Anthropic has started adding machine-readable marks to content generated by newer Claude models as part of its compliance with the European Union’s AI Act. Claude-generated text can now contain an imperceptible watermark, while supported image files can include digitally signed provenance metadata.
The policy applies to Claude models launched in the EU on or after August 2, 2026, and Anthropic says the markings are used worldwide across supported Claude products. Existing models released before that date are being updated during a transition period.
Anthropic detailed the system in its official guidance on AI-generated content. The company says text watermarks are embedded directly into generated text, meaning they can persist when content is copied, pasted, or lightly edited.
EU Rules Require Machine-Readable AI Marking
Article 50 of the EU AI Act took effect on August 2 and requires providers of generative AI systems to add machine-readable marks that help identify AI-generated or manipulated material. The European Commission’s transparency guidance says providers can comply through the voluntary Code of Practice or use another method that provides equivalent protection.
Anthropic has signed the Code of Practice on Transparency of AI-Generated Content. Its text watermark is invisible to readers and does not change the intended meaning or readability of Claude’s response, according to the company.
For supported image formats including PNG, JPG, and SVG, Anthropic uses digitally signed provenance metadata based on the C2PA standard. That metadata can indicate that a file was processed by Claude and whether it has subsequently been altered.
Some Claude Users Object to the Policy
The change has prompted criticism from some Claude users on Reddit, particularly over cases where AI is used for editing, rewriting, or coding rather than generating an entire piece of work. Some argue that a watermark could suggest greater AI involvement than actually occurred.
Anthropic acknowledges that limitation. A detected watermark indicates that content may have been processed by Claude, but does not prove that Claude originally wrote all of it because users may have asked the model to proofread, translate, summarize, or modify existing material.
The company also says the absence of a detectable watermark does not prove that material was human-written. Heavy editing, paraphrasing, translation, short passages, or stripped file metadata can make markings harder or impossible to detect.
Anthropic plans to provide detection tools for users and third parties and says more technical documentation is coming. The markings currently apply across supported uses of Claude, including Claude, Claude Code, Claude Cowork, the Claude API, and deployments through AWS, Google Cloud, and Microsoft Foundry.
Featured image credits: SlideTeam
For more stories like it, click the +Follow button at the top of this page to follow us.