Claude Watermark Checker

What this checker does

It measures your text and shows you the measurements. Word counts, vocabulary diversity, entropy, how much sentence lengths vary, which phrases repeat, how punctuation is distributed. It also scans for hidden characters and hidden markup.

What it does not do is tell you who wrote the text. That distinction is the whole point of the page, so it is worth being blunt about why.

How Claude’s text watermarking works

Anthropic published the mechanism in August 2026. It is worth understanding, because almost everything written about “AI watermarks” online describes something else entirely.

A language model generates one word at a time, picking from a list of candidates. Many of those choices are low-stakes — several words would work equally well. Watermarking uses those choices. Instead of picking randomly, the model uses a secret key plus the preceding few words to decide. Anyone holding the key can check whether a text’s word choices match what that key would have produced.

The third row is why this page says NOT AVAILABLE rather than giving you a score. Detection is a keyed operation. Anthropic has announced a detection API but has not released it. Until it exists, nobody outside Anthropic can check a text against the watermark — including us, and including every site currently claiming to.

What these statistics can and cannot tell you

The numbers on this page are real. They are also frequently misunderstood, so here is what each one is actually measuring.

On burstiness specifically

Burstiness has a documented association with machine-generated text: models tend to produce sentences of similar length, while human writing varies more. That association is real, and it is why the measure appears here.

It is still not a test. Uniform sentence length is normal in technical documentation, legal drafting, instruction manuals, and text translated from another language. Varied sentence length is easy for a model to produce if asked. A low burstiness score tells you the sentences are similar in length. That is all it tells you.

Why we do not show a percentage

Because it would be made up. There is no calculation available to this page that yields a defensible probability that text was machine-written, and presenting one would be inventing a number.

It would also be harmful. Evaluations of AI-text detectors have repeatedly found false-positive rates high enough to matter, and the errors are not evenly distributed — writing by people using English as a second language is flagged disproportionately, because it tends toward simpler and more uniform construction. People have been accused of cheating on the strength of those scores.

If you have been accused on the basis of a detector, that unreliability is the thing to raise. Ask which tool was used, what its published false-positive rate is, and whether the result was treated as evidence or as a prompt to have a conversation.

Hidden characters in AI text

This is a separate phenomenon that gets confused with watermarking constantly.

Text copied out of any web interface can pick up zero-width characters, non-breaking spaces, curly quotes, and occasionally HTML comments or invisible elements. These come from editors, content management systems and clipboard handling. They are not a watermark, they are not deliberate, and removing them has no effect whatsoever on Claude’s watermark.

They are still worth finding. They break CSV imports, database lookups, search matching and code identifiers, and bidirectional control characters can disguise what a line of text actually says. The checker reports them in its own section for that reason, and the Invisible Character Detector will remove them. If you only need the plain numbers for a piece of text, the word counter gives words, characters, sentences and reading time without any of the statistical analysis on this page.

Using the checker

1. Paste your text, or load a TXT, Markdown or HTML file. Everything is read in your browser.
2. Press Analyse text. The watermark status appears first, because it is the question most people arrive with.
3. Read the statistics alongside the note explaining each one.
4. Check the artifact section for hidden characters, and follow the link if you want them removed.

DOCX and PDF are not supported. Reading either needs a parsing library, and this site ships no third-party JavaScript at all. Open the document and paste the text instead.

Privacy

Nothing is uploaded, nothing is stored, and nothing is written to the URL. That last one is deliberate: people paste unpublished drafts, private correspondence and unreleased work into tools like this, and a querystring ends up in browser history, server logs and referrer headers.

If Anthropic’s detection API becomes available and we add it, that will involve sending your text to Anthropic — so it will be opt-in, per use, and labelled at the point of use. It will never be the default, and the local analysis will keep working with it switched off.

What happens when the detection API ships

The tool is built to accept it. There is an adapter interface waiting for a real endpoint, and the status panel is driven by that adapter rather than hardcoded — when a detector is connected, the panel changes on its own.

At that point this page can answer the question properly, for the first time. Until then, the honest answer is the one shown: not available, here is what your text actually measures, and be sceptical of anyone claiming otherwise.