The marking is part of Anthropic's work with EU rules for clearer provenance of AI content. In text, supported Claude models add an invisible pattern that can remain when the text is copied. Files such as PNG and JPEG can also receive Content Credentials – digital information showing that Claude processed the file.
The limitations matter. A person may have written the whole text and only asked Claude to proofread or translate it. The finished version can still carry a mark. In the other direction, extensive editing, translation or a very short text can make the mark difficult to detect.
Anthropic's text detector is also only available in a limited preview for certain organisations and companies.
Treat the mark as a clue, not a verdict. It can show that Claude was involved, but not how much of the thinking, writing or editing came from a person.