English: Anthropic has started embedding machine-readable watermarks into content generated by its Claude models. This isn’t a visible stamp — the signal is woven into the text at a statistical level, so to the human reader, the output looks completely normal. The mark can persist when the text is copied and pasted, and in some cases, it survives partial editing. For non-text files (images, documents), the company is using the C2PA standard to embed provenance metadata. The move is largely driven by the need to comply with the EU AI Act’s transparency requirements, though the policy applies globally. Anthropic plans to release technical documentation and detection tools later, allowing platforms to automatically identify such content. However, the company acknowledges limitations: heavy editing, translation, or mixing with other text can weaken or erase the signal, and a detected mark only indicates possible AI involvement, not definitive authorship.
About the Author
