Claude Watermark Detector: What Exists Today

This page is an inventory. It catalogs what the tools calling themselves Claude watermark detectors are actually checking, and sets out what changes on the day Anthropic's own detector opens.

The Anthropic API is coming, with no date

Anthropic's commitment, from the August 14, 2026 blog post, is one sentence long:

"We will soon be offering a watermark detection API."

The Decoder reported the same commitment on the day of the post, framed as an API that will let third parties detect Claude's AI texts. BetaNews repeated it on August 22, describing a plan to release a free detection tool. Neither report carried a date, because Anthropic has not given one.

There is a second obstacle behind the missing date. Anthropic has not published the exact scheme. Andrea Saez put the consequence plainly on August 11: without the published scheme there is no way of currently building a viable detector. A statistical watermark is only readable by a party holding the key, and the key is the entire security model. Anthropic cannot publish it without handing everyone the ability to forge the mark as well as read it.

So the detector is not late in the sense of a delayed feature. It is a service that has to exist on Anthropic's side, because the check cannot be done anywhere else.

What every current "detector" actually checks

The tools on the web today fall into three groups. Only one of them is doing what its name says.

Hidden-Unicode checkers wearing a Claude label

These scan a paste for zero-width characters, narrow no-break spaces, byte-order marks and similar non-printing code points, then report a verdict about "the Claude watermark."

They do find things, because pasted text often does carry stray characters from clipboards and rich-text sources. What they find has nothing to do with Anthropic's mark, which adds no characters at all. The correct reading of a green result from one of these tools is: no hidden characters were present. The statistical watermark is Unknown, and this check never looked at it.

Stylistic classifiers wearing a watermark label

These run an "AI likelihood" model over the text and return a score, sometimes with a percentage, sometimes with a confidence band. This is the GPTZero and Originality.ai shape of product, relabeled.

A classifier is not a watermark check. It has no key, it reads no embedded signal, and it can be wrong in both directions on text that is definitively marked or definitively unmarked. The score it returns is a guess about writing style. Presenting it as a watermark verdict is the single most misleading pattern in this category, because the number reads as evidence to whoever receives it.

Honest checkers that say what they check

A smaller group scans for hidden characters and clipboard fragments, states plainly that it cannot detect Anthropic's statistical watermark because nobody outside Anthropic can, and refuses to make a removal claim it cannot verify.

That is the correct shape, and it is the shape DetectMark uses. It also happens to be the version journalists cite, because it is the only one whose output survives a follow-up question.

What this site reports today

Every check on a text paste resolves to one of four labels, and the label always carries its reason:

• Clean: the check ran and found nothing. Applies to hidden characters, clipboard fragments and lookalikes.
• Found: the check ran and found something, listed by code point and position.
• Unknown: the check cannot run. Claude's statistical watermark is always Unknown, because there is no public detector to call.
• Unsupported: the check does not apply to this input, such as an image check on a text paste.

The Claude row never reads Clean. A Clean verdict would be a claim about a signal that nothing examined, and that is exactly the failure mode this page exists to describe.

What the detector will do the day Anthropic opens the API

The plan is specific, so it can be held to:

1. One check per model. Each row of the status matrix moves from Unknown to Found or Clean, dated per verification.
2. The verdict comes from Anthropic's key on Anthropic's service. No stylistic classifier is layered on top to pad the answer, and no percentage is invented to make it look precise.
3. The check runs on pasted text, behind an explicit consent step that states what is being sent and where it goes. It never fires automatically.
4. A minimum length threshold is shown before you submit. A statistical mark is measured across many word choices, so short samples resolve to Unknown rather than to a weak guess.
5. Every Claude page on this site updates the same day: this article, the detector page, the hub and the status matrix.
6. The tool's responses to a set of standard test inputs get published, so its behaviour can be checked by someone who does not trust the description.

What to do in the meantime

Check the things that can be checked. Hidden characters, clipboard HTML fragments and lookalike letters are real, they are common in pasted AI output, and they are the actual cause of most of the odd artifacts people notice in documents and CMS fields.

Then treat any "Claude watermark detected" verdict from anywhere on the web as a claim about something else until Anthropic's API exists. The claim may be true about hidden characters. It cannot be true about the watermark.