Anthropic text watermarking will mark new Claude output worldwide

Anthropic text watermarking will be built into output from new Claude models launched in the European Union on or after August 2, 2026. For people using Claude at work or school, the change means certain AI-generated text can carry a machine-readable signal even after it is copied elsewhere, although Anthropic says that signal cannot establish who originally wrote the material.

Anthropic said the rollout supports commitments it made after signing the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content. The company says the marking will apply worldwide for supported models, rather than only in Europe.

How will Anthropic text watermarking work?

Claude will use two different methods, according to Anthropic's support documentation. Text produced by supported models will contain an imperceptible watermark embedded in the text itself. Anthropic says it will not alter the meaning, quality or readability of the response.

Because the mark is embedded in the text, it should follow copied-and-pasted passages and may survive some editing, Anthropic says. The company has not published technical detection details yet, but says it plans to help users and third parties identify Claude watermarks.

Files are handled differently. When Claude generates supported file types, including SVG, PNG and JPG files, it will add digitally signed provenance metadata where supported. That metadata follows the C2PA open standard, which records information about a file's origins and whether it has been tampered with.

• Text: embedded, machine-readable watermark.
• Supported files: signed provenance metadata using C2PA.
• Coverage: supported output across Claude, Claude Platform API, Claude Code, Claude Cowork and Claude Tag.
• Cloud access: text watermarks also apply to supported models used through AWS, Google Cloud and Microsoft Foundry.

What can a Claude watermark tell you?

A detected mark is evidence that Claude may have processed content, not proof that Claude wrote it. Anthropic says a mark could appear after someone used Claude to proofread, translate, summarize or convert existing material. The underlying ideas, data or original text may have come from elsewhere.

The reverse is also true: no detected mark does not show that content was written without AI. Anthropic says marks can become undetectable after heavy editing, paraphrasing, translation or mixing with other writing. Short passages may not have enough text for a reliable signal, while older models, unsupported features and stripped file metadata can also leave no detectable mark.

Which Claude models are covered?

The commitment applies at launch to Claude models introduced in the EU on or after August 2. Anthropic says it is still working to add marking support to models released before that date during a transition period, so existing models should not be assumed to be marked yet.

The plan adds a provenance signal to Claude's output, but it does not create a definitive AI-content test. Until Anthropic releases its detection documentation, users and platforms will not know the precise mechanics for checking those signals.

This story draws on original reporting from TechCrunch.