In brief: Anthropic is equipping future Claude models with a text watermark imperceptible to readers in order to comply with the EU AI Act’s labeling requirement for AI content, which has been in effect since August 2, 2025.
Anthropic will equip future Claude models with an invisible text watermark to comply with the labeling requirement for AI-generated content under the EU AI Act, which has applied since August 2. Other major model providers that have signed the corresponding Code of Practice are also implementing their own watermarking procedures.
In a blog post, Anthropic explained how the watermarking procedure for Claude texts works. It is based on the fact that language models generate text word by word and, for many formulations, can choose between several equally valid candidates – for example, between “overcast” and “grey” in a sentence about the weather. In such low-stakes decisions, the choice is normally made by a random number generator. In the watermarking procedure, a cryptographic key is instead used together with the preceding words to steer the word choice. Anyone in possession of this key can subsequently check whether a text sequence matches the decisions Claude would have made using the key, and from that derive a probability that the text originated from Claude. The method is based on the SynthID-Text technique developed by Google DeepMind.
According to Anthropic, the procedure has no practical impact on the quality or content of Claude’s outputs. For readers, watermarked text is indistinguishable from unwatermarked text; no additional characters or hidden symbols are inserted, no additional tokens are required, and usage does not become more expensive. Furthermore, the watermark contains no identifying information and cannot be traced back to a specific person, organization, or particular chat. In internal tests as well as in the original SynthID-Text paper by Google DeepMind, which examined the technique using a portion of Gemini traffic and a comparison of user ratings (thumbs-up/thumbs-down), no statistically significant differences from non-watermarked outputs were found.
For compliance officers at companies using Claude or comparable models, it is relevant that the labeling requirement for AI-generated content has applied since August 2, 2025, to providers serving the EU market. Since several major model providers have signed the same Code of Practice and are each introducing their own watermarking implementations, the procedure is not specific to Claude but part of an industry-wide approach to implementing the transparency requirements of the EU AI Act. Companies should examine whether and how detection tools for such watermarks can be integrated into their own compliance and verification processes for AI-generated content, especially since detection requires the corresponding key and cannot readily be performed by third parties.
Source: www.anthropic.com · Published August 13, 2026
Lumi AI News — AI-assisted curation pursuant to Art. 50 EU AI Act. Paraphrasing and classification by Lumi News Pipeline v1.8.3.