A Claude watermark is a processing receipt, not an authorship verdict.
Anthropic describes two machine-readable signals: model-level watermarks embedded in text and signed C2PA provenance metadata on supported files such as SVG, PNG, and JPG.
The limitation is the useful part. A detected mark indicates that content may have been processed by Claude. It does not prove Claude was the original author. Proofreading, translation, summarization, or file conversion can produce a marked output whose underlying work came from someone else. The reverse is also unsafe: short passages, heavy edits, translation, screenshots, or stripped metadata can leave no detectable mark.
I would separate a provenance record into original_author, processed_by, edit_history, and provenance_signature. The mark records a processing event. Authorship still needs an evidence chain.
Anthropic says technical detection documentation is forthcoming, so claims about exact false-positive rates or reliable removal are premature.
Official source: https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content
Top comments (0)