Is there an AI checker for Claude?
By Claude Watermark Research · Updated
Partly. There are artifact checkers, which reliably find what Claude's interface leaves in copied text, and there are AI detectors like GPTZero and Turnitin, which guess at authorship from writing style and are wrong often enough to have ruined real careers. What does not exist is a Claude watermark detector: Anthropic holds the key to its statistical mark and has not released a way for anyone else to verify it. Any product advertising that third capability is selling one of the first two under a better name.
Check your own text for these artifacts. Runs in your browser, no account, nothing uploaded.
The three products, ranked by how much they can prove
Artifact checkers. Look for concrete residue: HTML class names from the web UI, zero-width Unicode, smart quotes, em dashes. Fully reliable, because there is no inference — the character is in the string or it is not. What they prove is narrow: the text passed through a particular interface.
AI detectors. GPTZero, Turnitin, Originality and the rest score how predictable and how uniform the writing is. They produce a percentage, and that percentage is a guess. They flag human writing regularly, and they misfire hardest on non-native English and on plain, tightly-edited prose.
Watermark detectors. Would verify Anthropic's keyed bias directly. None is available to the public. Anthropic has said it is working on enabling third-party detection; it has not shipped.
How to tell which one a product actually is
Read what it says it found. A tool reporting “4 zero-width characters, 2 class attributes” is an artifact checker and is telling the truth. A tool reporting “87% AI” is a detector making a guess. A tool reporting that it verified or removed the Anthropic watermark is describing something nobody can currently do.
The tell is the confidence. Artifact checkers give counts and positions. Detectors give percentages. Anything giving you a yes-or-no on the watermark specifically is inventing it.
What we do, stated plainly
This site is the first kind. It finds and removes artifacts, and it will show you exactly what it found and let you re-check the output afterwards. It does not detect Anthropic's watermark and does not claim to remove it.
That is a smaller promise than several competitors make, and we would rather make the smaller true one. If a public detector ever ships, the honest position changes and this page changes with it.
Questions
- Is there a Claude watermark detector?
- Not publicly. Anthropic's watermark is verified with a key that has not been released, so no third party can check for it. Anthropic has said it intends to enable third-party detection, but nothing has shipped.
- Can GPTZero or Turnitin detect Claude specifically?
- No. They classify writing style and cannot attribute text to a particular model. They return a likelihood that some AI was involved, and that likelihood is frequently wrong in both directions.
- Then what can I actually check for?
- Artifacts — the HTML class names, invisible Unicode and typographic characters that Claude's web interface leaves in copied text. Those are definitively checkable, because a checker reads them straight out of the text.
- If a tool says it removed the Claude watermark, what did it really do?
- Almost certainly it stripped invisible Unicode characters. That is worth doing, but those characters are not the watermark — Anthropic's method biases word choice, which deleting characters cannot affect.
Check your own text. Free, unlimited, no account, and it runs in your browser so nothing is uploaded.
Open the checker
