Anthropic is moving to make content generated by its Claude AI models more identifiable, with the company planning to introduce machine-readable markings across newer models and the products that use them.
The move means AI-generated text from supported Claude models will carry an invisible watermark, while certain files created by Claude will include digitally signed provenance information. Anthropic said the new marking system is being introduced to meet transparency commitments under the EU’s AI framework.
Claude models launched in the EU on or after Aug. 2 will support the marking system from launch. Anthropic is also working to bring the technology to models released before that date, which are covered by a transition period under the EU rules.
Under the new system, Claude will embed an imperceptible watermark directly into generated text. The watermark is designed to be invisible to users and should not affect the meaning, readability or quality of the content.
Because the marking is implemented at the model level, Anthropic said it will apply regardless of where users access Claude. Supported output across Claude Platform, Claude, Claude Code, Claude Cowork and Claude Tag will be covered.
READ: Anthropic to develop custom AI chips for Claude (August 5, 2026)
The company also plans to apply the marking globally wherever supported Claude models are available, rather than restricting the system to the European market.
Anthropic expects the watermark to remain with text when it is copied and pasted and survive some forms of editing. However, heavy editing, paraphrasing, translation or combining the text with other material could make the mark harder to detect.
Claude will also use a separate marking method for supported files.
When Claude generates files such as SVG, PNG and JPG images, Anthropic plans to attach digitally signed provenance metadata based on the Coalition for Content Provenance and Authenticity, or C2PA, standard.
The metadata can indicate that a file was processed by Claude and help determine whether it was subsequently altered.
This feature will depend on the file type and platform. Anthropic said signed provenance metadata may not be supported in every environment.
The company said embedded watermarks will also work when supported Claude models are accessed through cloud platforms including AWS, Google Cloud and Microsoft Foundry, although the availability of signed metadata may vary.
Anthropic is also working on tools that will allow users and third parties to check whether text or files contain Claude’s markings.
The company plans to publish technical documentation explaining how detection will work. A detected mark would indicate that content may have been processed by Claude.
However, Anthropic emphasized that the markings should not be treated as definitive proof of authorship.
For example, a person could write content themselves and then use Claude to translate, proofread or summarize it. The resulting material could carry a Claude mark even though the underlying content originated elsewhere.
READ: Anthropic signs $19 billion AI data center lease with TeraWulf (July 6, 2026)
Similarly, content may lose its detectable marking after extensive editing or other transformations.
Anthropic also warned against using the system as a simple test for determining whether something was created by AI.
However, the marking system will not be foolproof. Claude-generated text may not always return a detectable signal, particularly when it comes from older models that do not yet support the technology, has been substantially rewritten or paraphrased, or is too short for the watermark to be reliably identified.
For files, provenance information can also disappear through processes such as format conversion, screenshots or saving the file again.
That means the presence of a Claude mark could provide a useful signal about how content was processed, but its absence would not prove that the content was created entirely by a human.
While Anthropic is starting with models launched from August, the company said it is working to add marking capabilities to Claude models released before that date.
The move comes as regulators and AI companies increasingly focus on making AI-generated content easier to identify as generative AI becomes more common across text, images and other digital formats.
For developers using Claude to build their own products, Anthropic said they will still need to assess which transparency requirements apply to their services under the EU AI Act. The company plans to provide further technical guidance on its marking and detection systems as implementation progresses.