If Claude wrote it, it’s marked.
We sanitise every kind of AI watermark in seconds.
- Every kind of watermark
- 100% of detectable marks removed
- Free. No account needed.
We scan for every kind of watermark
2free- Hidden charactersAwaiting text or file
What it is
Invisible spaces and joiners: real characters that hold a position in your text and display as nothing.
Who puts it there
Many major AI models leave them behind, and they survive copy, paste and export.
How we remove it
Find every one, show you where it sat, remove it, and prove none are left.
- MetadataAwaiting a file
What it is
A signed record inside the file naming the tool that made it, and when.
Who puts it there
Every major AI model signs its files. Claude included.
How we remove it
Strip the record and show you the file before and after, byte for byte.
- Statistical watermarkAwaiting text
What it is
Nothing is added to your text. The mark is the pattern of word choices itself, and no tool can show it. Yet.
Who puts it there
Claude and other major models, everywhere, with no off switch.
How we sanitise it
A structurally engineered rebuild of every sentence that keeps your facts and your length, and hands you the receipts.
- Pasted text
- Word documents
- PNG and JPG
AI tools now mark what they make. Invisibly, and without telling you.
The story, as covered by:
Marked, invisibly
Claude watermarks what it writes. Universities, employers and anyone else will be able to see it soon.
A public detector is coming
Anthropic says its release of a checker anyone can use is imminent.
Marks don’t expire
What’s already submitted stays marked. Sanitise before checking starts.
Every major lab has signed up to this.
The EU AI Act forced the industry’s hand, and the labs applied it worldwide. Almost every major lab has committed to watermarking what it produces, and most already do. Un-Claude sanitises all of them, in text, files and images.
- Marking todayMarking todayClaudeby Anthropic
- Marking todayCommitted, comingChatGPTby OpenAI
- Marking todayMarking todayGeminiby Google
- Marking todayNothing yetGrokby xAI
- Marking todayCommitted, comingMeta AIby Meta
- Marking todayDoes not produce thisFireflyby Adobe
- Marking todayDoes not produce thisStable Diffusionby Stability AI
As of 20 August 2026, from each company’s own published material.
FAQ
Can my school or employer detect Claude’s watermark today?
For files, yes. For text, not yet. Anthropic’s text detector is not open yet, and the company has committed to releasing one imminently. The day it opens, anyone can check, and text that is already marked stays marked. We will be running our own sanitised output against Anthropic’s API from the day it opens, to verify our sanitisation works on every run.
What can you actually guarantee?
Two of the three layers, completely. Hidden characters and metadata are found, counted, removed and verified: we show you exactly what was in your file, and prove it is gone. The third layer, the statistical watermark in the word choices themselves, is different. It is present in everything Claude writes and in Gemini’s text too, no tool on earth can currently detect it, and so nobody can verify its removal, including us. We target it with an engine built for that one job: it dismantles the word sequences the mark rides on, holds every fact and your original length, and hands you the measurements from each run.
Why can’t I just ask another AI to reword it?
Because a generic rewrite protects exactly the wrong thing. Ask any model to reword and it hands your text back mildly altered, and unbroken runs of your original wording are precisely where the statistical watermark lives. In our tests generic rewrites left those runs intact, came back around a third shorter, and silently dropped facts. Ours is a purpose-built engine, not a prompt: it works to a three-word limit on what carries over, routed through a model that is not Claude so the mark cannot be reapplied mid-rewrite, every number, date and name checked against your original with a retry if one drifts, and length held within a tenth. Quotations and references are protected on purpose and come back exactly as you sent them. Across our test set it breaks over 90% of three-word sequences with zero figures lost, and every run hands you the numbers for your own document.
How do I know the rewrite actually worked?
Nobody can verify the removal of a statistical watermark yet, and any tool that claims otherwise is lying to you and selling a fraudulent product. What we give you is measured: every run returns the share of your wording replaced, the longest run of your original wording still standing, and every figure checked. The mark rides only on unbroken runs of your wording, so those are the numbers that describe the work. Read them together with what we protected: quotations and references come back exactly as you sent them, so they count towards that longest run by design.
Will it change my meaning, my facts, or my numbers?
No, and this is enforced rather than promised. Every number, date and name is checked against your original and the section retries if one drifts. Length is held within a tenth, and the receipt shows the figures carried through.
Is using Un-Claude cheating?
Un-Claude removes marks that AI companies place inside your work, invisibly, without asking you and without telling you. What you then submit, publish or send is your call and your responsibility, under whatever rules you have agreed to. Where we stand on the marking itself is on the mission page.
What happens to my document after I upload it?
It is processed and deleted. Uploads are held for the length of the request and removed when the response is sent. Nothing you paste or upload is stored.
What about Word documents and PDFs?
Upload a Word document and we clean its hidden characters and metadata. For the statistical rewrite, paste the text itself into the box, and paste it back into your document when it is done. More file types are coming soon.
What if the scan finds nothing?
Then your text or file is clean of hidden characters and metadata, and that is a real answer rather than a failed scan. The statistical watermark is the exception: no tool can show or count it yet, which is why it is not found at all but targeted, by an engine built to break the word sequences it rides on.