AI Watermark Detector & Remover
Paste AI text to check it for statistical-watermark patterns and hidden characters, then remove them. The only thing that actually disrupts a token-level watermark like Claude's or Google SynthID is rewriting — so this tool rewrites, on top of cleaning hidden characters.
Scanning happens entirely in your browser. Only the optional 'Rewrite' step sends text — to Plagly's humanizer — and only when you click it.
Paste some text above. We'll check it for statistical-watermark patterns and hidden characters, and let you remove them.
How it works
Check the signal
We measure the flat, low-entropy cadence that watermarking amplifies and flag hidden characters — all in your browser.
Clean the characters
Strip zero-width, tag and look-alike characters. This alone can break the exact token sequence a watermark hashes.
Rewrite to remove it
Reword the text through the humanizer — the documented, effective way to disrupt a statistical watermark.
Two different things people call "AI watermarks"
Hidden characters
Invisible Unicode — zero-width spaces, tag characters — injected into text. Deterministic and provable: we show you exactly what was there and remove it. This is a bonus; it is not the main watermark.
Statistical watermarks (Claude, Google SynthID)
The real thing: a signal woven into which words the model chooses (Anthropic began watermarking Claude text for models launched Aug 2, 2026). It's invisible to a character scanner, so cleaning won't touch it — the only way to disrupt it is to reword the text.
Open the full AI HumanizerWhat this tool can and can't do
What it does
- +Flags statistical-watermark patterns (flat, low-entropy cadence) and every hidden Unicode character, with the exact code points.
- +Removes hidden characters deterministically and folds look-alike letters, which perturbs the token sequence a watermark depends on.
- +Rewrites the text through Plagly's humanizer — the method research reports removes 90–100% of token-level watermarks like SynthID.
What it can't do
- –It cannot confirm a specific vendor's watermark (Claude, SynthID). Those are verified only with the vendor's secret key, which is not public.
- –It does not lower your AI-detection score by itself — detectors read writing style, not marks. Rewriting changes style; character cleaning does not.
- –It cannot guarantee complete watermark removal — heavier rewriting removes more, at the cost of changing your wording.
- –It is not a plagiarism or authorship guarantee.
Frequently asked questions
How does Claude's or Google's text watermark work?
It's a statistical, token-level watermark: during generation the model biases its word choices toward a secret 'green list' (the Kirchenbauer / SynthID approach). The pattern is invisible to read but statistically detectable by whoever holds the key. Anthropic began adding this to Claude text for models launched on or after August 2, 2026.
Can you detect the watermark for sure?
No — and no public tool can. Verifying a proprietary watermark needs the vendor's secret key and detector; Anthropic's is still 'forthcoming' and open-source toolkits (MarkLLM and others) only detect watermarks you created with a known key. We show an honest statistical indicator instead, and our AI detector scores the writing itself.
Does removing hidden characters remove the watermark?
Partly. A statistical watermark hashes exact token sequences, so changing characters (zero-width removal, homoglyph folding) can degrade it. But the reliable method is rewording — Anthropic itself says the mark 'may not survive heavy editing, paraphrasing, translation, or mixing.'
Does this make my text pass AI detectors?
Not on its own. Character cleaning does not change an AI-detection score, because detectors analyze style, not marks. Rewriting changes the style and lowers the statistical signal, but nothing can promise permanent undetectability.
Is my text sent anywhere?
Scanning and character cleaning run entirely in your browser. Only the optional 'Rewrite' step sends text — to Plagly's humanizer — and only when you click the button.
Want to know if text reads as AI-written?
The watermark signal is only an indicator — our AI detector scores the writing itself, sentence by sentence.
Try the free AI detector