Claude Watermark Detector
Anthropic has announced a detection API for Claude's statistical text watermark, but it is not publicly available yet. DetectMark can still inspect your text for hidden Unicode, suspicious characters, and clipboard artifacts, which are a different thing.
- Watermark type
- Statistical text watermark
- Hidden characters
- No
- Official detector
- Announced: Announced
- Public API
- Unknown: Not yet available
- DetectMark support
- DetectMark support: Ready when API ships
- Removal
- Not locally removable
How this page works
Claude's watermark, as announced, is statistical: it lives in the pattern of word choices, not in hidden characters. That means no local tool can verify or remove it today, and any site claiming otherwise is guessing. What DetectMark can do locally is scan for invisible Unicode, lookalike characters, and clipboard artifacts that copy events and chat UIs sometimes insert.
What a real detector will look like
When Anthropic opens the detection API, the verdict will come from Anthropic's key, on Anthropic's service, and it will be one of the four labels DetectMark uses for a scan verdict: Clean, Found, Unknown, or Unsupported. No percentage, and no stylistic classifier layered on top to pad the answer.
Length will matter. A statistical watermark is measured across many word choices, so a headline or a subject line may not carry enough signal to score at all. That case is Unknown, and it will say so before you submit anything.
What the check reports today
Every category on the result panel says which of the four states it is in and why. The statistical row for Claude reads Unknown, with the reason attached: there is no public detector for DetectMark to call. It does not read Clean, because a clean verdict would be a claim nobody can currently make.
FAQ
No. Anthropic said a detection API is coming and gave no date. Every "Claude detector" on the web today checks hidden Unicode, runs a stylistic classifier, or declines to make the claim. DetectMark declines.
Not with any local tool, including this one. The watermark is encoded in generation patterns. DetectMark removes hidden characters and metadata where present, and tells you plainly when the statistical watermark status is unchanged.
Clean, Found, or Unknown per check. The statistical watermark row is always Unknown until the API opens. Hidden characters, clipboard fragments and lookalikes come back Clean or Found with positions.
No. Those are stylistic classifiers, not watermark detectors. Anthropic's mark is invisible to them.
The day Anthropic opens the detection API. The plan is to ship it the same day, update every Claude page on this site, and publish how the tool responds to a set of standard test inputs so its behaviour can be checked.
Every check runs in your browser: how it works. The limits are written down too: what DetectMark cannot remove.