Free watermark check

Claude watermark checker

The tools claiming to detect Claude's watermark are counting zero-width spaces. Here is what is actually in your text, and what the watermark actually is.

Paste text

Paste, do not retype — invisible characters only survive a copy.

Upload PDF / DOCX / TXT

Click to choose a file

No account, no word limit

We never ask for a name, an email address or a university.

No invented number

Separate findings, not one percentage. Style is never counted as evidence.

Says what it cannot do

Claude's statistical watermark is checkable by nobody today, including us.

The short answer

Claude does watermark its text — Anthropic confirmed this on 11 August 2026 — but no public tool can currently detect it, including this one. The mark is a statistical pattern in Claude's word choice, readable only with Anthropic's secret key, and Anthropic has not released a detector. Tools advertising Claude watermark detection are scanning for invisible Unicode characters, which is a different thing entirely and has never been shown to be part of Claude's watermark. This checker does that Unicode scan too, accurately and with exact positions, and tells you plainly that it is not the same question. Free, no account needed.

Free, no accountNo fake detection claimExact character positionsNo word limit

What this check establishes

  • Every invisible Unicode character in your text, by class and position
  • Hidden messages encoded in zero-width characters or variation selectors
  • Mixed-alphabet words, which have no innocent explanation
  • What your DOCX or PDF records about how it was written

What it cannot establish

  • Whether Claude's statistical watermark is present — that needs Anthropic's key
  • Whether Claude wrote your text, or merely proofread or translated it
  • That a clean text is human-written

How to check text for a Claude watermark

  1. 1

    Paste your text, or drop in the file

    Copy rather than retype — invisible characters only survive a copy-paste. PDF, DOCX and TXT are read locally, and a DOCX also gives up its editing time and save count.

  2. 2

    No account, no word limit

    You are never asked for a name, an email address or a university. The text is analysed in your browser and the analysed text is then stored on our EU servers for twelve months, never linked to your identity.

  3. 3

    Read each finding separately

    You get one card per layer, each with its own confidence, and no combined percentage. Character findings are definite. Style habits are not evidence and are labelled as such.

What Anthropic actually announced

On 11 August 2026 Anthropic confirmed that Claude embeds an imperceptible watermark directly into the text it generates. The mark travels with the text when it is copied and pasted, and may persist through some editing. Separately, generated .svg, .png and .jpg files receive signed provenance metadata following the C2PA standard.

Coverage is global rather than EU-only, and spans Claude, the Claude API, Claude Code, Claude Cowork and Claude Tag, plus Claude models served via AWS Bedrock, Google Cloud Vertex and Microsoft Foundry. Models released from 2 August 2026 mark from day one; earlier models are being retrofitted during a transition period.

Anthropic also names the conditions under which the mark disappears: heavy editing, paraphrasing or translation, very short passages, output from models predating the rollout, and — for files — any metadata stripping through format conversion or screenshots.

Why no tool can read it without Anthropic

Statistical text watermarking biases the model's choice between equally good words. A secret key, combined with the words immediately preceding, seeds a pseudo-random partition of the vocabulary into a favoured half and a disfavoured half, and the model's scores are nudged toward the favoured half. Any individual sentence still reads naturally; across several hundred words the bias becomes measurable, and a detector computes how far the favoured-word count sits above chance.

The partition is re-derived at every position from the key and the local context. So the same word is favoured here and disfavoured three words later, and there is no fixed list of watermarked words for anyone to scan. Without the key, a detector cannot reconstruct a single partition — meaning it cannot even begin the count. This is not an engineering gap. It is the property that makes the scheme work at all: if the mark were readable without the key, it would also be removable by anyone who cared to.

What about Claude watermark removers?

They remove invisible Unicode characters. Whatever else that achieves, it does not remove a statistical watermark, because a statistical watermark is not made of characters. The only thing that reliably destroys one is substantial paraphrasing — rewriting the text with a different model — which changes what the text says and, in an academic context, is a considerably worse idea than the problem it claims to solve.

We have not built a remover and will not. The check we run is the same for a student before submission and for an examiner afterwards, which is only possible because it does not claim to defeat anything. A tool designed to beat a detector has no reason to tell you when it is wrong.

Frequently asked questions

Yes. Anthropic confirmed on 11 August 2026 that Claude embeds an imperceptible watermark directly in generated text, worldwide, across all Claude products and cloud partners. Models released from 2 August 2026 mark output from day one, with older models retrofitted during a transition period.
Not today. Anthropic has announced a detector but has not released one, and the watermark is only readable with their secret key. No public tool can check it, and any tool claiming to is scanning for hidden Unicode characters instead — which has never been shown to be part of Claude's watermark.
Partly. Anthropic says it may persist through some editing but is lost through heavy editing, paraphrasing or translation, and it is unreliable in very short passages. That is a documented weakness of statistical watermarking generally: light edits leave most word choices intact, while a full rewrite replaces them.
No, and Anthropic says so explicitly. A detected watermark means only that content may have been processed by Claude. Claude may have proofread, translated or reformatted text a person wrote, and the content may have been changed after the mark was applied.
They remove invisible Unicode characters, which is not the same as removing Claude's watermark — the watermark is a statistical pattern in word choice and is untouched by deleting characters. Only substantial paraphrasing degrades a statistical watermark, and that changes what your text says.

Sources

Related checks

ChatGPT watermarkInvisible charactersFree watermark checkerAI check

The real question: will your work be flagged as AI?

That is what universities actually run, and those detectors are wrong about human writing more often than they admit. Our AI check publishes its false-positive rate beside every verdict. Free for the first 1,500 words.

Run the free AI check