Claude Watermark Checker & Remover

Why Everyone Suddenly Cares About Claude Watermarks

In August 2026, Anthropic announced that Claude models launched on or after August 2, 2026 embed a statistical watermark in the text they generate. The news spread fast, and with it came a wave of confusion. People started pasting their writing into random detectors, worrying that an invisible stamp would follow every sentence they had ever touched with AI. Most of what circulates about this topic is wrong, so let's sort out what is actually in your text and what isn't.

There are two completely different things people call a watermark. The first is the statistical one: Claude subtly biases which words it picks, following a pattern generated from a cryptographic key that only Anthropic holds. The second is much more mundane: concrete characters and markup that ride along when you copy text out of a chat window. This checker deals with the second kind, because that's the kind that can actually be found and removed. It is honest about the first kind: nobody outside Anthropic can verify it, and any tool that claims otherwise is guessing.

What Actually Hides in Copied Text

When you copy text from a web app and paste it somewhere else, you often get more than the words you see. The checker scans four layers:

Invisible Characters

Zero-width spaces, zero-width joiners, word joiners, byte order marks, soft hyphens, directional marks. These render as nothing at all, but they sit in the file as real bytes. A reader will never spot them. A script, a plagiarism scanner, or a diff tool will. They're the strongest copy-paste tell there is, because almost nobody types them by hand. The checker names each one with its Unicode code point, counts it, and shows where it first appears.

Special Spaces

A non-breaking space looks exactly like a normal space. So do thin spaces, hair spaces, narrow no-break spaces, and about a dozen relatives. Chat interfaces produce them, but so does Microsoft Word, so treat these as a hint rather than a verdict. They still cause real problems: broken search, weird line wrapping, failed string comparisons in code.

Typography

Curly quotes, em dashes, and the single-character ellipsis. Claude produces these constantly, and em dashes in particular have become the tell people flag on sight. Here's the honest part: your phone's autocorrect produces the exact same characters, and so does every word processor with smart quotes on. The checker counts them and lets you decide. If you want them gone, the cleaner can straighten quotes and swap dashes for plain ones.

Claude Interface Markup

For Google's assistant, the Gemini watermark checker covers SynthID facts and Gemini copy artifacts.

Working with text from other assistants too? The ChatGPT watermark checker looks for the traces ChatGPT copies leave behind.

Copy from claude.ai into a rich text editor and the paste can carry HTML class names straight from the interface, including class attributes with claude in the name and font classes like font-claude-message. Paste that into a WordPress post or an email and the markup travels with it. The checker inspects your clipboard's rich text version at paste time and warns you when interface markup came along for the ride, even though the plain text box itself discards it.

Quick Reference: Characters Worth Knowing

You don't need to memorize Unicode to use the checker, but knowing the usual suspects helps you read the results. These show up most often in text copied out of Claude and other chat tools:

The full catalog the checker scans is larger: directional controls, variation selectors, Hangul fillers, tag characters, and more than a dozen exotic spaces. Every finding is reported with its exact code point, so nothing hides behind a vague label.

The Statistical Watermark, Honestly

Anthropic's watermark changes the source of randomness Claude uses when several words would work equally well. Over enough text, that bias forms a pattern that can be verified with the right key. Three facts matter here. First, only Anthropic holds that key, and until their planned detection API ships, nobody else can check for the watermark at all. Second, light editing leaves the watermark mostly intact, while a full rewrite where every word changes removes it. Third, it lives in word choice, not in characters, so no character-level cleanup can touch it, including this one.

That last point cuts both ways. Cleaning invisible characters won't strip the statistical watermark, and a tool claiming to detect the statistical watermark without Anthropic's key is selling you a coin flip. We'd rather tell you plainly what this tool can and cannot do.

Claude vs ChatGPT vs Gemini: Who Actually Watermarks?

The three big assistants took three different paths, and knowing which is which saves you from tools that promise the impossible:

Notice the pattern: whatever the statistical situation, the layer you can actually inspect yourself is the same everywhere. Characters and markup. That's the layer this page checks and removes.

How the Checker Works

Paste your text and hit Check. The engine scans every character against a catalog of invisible characters, exotic spaces, and typographic marks, plus regex patterns for Claude interface markup. You get a verdict card, per-character counts with first positions, character, word, and byte statistics, and a preview where every hidden character is rendered as a visible colored badge showing its code point. Text that comes back clean gets a green result, which is just as useful when you need to prove your writing carries nothing.

The whole thing is client-side JavaScript. Your text is never uploaded, never logged, never stored. Check something confidential without thinking twice.

How to Check Text for Claude Watermark Artifacts

1. Copy the text you want to inspect and paste it into the box above. The check runs automatically on paste, or press Check Text if you typed or edited it.
2. No text of your own handy? Hit Try Sample and the tool loads a demonstration paragraph seeded with real artifacts so you can watch every layer light up.
3. Read the verdict card first. Red means invisible characters or Claude markup, amber means spaces or typography only, green means the text is byte-clean.
4. Scan the findings by layer. Each entry names the exact character, its code point, how many times it appears, and where it first sits. The preview below renders every hidden character as a visible badge so you can see it in context.

How to Remove Claude Watermark Artifacts

Three switches control the cleanup. Removing invisible characters is on by default, and it's careful around emoji: zero-width joiners that hold a family emoji together stay put, while the ones lurking between ordinary letters get removed. Converting special spaces to regular spaces is also on by default. Straightening quotes and replacing dashes is off by default because it visibly changes your text; flip it on when you want curly quotes turned into straight ones, em dashes into plain hyphens with spacing, and the ellipsis character into three periods. The cleaned text lands in its own box with a count of what changed and a copy button.

Who Reaches for This

• Writers accused unfairly. An editor flags your article because a detector didn't like your em dashes. Run it through, show the green result for hidden characters, and make your case with evidence.
• Fanfic and forum posters. Communities like AO3 have strict AI policies, and invisible characters in a paste can start arguments nobody enjoys. Check before you post.
• Developers. A zero-width space inside a config value or a code snippet produces errors that take hours to find by eye. Paste the offending line here and the byte is exposed in seconds.
• Editors and SEO teams. Content arrives from a dozen freelancers and half a dozen tools. Normalizing spaces and stripping stray markup before publishing saves cleanup later.
• Anyone pasting into spreadsheets or databases. Two strings that look identical but compare as different usually means an invisible character. This finds it.

Google Docs, Word, WordPress, and Email: Where Artifacts Bite

Google Docs

Docs accepts rich text pastes wholesale, which means Claude markup and every special character land in your document intact. Paste with Ctrl+Shift+V (Cmd+Shift+V on Mac) to force plain text, or better, clean the text here first so the special characters go too. Plain-text paste strips markup but keeps every invisible character.

Microsoft Word

Word both receives and creates these characters. It keeps whatever your paste carries, then adds its own non-breaking spaces and curly quotes as you edit. If a downstream system chokes on your Word document, run the exported text through the checker to see which characters are actually in it.

WordPress and Other CMS Editors

The block editor preserves pasted HTML, class names included. A claude-response class sitting in your published markup is visible to anyone who hits view-source, and it survives theme changes. Clean before you paste, and use the clipboard warning this tool shows you as your early alarm.

Email Clients

Gmail, Outlook, and Apple Mail all compose in HTML, so a rich paste keeps its markup and the recipient's client happily renders and forwards it. If the email matters, paste the cleaned version.

A Word on False Positives

Finding a non-breaking space does not mean Claude wrote your text. Word inserts them between numbers and units. French typography requires them before question marks. Emoji legitimately use variation selectors and joiners. That's why the checker separates its findings into layers and why the verdict language distinguishes hard evidence from hints. Read the categories, not just the count.