Claude AI text to get hidden watermarks and signed metadata in the EU

Anthropic’s Claude will soon leave a hidden trail behind its work as the company has confirmed that newer Claude models will mark generated content from day one. Wondering why Anthropic is making this change? Well, the company has signed the EU AI Act’s transparency code for AI-generated content. It now plans to apply machine-readable watermarks across supported Claude products.

Claude’s AI text will carry an invisible mark

New Claude models launched in the EU from August 2, 2026, will support watermarking at launch. Anthropic says generated text will contain an imperceptible watermark. The mark will not change how the text looks, reads, or behaves.

The company says the watermark can survive copying and pasting. It may also remain after some forms of editing. The marking happens at the model level. That means supported output gets the mark across Claude’s different interfaces.

Anthropic plans to support the system across Claude, Claude Platform, Claude Code, Claude Cowork, and Claude Tag. Supported models accessed through AWS, Google Cloud, and Microsoft Foundry will also receive text watermarks.

Older Claude models are not being ignored. Anthropic says it is working to add marking support during the EU law’s transition period.

Files get a different kind of AI fingerprint

Text is only part of the change. For supported files, Claude will attach signed provenance metadata using the C2PA standard. Supported formats include files such as SVG, PNG, and JPG. This metadata can indicate that Claude processed the file. It can also help detect whether that file was altered afterward.

Anthropic says detection tools are also being developed. More technical details will arrive through future documentation. That said, these marks won’t provide a perfect AI detector. A detected mark means content may have been processed by Claude. It doesn’t prove Claude created the original ideas or material.

Claude could simply have translated, summarized, edited, or reformatted existing content. Heavy editing, screenshots, conversions, and short passages can also remove or weaken detectable signals. The absence of a mark won’t prove content was never AI-generated either. Anthropic itself acknowledges those limitations.

For developers using Claude inside their own products, Anthropic recommends reviewing their own Article 50 obligations. Moreover, the company says further guidance will explain how its marking system can support those requirements.