How to check text for a Claude watermark
By Claude Watermark Research · Updated
You can check text for everything Claude leaves behind except the watermark itself. Copied Claude output carries HTML class names, invisible Unicode characters, smart quotes and em dashes, and all of those are verifiable by reading the bytes — paste the passage into a checker and they are either present or they are not. Anthropic's statistical watermark is different: it is a keyed bias in word choice, the key is not public, and there is no released detector, so no tool can honestly tell you whether it is there.
Check your own text for these artifacts. Runs in your browser, no account, nothing uploaded.
The two different things people mean by “check”
Artifacts. Concrete things sitting in the text: `class="font-claude-response-body"` and similar attributes from the web UI, zero-width Unicode, non-breaking spaces, curly quotes, em dashes. These are facts about the file. A checker finds them by looking, and the answer is not a probability.
The watermark. A statistical bias woven into which words the model picked. Verifying it requires Anthropic's key. Anthropic has said it is working toward letting third parties detect it; until that ships, nobody outside the company can, and that includes us.
Nearly every dispute about “Claude watermark checkers” comes from collapsing these two into one word. The first is trivially checkable. The second is not checkable by anyone you can currently buy a subscription from.
What a check will and will not tell you
It will tell you: this passage contains three zero-width characters, four em dashes, curly apostrophes, and markup naming Claude's web interface. That is a complete, literal answer.
It will not tell you: whether a human wrote this. Artifacts prove the text passed through a particular interface, not who composed it — pasting a human-written paragraph through Claude's editor picks up the same markup, and typing an em dash on a Mac takes one keystroke.
It will also not tell you what a detector like Turnitin or GPTZero will conclude, because those judge writing style, not bytes. Clean text can still be flagged, and that is a separate problem.
Why the invisible characters get the most attention
Zero-width joiners and non-breaking spaces render as nothing, survive most copy-paste, and show up in a hex view — which makes for a convincing screenshot. That is why a wave of posts described them as “the Claude watermark”.
They are not it. Anthropic's published description is a distributional method operating on token selection, which is a different mechanism entirely, and one that stripping invisible characters cannot touch. Removing them is still worth doing, because they are genuinely a tell. Just not that tell.
Questions
- How do I check if text has a Claude watermark?
- You can check for Claude's artifacts — HTML class names, invisible Unicode, smart quotes, em dashes — by pasting the passage into a checker, and those results are definitive because they are read directly from the text. You cannot check for Anthropic's statistical watermark, because the detection key has not been made public and no detector exists outside Anthropic.
- Is checking free?
- Yes, and it needs no account. The check runs in your browser, so the text never leaves your machine. Only the optional rewrite step costs credits, because that one runs a model.
- Does a clean result mean I will not be flagged as AI?
- No. Artifact checking and AI detection are unrelated. Turnitin and GPTZero judge writing style and can flag text with no artifacts at all — including text written entirely by a person.
- Can I check text that did not come from Claude?
- Yes. Most of the artifacts are not Claude-specific — ChatGPT, Gemini and ordinary word processors all leave smart quotes, em dashes and stray Unicode. Only the class-name markup is particular to Claude's web interface.
Check your own text. Free, unlimited, no account, and it runs in your browser so nothing is uploaded.
Open the checker
