Anthropic will watermark Claude's text to meet EU rules

Anthropic says text generated by Claude will now carry a watermark, a response to the EU AI Act's Transparency Code that took effect on August 2. The watermark is embedded at the model level and, per Anthropic's support documentation, becomes part of the text itself, so it travels when content is copied and pasted elsewhere and can survive some editing. For files rather than raw text, the company uses C2PA, the open provenance standard already used for images. Every model released after August 2, 2026 includes the technology automatically, with support for older models to follow.

The open question is durability. TechCrunch notes it is not clear how much editing removes the mark, and Anthropic had not answered that by publication. The stated aim is to make undisclosed AI text detectable, the practice some call "Claudefishing," passing off machine-written text as your own.

Why it matters

If you paste Claude's output into your own work, assume it may be identifiable as AI-generated, which matters for students, writers, and anyone in a setting where AI use must be disclosed. How easily a light edit strips the mark is still unknown, so do not treat either its presence or its absence as proof of anything yet.