Anthropic said it will add machine-readable watermarks to text and provenance metadata to files generated by Claude, complying with the European Union AI Act's transparency rules, TechCrunch reported.
The watermarking covers models released on or after Aug. 2, 2026, including Claude Sonnet 4.6 and Claude Haiku 4.5, and applies globally rather than only to European users, according to Interesting Engineering. Text watermarks are invisible to readers, persist through copying and pasting, and may survive some editing, TechCrunch reported. Files carry signed provenance metadata under the C2PA standard.
The requirement stems from Article 50 of the EU AI Act, which took effect Aug. 2 and requires AI companies to mark generated or edited content in machine-detectable ways, according to Interesting Engineering. OpenAI, Google, Meta and Microsoft have also committed to the EU's code of practice, TechCrunch reported.
Anthropic cautioned the marks are not proof of authorship: Claude can process human-written text, and heavy editing or format conversion can strip a text watermark or a file's metadata, according to Interesting Engineering.
For anyone building content-provenance or AI-detection tooling, this gives a concrete standard to build against now: C2PA for files and Anthropic's own watermark for text, rather than a future promise to wait on.