New Claude models will embed invisible watermarks in generated text
Anthropic has announced that starting from August 2, 2026, new Claude models deployed in the European Union will incorporate machine-readable invisible watermarks into generated text. This measure will apply globally wherever Claude is offered, including via its API, applications, and through partners like AWS, Google Cloud, and Microsoft Foundry.
The watermarking mechanism for text is distinct from file provenance. For text, invisible watermarks will be woven directly into the model's responses without altering meaning, quality, or readability, and are designed to persist through copying and pasting and partially through editing. This system is separate from the C2PA standard. For files (e.g., .png, .jpg), Claude already attaches C2PA-compliant metadata signatures to indicate AI origin.
Models released before August 2, 2026 (including the current Claude Opus 5) are in a transitional period and do not embed text watermarks, though Anthropic is working to add support. Currently, there are no public tools to detect these text watermarks; Anthropic states it is developing such tools and will release technical details later.
The timing aligns with the EU AI Act's Article 50, which from August 2, 2026, mandates labeling of AI-generated content, with potential fines for non-compliance. A key technical question remains regarding the watermark's resilience to paraphrasing or translation by other AI models, which in other industry experiments has significantly reduced detection accuracy.
cryptonews.ru11m ago
