How to remove the Claude watermark (what works, what doesn't)
By Claude Watermark Research · Updated
There are two different things people mean by "the Claude watermark", and they need opposite fixes. The copy-paste traces — hidden HTML class names, invisible characters, exotic spaces — come out completely by cleaning the text or pasting through a plain-text editor. Anthropic's statistical mark lives in the word choice itself, so nothing character-based touches it; only resampling the text through a different model does.
Method 1: paste through plain text
Copy your text, paste it into Notepad, TextEdit in plain mode, or any code editor, then copy it back out. This strips all HTML, including the `font-claude-response-body` class names that the AO3 detector and anyone inspecting your document can see.
Fast, free, and it handles the single most detectable artifact. What it misses: zero-width characters and byte-order marks sometimes survive, smart quotes and em dashes always do, and it obviously does nothing to word choice.
Method 2: clean the characters properly
A checker that lists each artifact with its position beats a blind strip, because a few of these are load-bearing — zero-width joiners are meaningful inside some emoji sequences and scripts. You want to see what is being removed.
This is what our free checker does: HTML classes and AI `data-` attributes, zero-width space, non-joiner, joiner and byte-order mark, non-breaking and exotic spaces, smart quotes, and em dashes. Checking is unlimited and runs in your browser, so nothing is uploaded.
Method 3: resample the text
A distributional watermark is carried by which words were chosen, so the only thing that disturbs it is choosing different words. Running the passage through a different model regenerates those choices. Anthropic's own documentation concedes marks may not survive heavy editing, paraphrasing or translation — that is this method, described from the other side.
One constraint worth knowing: the model you rewrite with must not itself watermark, or you have swapped one mark for another. Gemini has applied SynthID to text since 2024, and Claude models launched from 2 August 2026 mark at launch.
What does not work
Asking the model to "remove its watermark" or "write without a watermark". The bias is applied during sampling, below the level the model can introspect or control.
Retyping the text by hand. You reproduce the same word choices, so you preserve the signal exactly while wasting an afternoon.
Any tool promising to verify a passage is now unmarked. Detection needs the key Anthropic holds and has not published, so nobody outside Anthropic can confirm a mark is present or absent — which means no removal can be verified either, by anyone.
Questions
- Does changing a few words remove it?
- Light edits barely move a distributional signal, because it is spread across the whole passage rather than sitting in any one place. Heavy paraphrase is what degrades it.
- Is removing a watermark against Anthropic's terms?
- Anthropic's usage policy prohibits misrepresenting AI-generated content as human where that deceives someone. Cleaning formatting artifacts from your own writing is a different act from passing off a generated essay. Read the policy and your institution's rules.
- Do I need an account to check?
- No. Checking is free, unlimited and runs entirely in your browser. Only the rewrite costs credits, and new accounts get 10 free with no card.
Check your own text. Free, unlimited, no account, and it runs in your browser so nothing is uploaded.
Open the checker
