Anthropic says Claude output can carry two kinds of marks: invisible text watermarks and signed C2PA Content Credentials on supported files. A detected mark suggests Claude processed the content, but does not fully confirm its origin. A missing mark does not prove the content was written by a person.
The two methods
According to Anthropic's help article on marking AI-generated content:
Text watermarks are woven into the text and are not visible. They do not change meaning, quality or readability, and can persist when text is copied or lightly edited.
Content Credentials use the C2PA open standard to attach signed provenance metadata to supported files such as PNG and JPEG. If a signed label is present, it signals the file was processed by Claude.
Where this comes from
Anthropic says it has signed the EU AI Act Article 50(2) Code of Practice on Transparency of AI-Generated Content. Claude models launched in the EU on or after August 2, 2026 are to support machine readable marking at launch, and generated text will carry embedded watermarks. Support for earlier models is in progress, with a transition period in the law for systems launched before that date.
Checking a file
A free Claude Content Checker verifies Content Credentials on files. Watermark detection for text is in private preview for eligible organizations, with access requested through a registration form. So an ordinary reader cannot check text for a watermark today, according to the page.
What a result means
A detected mark suggests Claude processed the content. It does not prove Claude created all of it: Claude may have only edited or translated it, and the content may have been changed afterward.
A missing mark proves nothing. Anthropic lists reasons it can be absent: older models, heavily edited or very short text, stripped file metadata, or unsupported platforms.
What this means for ordinary people
- Do not use "no mark" as proof that a message, essay or image is human.
- Do not accuse a student or colleague on the strength of a mark alone.
- For images, a Content Credentials label is a useful hint. Anthropic notes that stripped file metadata is one reason a mark can be missing.
- A scam message does not become safe because it lacks a mark, and does not become fake because it has one. Check the sender and the request instead.
Good habits still beat detection: verify through a second channel, look at the source, and ask for the original file.
Quick check
Can I check text for a watermark myself?
Anthropic says watermark detection is in private preview for eligible organizations. The free checker is for Content Credentials on files.
Does a mark prove Claude wrote everything?
No. Claude may have only edited or translated it, and the content may have changed afterward.
Sources
Last reviewed: 19 October 2026.
