Provenance Lens
Detect the Claude Text Watermark
See what can be verified about Claude text watermarks today. Anthropic documents a mark, but no public detection mechanism exists yet.
Text: not publicly detectable. Anthropic states that Claude models released from August 2026 embed a statistical watermark in generated text, but has not published a detection mechanism. No tool — including this one — can currently verify a Claude text watermark. Any service claiming otherwise cannot substantiate it.
Images: verifiable now. Anthropic applies signed C2PA provenance metadata to supported generated image formats. That metadata is checkable, and this page checks it.
What the mark would prove, if it were detectable
Anthropic's own framing is that the mark indicates Claude processed the text, not that Claude authored it. Text that a person wrote and then asked Claude to lightly edit would carry the same mark as text Claude wrote entirely.
Reported limitations
- Paraphrasing or substantial rewriting is reported to destroy the mark.
- Reliable detection is reported to require roughly 100 or more tokens.
- The mechanism has not been independently verified, because it has not been published.
Support status: UNSUPPORTED (text) · DETECTED where present (image C2PA) · Last reviewed: · Source: Anthropic
Why we do not offer a guess
We could ship a statistical estimator and call it a Claude detector. It would produce confident numbers with no verifiable basis, and people would make decisions about students, employees and contributors on the strength of them. We report UNSUPPORTED until a public mechanism exists.