Remove the AI watermark from any text

Google’s SynthID, ChatGPT’s new textGrain and Claude’s watermark hide a detectable pattern in AI text. Unmark rewrites it away while keeping your meaning, strips invisible characters, and flags the formulaic AI phrases worth editing out.

SynthID, textGrain & ClaudeInvisible characters60 free rewrites/hr · no sign-upFlags AI clichés
Paste AI text0 chars
Watermark-free text0 chars
Your watermark-free text appears here.
Text from
Pick where it came from to see what it needs.
Strength
Humanise — natural rewrite, same length & meaning.
Paste AI text, then remove its watermark · 60 free rewrites an hour
▸Hidden-character optionscore invisibles, tags & private-use are always removed
Original research

Does it work? We tested it

We watermarked 120 passages with Google’s open-source SynthID-Text, edited each one six ways, and checked which were still detected. Light edits leave the watermark in place; a full rewrite takes it out.

100%

of watermarked passages detected before any edit

92%

still detected after swapping 10% of the words

8%

after a full paraphrase by another AI model

3%

after Unmark’s default rewrite

Three steps

How to remove an AI watermark

No account, no setup — paste, remove, copy. Works with text from ChatGPT, Claude, Gemini and any other AI. New here? Read the full guide to removing AI watermarks.

01

Paste your AI text

Drop in text from ChatGPT, Claude, Gemini, Copilot or anywhere else. Unmark scans it for hidden characters instantly, in your browser.

02

Remove the watermark

Pick a strength and hit Remove watermark. Unmark rewrites the wording to break statistical watermarks like SynthID and textGrain, and strips every invisible character in the same step.

03

Review it, then copy

See each rewritten paragraph beside its original, put one back or have it redone, then copy or download text that reads the same to a human, minus the machine fingerprint.

Two watermarks, one tool

What Unmark removes

Token-level fingerprint

Statistical watermarks

Google’s SynthID, OpenAI’s textGrain and Claude’s watermark all hide in the model’s word choices and survive light editing. Unmark rewrites your text to scramble the pattern while the meaning stays intact.

Invisible fingerprint

Hidden-character watermarks

Zero-width spaces, bidi controls, tag characters and look-alike glyphs woven between letters — revealed and stripped instantly, in your browser.

Background

Understanding AI text watermarks

As AI writing tools became mainstream, providers began embedding watermarks in their output so it can later be identified as machine-generated. These marks come in two very different forms, and removing one does nothing about the other.

Statistical watermarks (SynthID, textGrain, Claude)

A statistical watermark such as Google’s SynthID-Text subtly biases which words and tokens the model picks, creating a pattern that a matching detector can measure across a passage. Because it lives in the word choices themselves — not in any visible mark — trimming a sentence or swapping a few words rarely clears it. The dependable way to remove it is to rewrite the text enough that the token sequence no longer matches, which is exactly what Unmark does while preserving your meaning.

This went mainstream in 2026, as the EU AI Act began requiring machine-detectable marks on AI-generated text. Anthropic now watermarks text from Claude models launched since August 2, worldwide, and OpenAI announced textGrain on October 5 for ChatGPT and Codex text in the EU. Both say light edits weaken the mark without reliably removing it: in OpenAI’s tests, swapping 25% of the words cut detection to 17%, and Anthropic says a complete rewrite removes its watermark. None of the three offers a public text detector, so treat any removal — including Unmark’s — as best-effort.

Invisible-character watermarks

The second kind hides invisible Unicode between the letters: zero-width spaces and joiners, byte-order marks, bidirectional controls, tag characters and look-alike glyphs. They render as nothing on screen but survive copy-paste, so they can fingerprint or track text. Google’s, OpenAI’s and Anthropic’s watermarks don’t use them, but Microsoft’s Windows AI text APIs do: since October 2026 they insert the sequence U+200B U+034F U+2060 U+200B into generated text. Others come from formatting, copy-paste pipelines and deliberate tracking. Unmark detects and strips them in your browser — while carefully preserving load-bearing marks such as emoji sequences — so the readable text is untouched but the hidden payload is gone.

Questions

Frequently asked questions

What is an AI text watermark?+
It's a hidden signal that flags text as AI-generated. There are two families: statistical watermarks (Google's SynthID, OpenAI's textGrain, Claude's watermark) that bias the model's word choices into a pattern a detector can measure, and hidden-character marks that weave invisible Unicode between the letters. Unmark removes both.
Does ChatGPT watermark its text?+
Yes, for users in the EU. On October 5, 2026 OpenAI announced textGrain, a statistical watermark rolling out to eligible ChatGPT and Codex text in the EU; elsewhere it is off by default, though API developers can opt in. OpenAI's own tests show detection falling to 17% once a quarter of the words are swapped, and a full rewrite changes far more than that.
What is SynthID, and can Unmark remove it?+
SynthID-Text is Google's watermark for Gemini text. It nudges the model's word choices into a statistical pattern that survives light editing. Google notes that detection confidence drops sharply when text is thoroughly rewritten, which is what Unmark does while keeping your meaning.
Does Claude watermark its text, and can Unmark remove it?+
Yes. Anthropic watermarks text from Claude models launched since August 2, 2026, worldwide. Anthropic says light editing probably won't remove it completely, but a complete rewrite that replaces every word will. Unmark's rewrite changes most of the wording while keeping your meaning; treat it as best-effort, since only eligible organisations can run Anthropic's detector.
How do I remove an AI watermark from text?+
Paste the text, choose a rewrite strength, and click Remove watermark. Unmark rewrites the wording to disrupt the statistical/SynthID watermark and strips invisible characters at the same time, then you copy the clean result. It's free and runs in your browser.
Do I always need a rewrite?+
No. A rewrite is for watermarks that live in the wording: Gemini, Claude, and ChatGPT text written in the EU. Text from an assistant that marks nothing in the wording only needs its hidden characters stripped, which leaves your wording exactly as it is. Pick where the text came from under the tool and it says which applies, with the date that answer was last checked.
How do I know the watermark is gone?+
You can't know for certain, because no provider offers a public text detector. What Unmark shows after each rewrite is how much of the original wording survives, as the share of its five-word runs that are still there, next to what we measured: in our test of open-source SynthID-Text on 120 passages, rewrites that kept under 10% of those runs were detected 6 times in 405, and edits that kept more than half were detected 110 times in 120.
Can Unmark make AI text undetectable?+
Rewriting substantially degrades statistical watermarks and AI-detector signals, and stronger settings degrade them more. But no tool can promise a text is permanently undetectable against every future detector — treat Unmark as best-effort de-watermarking, and always re-read the result.
Does it remove invisible or zero-width characters?+
Yes. Unmark scans in your browser for zero-width spaces, joiners, byte-order marks, bidirectional controls, the Unicode tag block, exotic spaces and look-alike glyphs, and strips them — while preserving load-bearing marks like emoji sequences. This runs instantly and is applied to every rewrite.
Does Unmark remove Microsoft's Windows AI text watermark?+
Yes, completely. Since October 2026, text from Microsoft's Windows AI text APIs can contain the invisible sequence U+200B U+034F U+2060 U+200B. Because it is made of characters rather than word choice, Unmark finds every copy and strips it — no rewrite needed.
Which AI tools does it work with?+
Any of them. Unmark works on plain text, so it removes watermarks from ChatGPT, Claude, Gemini, Copilot, Llama and any other AI output — just paste the text in.
Is Unmark free?+
Yes. You get 60 free rewrites an hour with no account or sign-up, and the hidden-character scan runs in your browser with no limit at all.
Is my text private?+
The hidden-character scan runs 100% locally in your browser — nothing leaves your device. For a rewrite, the text is sent to our AI model provider to generate the result and sent straight back; Unmark itself doesn't store or log it. The site counts page views and button clicks to see which pages get used. Those counts hold no text, no cookie and nothing that identifies you.
Will rewriting change my meaning or formatting?+
Your layout stays: Unmark rewrites each paragraph and list item on its own, so headings, bullets, numbering and paragraph breaks come back where they were, and headings, code blocks and tables are kept exactly as written. The rewrite keeps the original language and roughly the same length, and is instructed to preserve facts, numbers and names. A paragraph that comes back without one of its names or numbers is rewritten again automatically, and a check after every rewrite flags anything still missing. Review changes then shows each paragraph beside its original, where you can put one back as you wrote it or have it rewritten again. Heavier strengths reword more, so re-read the result to confirm it kept your intent.
Does removing the invisible characters change how the text reads?+
No. Every hidden character Unmark strips is invisible or a look-alike of a standard character, so the visible, readable text is identical — just without the hidden payload.
Does Unmark detect AI-written text?+
No — and be wary of any tool that claims to. Unmark flags formulaic 'AI-tell' phrases (like 'delve into', 'leverage', or 'in today's fast-paced world') so you can edit them out, and shows how many remain after a rewrite. But statistical style analysis cannot reliably tell clean AI text from human writing, so a low count doesn't mean text is human-written.