Anthropic adds invisible watermarks to Claude texts to comply with EU AI Act

Anthropic has introduced hidden watermarking for text generated by its Claude language models. The feature is now active across all versions of the AI released after August 2 and comes in response to the European Union's Artificial Intelligence Act, which mandates transparency for algorithm-generated content.

The mark is embedded directly into the text structure and remains invisible to humans, while preserving the original meaning and quality. The machine-readable label is tied to the text itself, survives copying into other programs, and withstands editing. Because the mechanism operates at the model level, it covers all Anthropic products, including Claude API, Claude Code, and third-party cloud platforms such as AWS or Google Cloud.

The company highlighted a key limitation: the presence of a watermark only confirms that the neural network participated in creating or processing the text, such as translation, proofreading, or summarization. Its absence does not guarantee human authorship, as heavy rewriting or combination with other algorithms can remove the mark.

The new rules apply globally, not just in Europe. Google and Meta, which is designated as an extremist organization and banned in Russia, have also announced their readiness to comply with European regulations. A Russian court previously outlawed Meta's activities in the country.