Check
Claude Watermark Detector
Claude watermark detector searches have one accurate answer as of August 2026, and it is that no public detector exists.
Anthropic has said it will support detection by users and third parties. It has not published the tooling, the documentation, or a timeline.
No public detector exists. Nobody outside Anthropic can currently verify a text mark.
Runs in your browser. No signup, no upload, nothing stored.
Why nobody else can build one
A statistical watermark is detectable only with the key that seeded it. The test is a likelihood computation against a known pseudo-random rule, and without the rule there is no computation to perform.
This is a deliberate design property rather than an oversight. A mark anyone could detect is a mark anyone could strip, and the system would defeat itself on the first day.
So the absence of third-party detectors is not a gap in the market waiting to be filled. It is the design working as intended.
What Anthropic has committed to
Its documentation says it will support users and other third parties in detecting Claude's marks, with technical guidance to follow.
For enterprise and API customers it goes further, advising them to independently assess what EU AI Act Article 50 requires of them as deployers — which implies detection tooling those customers can actually run.
The consequence people miss
If a detector ships, it works retroactively. The mark is already present in every piece of text produced since 2 August 2026.
Anything you publish today that carries the mark stays markable whenever the tooling appears, however long that takes. That asymmetry — marking started, detection did not — is the thing actually worth planning around.
What a detector will and will not say
Anthropic is explicit that a detected mark is not fully conclusive. It indicates content may have been processed by Claude, and nothing stronger:
- It does not prove Claude authored the content
- It does not prove the content is unchanged since generation
- Absence does not prove Claude was not involved
- It returns a likelihood, not a verdict
The risk in how it gets read
The technical caveats above are careful and correct. They will not survive contact with a spreadsheet column labelled 'AI: yes/no'.
The realistic danger is not that detection will be inaccurate. It is that a probabilistic score will be treated as proof by someone who stopped reading at the headline number.
What to watch for
The signal that this has changed will be Anthropic publishing detection documentation, not a third-party tool appearing.
Any product claiming to detect the Claude text watermark before that documentation exists is running a style classifier and relabelling the output.
Source
The claims on this page are drawn from primary documentation and reporting rather than from other tools’ marketing copy.
Anthropic: How Claude marks AI-generated content →Related
Check what your text and files actually carry
Runs in your browser. No signup, no upload, nothing stored.