Text method, originally published by Google DeepMind in Nature in 2024. By tweaking the randomness source during word selection, it creates a traceable pattern that Anthropic claims does not affect content quality or readability. For file-based outputs, the company employs the open C2PA standard to attach metadata without altering the document itself.
Technical limitations and vulnerabilities
The watermark is not infallible. Its reliability decreases significantly on short texts, fact-heavy passages, or code where alternative phrasings are limited. Furthermore, heavy human rewriting can strip the signal, whereas translations retain it as Claude selects all words in that process. The system cannot distinguish between fully AI-written text and heavily edited content, nor can it identify if a different model was used. The lack of a public detector has already sparked a race to unverified removal tools, including an open-source project with over 4,500 GitHub stars.
EU AI Act compliance and global rollout
Anthropic's move is a direct response to the obligations of the EU AI Act. The company signed the EU Code of Practice on transparency for AI-generated content, joining roughly 190 other signatories. Since there is currently no technical way to limit the feature by region, watermarking launches worldwide. All models released after August 2, 2025, support the feature out of the box, while older models will receive the update in coming months.
Impact on the European market
The introduction of this API aligns with a regulatory framework that, as highlighted by new EU sanctions up to 15 million euros for AI transparency failures, makes content origin verification a legal priority. For European companies operating in regulated sectors, access to this third-party tool is a crucial step toward demonstrating compliance with disclosure obligations.

AI-generated comment
AI-generated comment
AI-generated comment
AI-generated comment