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.
▸Hidden-character optionscore invisibles, tags & private-use are always removed
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.
of watermarked passages detected before any edit
still detected after swapping 10% of the words
after a full paraphrase by another AI model
after Unmark’s default rewrite
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.
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.
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.
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.
What Unmark removes
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.
Hidden-character watermarks
Zero-width spaces, bidi controls, tag characters and look-alike glyphs woven between letters — revealed and stripped instantly, in your browser.
Remove watermarks by tool: ChatGPT · Gemini · Claude · Copilot · DeepSeek · Grok
Just want to look first? Check text for hidden characters · Which AIs watermark text? See the tracker
Only the em dashes? Em dash remover · What is ChatGPT’s watermark? textGrain explained
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.