Claude watermark checker
Two checks that actually work, and one straight answer about the check nobody can run. Paste text to find hidden characters, or pick a file to look for a C2PA provenance manifest. Everything runs in your browser — nothing is uploaded.
Drop a PNG, JPG, SVG, or PDF here — or choose one. It is read on your device only.
Privacy: both checks are client-side. Your text and your file never leave your device — this page has no server-side processing and keeps no copy.
What can be checked, and what can't
"Claude watermark checker" is a fair thing to search for, but the honest answer has three parts, because Claude marks content in more than one way — and only some of those marks are readable by anyone other than Anthropic.
| Mark | Where it lives | Checkable here? |
|---|---|---|
| Hidden characters | Invisible Unicode inside the text — often a side effect of AI or word-processor formatting rather than a deliberate mark | Yes — exactly |
| C2PA provenance metadata | Signed manifest attached to supported files (.png, .jpg, .svg) recording that Claude produced or edited them | Yes — presence detection |
| Statistical text watermark | The pattern of word choices itself, embedded during generation | No — nobody public can |
That last row is the one every "Claude watermark detector" site quietly skips. Anthropic has not released a public detector, so no third-party tool can read the statistical mark — not this one, not the ones charging for it. If a checker returns a confident "watermark detected / not detected" verdict on plain text, it is reporting something else, or nothing at all. The mechanics are in our explainer on the Claude watermark.
How the file check works
C2PA is an open provenance standard: a signed manifest embedded in
the file that records how it was made. Anthropic attaches C2PA data
to supported file types, and so do a growing number of camera and
editing tools. This checker reads your file locally and looks for
the structural markers of such a manifest — the PNG
caBX chunk, JUMBF container boxes, and C2PA labels.
Two caveats worth stating plainly. It detects presence, not validity: verifying the cryptographic signature against a trust list is a separate job this page doesn't attempt. And a missing manifest proves nothing — most platforms strip metadata the moment they re-encode an upload, which is exactly why in-content watermarks exist at all.
Want to know when a real detector exists?
If Anthropic ships a public detector — or if we build something more capable than presence detection — the people on this list hear first. One email, nothing else.
Frequently asked questions
Is there a Claude watermark checker that actually works?
Not for the statistical text watermark. Anthropic has not released a public detector, so no third-party tool — including this one — can read it. What can be checked is the C2PA provenance metadata Claude attaches to files, and the hidden formatting characters that often travel with AI text. This page checks both.
What does this checker scan for?
Two things. In text, it finds invisible and non-standard Unicode characters such as zero-width spaces, narrow no-break spaces, and soft hyphens. In files, it scans for a C2PA provenance manifest — the signed metadata standard Anthropic uses to mark files that Claude generated or edited.
Is my text or file uploaded anywhere?
No. Both checks run entirely in your browser. The file is read from your disk into memory on your own device, nothing is sent to any server, and there is no account and no stored copy.
Does a clean result mean the text is not from Claude?
No. A clean result means only that no hidden characters and no C2PA manifest were found. A statistical watermark would still be present and unreadable by any public tool, and file metadata is stripped whenever a platform re-encodes an upload. Absence of evidence is not evidence of absence here.
Can this remove the Claude watermark?
No. This page only inspects; it changes nothing. For the character-level cleanup that can honestly be done, use our ChatGPT watermark remover — though that strips hidden characters only and cannot touch a statistical watermark either.