Free watermark check

AI watermark checker

Find every invisible character hiding in your text — and get an honest answer about the watermark you actually came here for.

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

This free checker finds invisible Unicode in your text: zero-width characters, Unicode tag characters, variation selectors, bidirectional controls and mixed-alphabet words, each shown with its exact position. It is free, needs no account and has no word limit. It does not detect Claude's own watermark, and neither does any other public tool: Anthropic's mark is a statistical pattern in word choice that can only be read with Anthropic's secret key, which has not been released. Any tool telling you it found a Claude watermark by scanning for hidden characters is measuring something else and calling it detection.

Free, no sign-upNo account, everNo invented percentageSays what it cannot check

What this check establishes

  • Every invisible Unicode character in the text, by class and exact position
  • Words that mix alphabets — a Cyrillic 'о' inside an otherwise Latin word
  • Hidden messages encoded in zero-width characters or variation selectors, decoded and shown to you
  • What a DOCX or PDF records about itself: total editing time, save count, the software that produced it
  • A clean copy with the invisible junk removed, so LaTeX, citation managers and text search stop breaking

What it cannot establish, and neither can anything else

  • Whether Claude's statistical watermark is present. That needs Anthropic's key. No public detector exists.
  • Who wrote the text. A watermark, if one is ever readable, shows a model touched the text — not that it wrote it.
  • That text is human-written. A clean result means no hidden characters, nothing more.
  • Whether a file once carried C2PA provenance. Any re-save, screenshot or format conversion strips it.

How to check a text for AI watermarks

  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.

Does Claude watermark its text?

Yes. Anthropic announced on 11 August 2026 that Claude embeds an imperceptible watermark directly into generated text, and attaches signed C2PA provenance metadata to generated .svg, .png and .jpg files. It applies worldwide — not only in the EU — across Claude, the Claude API, Claude Code, Claude Cowork and Claude Tag, as well as Claude models served through AWS Bedrock, Google Cloud Vertex and Microsoft Foundry. Every Claude model released from 2 August 2026 marks its output from day one; older models are being retrofitted.

The trigger is Article 50(2) of the EU AI Act, applicable since 2 August 2026, which requires providers of generative AI systems to mark synthetic output in a machine-readable format.

Anthropic is also unusually clear about the limits, and those limits are worth repeating because most coverage drops them: a detected watermark means only that content may have been processed by Claude. It does not prove Claude authored it — Claude may have proofread or translated someone else's writing — and it does not prove the content was unchanged afterwards.

Why can't this tool detect it?

Because nobody can, yet. The watermark is not a hidden character you could search for. The established approach — Kirchenbauer et al.'s green-list scheme, and Google DeepMind's SynthID-Text — works by nudging the model's word choice: at each step, a secret key and the preceding words seed a pseudo-random split of the vocabulary, and the model is biased toward one half. Over a few hundred words the bias becomes statistically obvious, and a detector measures it as a z-score.

The critical property is that a word is never watermarked in itself. The same word is favoured in one sentence and disfavoured two sentences later, because the split is re-derived from the local context and the key. Without the key there is no word list to scan for, no character to find, and no way to reconstruct which half of the vocabulary was favoured. This is not a limitation we are working around — it is the mathematical point of the design.

Anthropic says technical documentation on detection is forthcoming. This tool is built so that when it arrives, only one function changes. Until then the answer here stays 'not checkable', which is the truth.

Then what are all those Claude watermark removers actually removing?

Invisible Unicode characters — mostly zero-width spaces and the odd em dash. That is a real thing to find, and this page finds it too. But it is not Claude's watermark, and there is no published evidence that Claude's text watermark involves Unicode characters at all. Technically it would be a poor choice for exactly the reason those tools exist: any Markdown renderer, form field or trim() would destroy it, which would make it useless as an AI Act compliance measure.

So a tool that strips invisible characters and tells you the watermark is gone has removed something that was probably never the watermark, and left untouched the thing that was. If your reason for wanting it removed is that you are about to submit AI-written work as your own, the honest answer is that nothing on this page helps with that, and we have not built anything that does.

We do offer a clean copy — because zero-width characters genuinely break LaTeX builds, bibliography parsers, text search and screen readers, and a document going to a supervisor should not contain them. That is copy-editing hygiene, and we label it as exactly that.

What should you do instead if you are worried about being flagged?

Check the thing your university actually runs. Institutions screen for AI-generated text with statistical detectors, and those detectors get human writing wrong at rates that are rarely published — which is how students who wrote every word themselves end up in misconduct meetings, disproportionately those writing in a second language.

Our AI check publishes its false-positive rate next to every verdict, because a score without an error rate is not a measurement. It is free for the first 1,500 words and needs no account.

Frequently asked questions

No, and no public tool can. Claude's watermark is a statistical pattern in word choice that requires Anthropic's secret key to read, and Anthropic has not released a detector. What this tool does detect — definitely, and with exact positions — is invisible Unicode: zero-width characters, tag characters, variation selectors, bidirectional controls and mixed-alphabet words. Those are real findings, but they are a different question from 'did Claude write this'.
There is no published evidence that it does. Anthropic says the watermark is embedded in the text itself and survives copy-paste, but has not said it uses Unicode characters, and specialists consider it unlikely — invisible characters would not survive a Markdown renderer or a form field, which would make them useless for compliance. If this tool finds invisible characters in your text, the likeliest sources are a website you copied from, a chat interface, or a word processor.
Yes, entirely, with no sign-up and no word limit. The detection runs in your browser using your own computer's processing, so it costs us nothing per check and there is nothing to meter.
The analysis itself happens in your browser. The analysed text is then stored on our servers inside the EU for twelve months, so we can see which kinds of documents the tool is used on and improve it, after which it is deleted automatically. It is never linked to your identity: we do not ask for a name, an email address, a university or an account, so there is nothing to link it to.
PDF, DOCX and TXT, or paste text directly. DOCX and PDF files are also read for their own metadata — total editing time, save count and the software that produced them — which is often more informative than anything in the text.
Not from the watermark today, because Anthropic has not released a detector to anyone. What universities do run is AI-detection software based on writing statistics, which is a different technology with a real and often unpublished false-positive rate. If that is your actual concern, check your text against a detector that publishes how often it is wrong.
No. The two are unrelated. Claude's watermark lives in the pattern of word choice across the whole text, so deleting characters cannot reach it. Removing invisible characters is worth doing for a completely different reason: they break LaTeX compilation, citation tools, text search and screen readers.

Sources

Related checks

Claude watermarkChatGPT 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