Anthropic details how it will hide watermarks in Claude's writing to meet EU AI Act

Anthropic has published a blog post explaining how it will watermark the text generated by its chatbot Claude to comply with the EU AI Act's Transparency Code, which requires AI companies to use systems that make AI-generated content identifiable. The company says it will use SynthID Text, an approach released by Google DeepMind in 2024, and plans to release a watermark detection API.

The watermark works by biasing Claude's choices in low-stakes decisions — such as picking between words like 'overcast' and 'grey' — to create a pattern that is undetectable to a reader but visible to anyone holding the key that encodes it. Anthropic says the watermarking does not change output quality, and that a watermarked response is indistinguishable from an unwatermarked one to a reader.

On whether a user could hide the watermark by editing, Anthropic said light editing will probably not remove it completely, while a complete rewrite where every word is replaced will. In the second case, the company added, it is arguable whether the text can any longer be described as AI-generated at all.

For text that Claude only proofread or lightly edited, the watermark's detectability depends on the length of the text and how heavily Claude reworked it — in a lightly edited passage, most words are the author's and there is little for the watermark to attach to. Code carries much less of a watermark than prose, because the model has less freedom to swap between equivalent options while still producing working code; the watermark can, however, appear in comments, where wording is arbitrary.

Anthropic said Claude will not be the only chatbot to carry a watermark, because other major model developers have signed the same EU Code of Practice and will roll out their own systems. The company distinguished watermarking from stylistic AI-detection tools that scan for tells like the 'this isn't X, it's Y' construction, saying pattern-spotting of that kind is fundamentally different from checking for a watermark.