Reading your result
How to read a Claude watermark check result
You ran a check and got an answer. This page explains what that answer actually establishes, what it cannot, and how to act on it depending on whether you are publishing text, reviewing someone else's work, or documenting a process.
Available now in this browser
- Locate supported hidden Unicode and control characters.
- Read rich clipboard HTML after an explicit permission action.
- Preview and remove only selected text artifacts.
- Validate readable C2PA Content Credentials in selected files.
Everything in that list is verifiable by you, not just asserted by us. Each finding names the character, gives its code point and position, and shows the words around it. If you disagree with a finding you can check it against any Unicode reference and see who is right. That is the standard we hold results to.
What each check actually covers
Hidden Unicode
Zero-width characters, word joiners, bidirectional controls, Unicode Tags, and selected control characters. These are real code points present in your string. They travel with copied text and are invisible in most editors, which is precisely why a tool is useful here.
Clipboard formatting
When you copy rich text, the clipboard carries an HTML layer beside the plain text. That layer can hold inline styles, class names, comments, data attributes, and office-suite markers. It is inspected as text and never rendered.
C2PA Content Credentials
For supported files, a signed manifest can be read and validated locally. This is the strongest evidence available in the whole category, because signatures either verify or they do not.
Not available as a general third-party check
Anthropic describes an imperceptible model-level mark in supported Claude text, but says details about third-party detection are forthcoming. This site therefore does not label zero-width characters, clipboard formatting, writing style, or detector scores as that mark.
Specifically unavailable to anyone publicly:
- Reading Anthropic's model-level text mark.
- Removing that mark, or confirming it has been removed.
- Proving a passage was or was not produced by a language model.
- Reliable stylometric detection. Style-based classifiers produce false positives on non-native writers and formal prose, and should not carry consequences on their own.
How to interpret a negative result
“No inspected text artifacts found” means only that this checker did not find the signals it explicitly supports. It does not prove that text was written by a human, lacks a model-level mark, or never passed through Claude.
The useful mental model: this tool answers “is my text clean?”, not “who wrote this?”. The first question it answers completely. The second it does not attempt, because no public tool can answer it honestly today.
What would change this page
If Anthropic publishes a third-party detection mechanism, this status page and the tool will be updated to reflect exactly what it can and cannot establish — including its error rates. Until then, treating character findings as a proxy for official detection would be a fabrication, and this site will not make that trade.
Last reviewed: August 12, 2026.
Reading a result correctly, by situation
You are publishing or committing text
This is the case the tool serves completely. You want no stray invisible characters in what you ship. Paste, review, remove, copy. A clean result here means exactly what you need it to mean: the text contains none of the artifacts that break formatting, word counts, diffs, or string comparisons.
You are an educator or reviewer
This is the case where misreading a result does real damage. Nothing on this site — and no public tool at the time of writing — can tell you whether a student or contributor used AI. Anthropic has announced a watermark detection API but has not released it; even once it exists, Anthropic is explicit that a detected mark only means Claude may have been involved somewhere, and that asking Claude to proofread your own writing can carry the mark too. A clean result is not exoneration and a finding is not an accusation. Character artifacts say something about how text moved between applications, and nothing about who composed it.
You are on a compliance or trust team
The defensible output is a record, not a verdict: what was inspected, what was found, what was changed, and what remains outside the method's scope. That record holds up because every claim in it is independently checkable. A confidence percentage from a stylometric classifier does not hold up the same way.
You are checking an image or media file
Use the C2PA checker instead. File provenance is cryptographic and gives firmer answers than anything available for plain text — with the caveat that most files legitimately carry no credential at all.
Why this site states limits so insistently
Most tools in this category do the opposite. They return a percentage, a verdict, or a reassuring green tick, because a confident answer converts better than an honest one. The cost lands on people who get accused of something they did not do, or who ship text believing it was cleared when nothing of the sort was established.
The trade here is deliberate: a narrower claim that is actually true. Every finding names a character, gives its position, and shows its context, so you can check the tool rather than trust it. Where the method ends, the page says so instead of guessing.