# Claude Watermark

> Free tool to check whether your Claude or ChatGPT text carries detectable artifacts, and clean them. Straight answers on Anthropic's text watermark — what it is, which models carry it, and what actually removes it.
> Independent. Not affiliated with Anthropic, OpenAI or Google.

## Key facts (reviewed 2026-09-02)

- Anthropic embeds a statistical text watermark in Claude models launched on or
  after 2026-08-02. Marked models now include Claude Fable 5.1 and Claude Mythos 5.1; see the linked model list below.
- The mark is distributional (keyed bias on token choice, SynthID-Text family).
  It is not metadata and not invisible characters. It survives copy-paste, and
  stripping characters cannot remove it.
- Anthropic's detection API is in private preview for eligible organisations.
  There is no public detector, so public tools cannot verify whether a passage
  carries the mark or certify that removal worked. We say this about our own
  product too.
- What is detectable today is copy-paste artifacts: provider-specific HTML class
  names (Claude stamps classes containing "claude"), zero-width characters,
  non-breaking and exotic spaces, and typographic punctuation.
- Detection of short passages is unreliable by construction. Every method in use
  is a statistical test, and confidence grows with the square root of length.
- The AO3 "Claude detector" skin is a single CSS rule matching paragraphs whose
  class contains "claude". Effectively no false positives, and it only fires on
  text pasted directly from Claude into AO3's rich-text editor.

## Model status

- [Claude Fable 5.1](https://claudewatermark.xyz/models/claude-fable-5-1): launched 1 September 2026. Watermarked.
- [Claude Mythos 5.1](https://claudewatermark.xyz/models/claude-mythos-5-1): launched 1 September 2026. Watermarked.
- [Claude Opus 5](https://claudewatermark.xyz/models/claude-opus-5): launched 24 July 2026. Not watermarked.
- [Claude Sonnet 5](https://claudewatermark.xyz/models/claude-sonnet-5): launched 30 June 2026. Not watermarked.
- [Claude Fable 5](https://claudewatermark.xyz/models/claude-fable-5): launched 9 June 2026. Not watermarked.
- [Claude Mythos 5](https://claudewatermark.xyz/models/claude-mythos-5): launched 9 June 2026. Not watermarked.
- [Claude Opus 4.8](https://claudewatermark.xyz/models/claude-opus-4-8): launched 2026, before 2 August. Not watermarked.
- [Claude Haiku 4.5](https://claudewatermark.xyz/models/claude-haiku-4-5): launched 1 October 2025. Not watermarked.

## Tools

- [Checker](https://claudewatermark.xyz/): paste text and see every detectable artifact with its
  position. Runs entirely in the browser. No account, nothing uploaded.
- [Pricing](https://claudewatermark.xyz/pricing)

## Guides

- [How to check text for a Claude watermark](https://claudewatermark.xyz/guides/how-to-check-for-claude-watermark): Paste it and see in one second. Check concrete Claude artifacts, and understand why the statistical watermark still has no public detector.
- [AI watermark detectors: what actually exists](https://claudewatermark.xyz/guides/ai-watermark-detector-what-exists): Anthropic's detector is in private preview, but no public Claude watermark detector exists. What public tools actually check instead.
- [Is there an AI checker for Claude?](https://claudewatermark.xyz/guides/is-there-an-ai-checker-for-claude): Three different products get sold under that name. Two are public; Anthropic's true watermark detector is restricted. Which is which.
- [Does the GitHub Claude watermark remover actually work?](https://claudewatermark.xyz/guides/does-the-github-claude-watermark-remover-work): The open-source remover has over 20,000 GitHub stars. Compare its self-managed workflow with browser cleanup, a $3 document rewrite and our Agent Skill.
- [Anthropic's watermark detection API: access and limits](https://claudewatermark.xyz/guides/anthropic-watermark-detection-api): Anthropic's watermark detector is in private preview. Who can access it, what it detects and why there is still no public checker.
- [How does Anthropic's watermark work?](https://claudewatermark.xyz/guides/how-does-anthropic-watermarking-work): The mechanism behind Claude's text watermark, in plain terms: what gets marked, what survives editing, and the limits Anthropic states itself.
- [The AO3 Claude AI detector skin, explained](https://claudewatermark.xyz/guides/ao3-claude-ai-detector-skin): What the AO3 Claude detector site skin actually checks, why it has no false positives, and how the artifact gets into a fic in the first place.
- [Does Claude watermark text? (September 2026)](https://claudewatermark.xyz/guides/does-claude-watermark-text): Yes. Claude Fable 5.1 and Claude Mythos 5.1 carry Anthropic's statistical text watermark. Here is which models are marked, how detection works and what removes it.
- [What Claude actually leaves in text you copy and paste](https://claudewatermark.xyz/guides/what-claude-leaves-in-copied-text): Hidden HTML, invisible characters, smart quotes. Exactly what survives a copy-paste out of Claude, which parts are provable, and which mean nothing.
- [Best Claude watermark checkers & removers (2026)](https://claudewatermark.xyz/guides/best-claude-watermark-checker): Hands-on ClaudeWatermark.xyz vs .com vs .org cleanup tests with downloadable inputs and outputs, plus a comparison of free checkers and paid rewrites.
- [AI watermarks vs AI detectors: not the same thing](https://claudewatermark.xyz/guides/ai-text-watermark-vs-ai-detector): They work in opposite directions and fail in opposite ways. Mixing them up is why most advice about AI detection is wrong, and why people get accused.
- [Most Claude watermark removers cannot work](https://claudewatermark.xyz/guides/why-most-claude-watermark-removers-cannot-work): Tools selling Claude watermark removal describe their own method as stripping zero-width characters. Anthropic's mark is word choice. Here is the proof.
- [How to remove Claude's HTML from a fic on AO3](https://claudewatermark.xyz/guides/how-to-remove-claude-html-from-ao3): Pasting from Claude into AO3's editor carries hidden HTML class names, which the detector skin flags. How to strip them from a work already posted.
- [The AO3 AI skin flagged my fic and I did not use AI](https://claudewatermark.xyz/guides/ao3-ai-skin-false-positive): A flag is not proof, and the skin cannot tell what people think it tells. What it actually checks, and what to do if your fic was flagged unfairly.
- [How to paste from Claude without leaving artifacts](https://claudewatermark.xyz/guides/paste-from-claude-without-artifacts): Copying out of a chat window carries hidden HTML and Unicode with it. What survives, where it gets you flagged, and the one habit that prevents it.
- [Which Claude models are watermarked? (live list)](https://claudewatermark.xyz/guides/which-claude-models-are-watermarked): Claude Fable 5.1 and Claude Mythos 5.1 are watermarked. A current model-by-model list against Anthropic's 2 August 2026 cutoff.
- [No invisible characters in Claude, ChatGPT or Gemini output](https://claudewatermark.xyz/guides/no-invisible-characters-in-ai-output): We copied real output from three chat interfaces and measured it. Zero zero-width characters, zero non-breaking spaces, zero exotic spaces.
- [Does Grok watermark text? No.](https://claudewatermark.xyz/guides/does-grok-watermark-text): Grok does not watermark text. xAI never signed the code. So every "Grok watermark remover" is selling a fix for something that was never there.
- [Does Gemini watermark text? Yes — and it has for longer](https://claudewatermark.xyz/guides/does-gemini-watermark-text): Google's SynthID-Text has watermarked Gemini's text output since 2024. How it compares to Anthropic's scheme, and why it matters when choosing a model.
- [Why the em dash became the AI writing tell](https://claudewatermark.xyz/guides/em-dash-ai-writing-tell): The em dash is the most-cited AI giveaway. Why models overuse it, why it is weak evidence on its own, and how to fix it without flattening your prose.
- [Will Turnitin detect text written with Claude?](https://claudewatermark.xyz/guides/will-turnitin-detect-claude): It can, but not by finding Anthropic's watermark. Turnitin has no access to it. It guesses from how predictable your wording is, and it flags humans too.
- [How to tell if text was written by AI (what actually works)](https://claudewatermark.xyz/guides/how-to-tell-if-text-was-written-by-ai): A practical ranking of the signals people use to identify AI writing, from the ones that are real evidence to the ones that are guesswork.
- [Zero-width characters in AI text, and how to find them](https://claudewatermark.xyz/guides/zero-width-characters-in-ai-text): Invisible Unicode characters turn up in copied AI output. What they are, why they are not a watermark, and how to see and remove them.
- [Claude watermark remover: what actually works](https://claudewatermark.xyz/guides/remove-claude-watermark): A practical Claude watermark remover guide: what strips copy-paste artifacts, what changes the statistical mark, and which popular fixes do nothing.
- [An AI detector flagged your writing. Here is what to do](https://claudewatermark.xyz/guides/ai-detector-false-positive-what-to-do): Detector scores are guesses, not evidence. What to say when you are accused, and what actually proves you wrote it. Written for people being punished.
- [Is Claude-generated code watermarked?](https://claudewatermark.xyz/guides/is-claude-generated-code-watermarked): Anthropic's watermark applies to text output, including code. But code is a poor host for a distributional mark, and here is the technical reason why.
- [Does ChatGPT watermark text? (August 2026)](https://claudewatermark.xyz/guides/does-chatgpt-watermark-text): OpenAI has shipped SynthID watermarking for ChatGPT images and voice, but not for text. Here is what is marked, what is not, and why text is the hard case.
- [Hidden text in assignment prompts: the trap that actually works](https://claudewatermark.xyz/guides/hidden-text-in-assignment-prompts): Some instructors hide invisible instructions inside assignment files. Unlike AI detectors this produces real evidence. How it works and how to see it.
- [Can my employer tell you used Claude at work?](https://claudewatermark.xyz/guides/can-my-employer-tell-i-used-claude): Not through a generally available watermark detector. Anthropic limits detection to an eligible private preview, while a managed work device exposes far more direct evidence. What is visible.
- [How to make AI text sound human (what actually changes it)](https://claudewatermark.xyz/guides/how-to-make-ai-text-sound-human): The exact patterns that make writing read as machine-written, and the edits that remove them. Concrete changes you can make, not vague advice.
- [Does Google penalise AI content?](https://claudewatermark.xyz/guides/does-google-penalize-ai-content): No. Google's published position is that how content was made is not the issue, scaled low-value content is. What the policies actually say.
- [Can recruiters tell you used AI on your application?](https://claudewatermark.xyz/guides/can-recruiters-tell-you-used-ai): Not from a watermark, nobody can read one. But a CV carries metadata, pasted markup and invisible characters anyone can check, and one trick backfires.
- [Moby Dick: 26 em dashes per 1,000 words. Austen: none.](https://claudewatermark.xyz/guides/moby-dick-87-em-dashes): Melville used 87 em dashes in 3,400 words, in 1851. We ran ten pre-computer novels through an AI artifact checker. Here are all the numbers.
- [Why short text cannot be reliably detected as AI](https://claudewatermark.xyz/guides/why-short-text-cannot-be-detected): Watermarks and AI detectors are both statistical tests, and statistics need sample size. Why a paragraph carries almost no signal either way.
- [Is Claude's watermark invisible characters?](https://claudewatermark.xyz/guides/is-the-claude-watermark-invisible-characters): No. Headlines calling it invisible mean imperceptible, not hidden characters. The two get confused, and the difference decides what you can check.
- [If Claude only edited my writing, is it watermarked?](https://claudewatermark.xyz/guides/does-claude-watermark-writing-it-edits): Proofreading is not generation. What a word-choice watermark can and cannot reach when you hand Claude your own prose and ask it to fix it.

## Cleaning by platform

- [AO3](https://claudewatermark.xyz/clean/ao3): keeps pasted HTML. Edit the work, switch the editor to HTML mode, and delete any class attribute containing “claude”.
- [Google Docs](https://claudewatermark.xyz/clean/google-docs): keeps pasted HTML. Paste with Ctrl+Shift+V (Cmd+Shift+V on macOS) to paste without formatting.
- [WordPress](https://claudewatermark.xyz/clean/wordpress): keeps pasted HTML. Switch the block to Edit as HTML, or use the Code editor view, and remove the class attributes.
- [Notion](https://claudewatermark.xyz/clean/notion): discards pasted HTML. Paste as plain text with Ctrl+Shift+V (Cmd+Shift+V), or paste normally and then check for invisible characters.
- [Substack](https://claudewatermark.xyz/clean/substack): keeps pasted HTML. Paste into a plain-text editor first, then into Substack.
- [Medium](https://claudewatermark.xyz/clean/medium): discards pasted HTML. Paste normally, then check for invisible characters and smart punctuation.
- [Microsoft Word](https://claudewatermark.xyz/clean/word): keeps pasted HTML. Use Paste Special → Unformatted Text, or Ctrl+Alt+V and choose Unformatted.
- [Canvas](https://claudewatermark.xyz/clean/canvas): keeps pasted HTML. Use the HTML editor toggle in the Rich Content Editor and remove the class attributes, or submit a file instead.
- [Gmail](https://claudewatermark.xyz/clean/gmail): keeps pasted HTML. Paste with Ctrl+Shift+V (Cmd+Shift+V) to paste without formatting.
- [LinkedIn](https://claudewatermark.xyz/clean/linkedin): discards pasted HTML. Paste normally — but strip invisible characters first, and reconsider the em dashes.
- [Ghost](https://claudewatermark.xyz/clean/ghost): keeps pasted HTML. Use the HTML card, or paste through a plain-text editor before pasting into the post.
- [Obsidian](https://claudewatermark.xyz/clean/obsidian): discards pasted HTML. Paste as plain text with Ctrl+Shift+V (Cmd+Shift+V), or clean the invisible characters afterwards.
- [Outlook](https://claudewatermark.xyz/clean/outlook): keeps pasted HTML. Paste with Ctrl+Shift+V, or use Home → Paste → Keep Text Only.
- [Slack](https://claudewatermark.xyz/clean/slack): discards pasted HTML. Paste normally — then check for invisible characters before sending.
- [Discord](https://claudewatermark.xyz/clean/discord): discards pasted HTML. Paste normally. Nothing structural survives, but strip invisible characters anyway.
- [Confluence](https://claudewatermark.xyz/clean/confluence): keeps pasted HTML. Use the source editor, or paste with Ctrl+Shift+V to strip formatting.
- [Overleaf](https://claudewatermark.xyz/clean/overleaf): discards pasted HTML. Paste normally, then replace smart quotes and em dashes — LaTeX wants ``quotes'' and ---.
- [Blackboard](https://claudewatermark.xyz/clean/blackboard): keeps pasted HTML. Open the HTML view in the content editor and remove the class attributes, or attach a file instead of pasting.
- [Moodle](https://claudewatermark.xyz/clean/moodle): keeps pasted HTML. Switch the editor to HTML source view and delete the class attributes, or paste through a plain-text editor first.
- [Microsoft Teams](https://claudewatermark.xyz/clean/teams): keeps pasted HTML. Paste with Ctrl+Shift+V to paste without formatting.
- [Google Slides](https://claudewatermark.xyz/clean/google-slides): keeps pasted HTML. Paste with Ctrl+Shift+V (Cmd+Shift+V) to paste without formatting.