Anthropic adds "watermarks" to Claude's text: a response to EU AI rules

Anthropic has started embedding invisible watermarks — digital markers hidden within the text itself — into content generated by its Claude models, and adding signed provenance metadata to generated files. The measure is being rolled out not just in the European Union but worldwide. The company stresses that detecting the watermark indicates Claude may have processed the content, not necessarily that Claude wrote it.

The move follows Anthropic's decision to sign the EU's Code of Practice on transparency for AI-generated content, under Article 50(2) of the EU AI Act. According to the European Commission, roughly 190 organizations had joined the code by the end of July, including Google, Meta, Microsoft, Mistral, and OpenAI, all registered as "providers" — the party responsible for machine-readable marking and detection. Notably, OpenAI scrapped its own watermarking plan back in September 2024 after a survey found that nearly 30% of ChatGPT users would use the service less if watermarking were added.

Technically, the embedded text marker doesn't affect the meaning, quality, or readability of a response, and it survives copy-pasting since it's applied at the model level. Generated .svg, .png, and .jpg files carry signed metadata conforming to the open C2PA standard, showing how a file was created and whether it was subsequently altered. Anthropic also notes that even simple editing, translation, shortening, or reformatting of text can leave the mark behind — even when the original idea and wording come from another source.

Article 50 of the EU law carries four exemptions, two of which apply specifically to edited text: a system that merely assists with standard editing and doesn't substantially alter the meaning of the input is exempt from the labeling requirement, and text that has undergone human editorial review and revision can likewise be exempt from disclosure. In practice, that means text bearing a Claude watermark can be published without any labeling obligation on the publisher's part.

Alex Cui, CTO of the AI-detection company GPTZero, notes that text watermarks can be "broken" through heavy paraphrasing, and that free tools have already managed to defeat even Google DeepMind's SynthID system. Jonas Geiping, who studies watermarking at the ELLIS Institute, counters that while paraphrasing can strip a watermark, not every paraphrase will do so — fully removing a mark from a long document requires substantially rewriting most of the original text.

The practical stakes are significant: an author who writes their own text and simply runs it through Claude for a "polish," and a translator rendering someone else's article into another language, both end up with watermarked text. Anthropic flags both scenarios separately among the system's limitations, so the presence of a watermark only indicates that Claude was involved somewhere in the process — it says nothing about who actually wrote the piece. The company has pledged to extend watermarking to models released before August, though no firm timeline has been announced.