Anthropic plans invisible text watermarks for new Claude models

Anthropic plans to add machine readable marks to content generated by future Claude models. The company says models introduced after August 2, 2026 will embed invisible watermarks in generated text and attach signed provenance metadata to supported files.

The policy is part of Anthropic’s commitments under the EU AI Act’s Article 50(2) Code of Practice on Transparency of AI Generated Content. The company says the marks will apply worldwide when people use supported Claude models, not only within the EU.

Anthropic’s approach uses two different methods. Text generated by covered models will receive an embedded watermark. Files such as SVG, PNG and JPG images will carry signed provenance metadata based on the C2PA standard, short for Coalition for Content Provenance and Authenticity.

The company says the text watermark will not be visible and will not alter the meaning, quality or readability of an answer. Because it is embedded in the text, Anthropic says it can remain present when someone copies and pastes content and may survive some editing.

Marks across Claude products

The planned marking system will cover output from supported models across Anthropic’s products. This includes the Claude API, the Claude consumer app, Claude Code, Claude Cowork and Claude Tag. Text watermarks will also apply when customers access supported Claude models through cloud providers including AWS, Google Cloud and Microsoft Foundry.

Metadata support may vary by platform and file feature. Anthropic says signed provenance data will be added only where a product supports file processing and the relevant format. A C2PA label can indicate that Claude processed a file and can help show whether the file has later been tampered with.

Older Claude models are not yet covered in all cases. Anthropic says it is working to add marking support to models released before the August 2026 threshold, which are covered by an EU transition period.

The company also plans to provide tools or technical documentation that allow customers and outside parties to check for Claude’s marks. It has not yet described the detection method for text watermarks or said when it will become available.

A label is not proof of authorship

Anthropic stresses that a detected mark should not be treated as conclusive proof that Claude authored content. A user may ask Claude to translate, proofread, summarize or reformat material created by a person. In that case, the result could carry a Claude mark even though the underlying ideas and source material came from elsewhere.

Likewise, content can change after Claude has processed it. Users may edit it, publish only an excerpt or combine it with writing, images or data from other sources.

The absence of a mark also does not prove that content was created without AI. Anthropic says a watermark may become difficult to detect after heavy editing, paraphrasing, translation or mixing with other text. Very short passages may not contain enough material for a reliable signal. File metadata can also disappear after format conversion, re saving or taking a screenshot.

Anthropic’s Claude Help Center explains its planned marking system as a transparency measure rather than a definitive record of origin. For content teams, the change means Claude output may increasingly carry a technical signal of AI involvement, but that signal will still require context and human judgment.

Sources

• How Claude marks AI-generated content – Claude Help Center
• Anthropic says Claude models in the EU will now add invisible watermarks to generated text and C2PA metadata to generated files, to comply with the EU AI Act – The Register