Claude watermark detector: check whether a text carries Claude's watermark

What the Claude watermark detector tells you

Paste a text and the detector answers one question: does this text carry the statistical watermark that Anthropic applies to Claude output since August 2026, and how strongly. You get a probability band (unlikely, possible, likely, very likely), a short explanation of what drove the estimate, and the length caveat that every honest detector must give. It will be free at the entry tier, and it is the same measurement Unmarker shows before and after a rewrite on the Claude watermark remover.

It is not an "AI or human" verdict. A watermark detector and an AI-content detector are different instruments, and confusing them is the most common mistake in this space. The sections below explain the difference, because it changes what you should do with the result.

What Claude's watermark is, in one paragraph

Since August 2, 2026, Claude models pick among near-equivalent words according to a secret key held by Anthropic. Over a paragraph, the text ends up slightly enriched in "favoured" words. There is no hidden character, no metadata, nothing to see. Only a scoring function that knows the key, or a good estimate of its footprint, can tell. Anthropic announced it on August 11, 2026 and says the mark survives copy-paste and light editing. Our guide to the Claude watermark covers the scope, the sources and what survives.

How a statistical detection works

A detector recomputes, for every word position, whether the word that was chosen belongs to the "favoured" set for that position. In a human text, or in text from an unmarked model, favoured and unfavoured words show up at the rate chance predicts. In a marked text they show up more often, and the longer the text, the more clearly the excess stands out from noise. The result is a score, comparable to a z-score, that says how unlikely the observed excess is under the assumption of no watermark. Anthropic's own detector holds the key and can be precise. We do not hold the key, and we do not pretend to: our detector estimates the footprint of a keyed bias under a family of scoring functions calibrated on three sets, unmarked human text, text from marked Claude models, and human text lightly edited by Claude. That gives a useful band, with an explanation, and it is refreshed when Anthropic changes anything public. If Anthropic publishes a public detector, we will link to it as the authoritative source.

Why short texts stay uncertain

A statistical watermark needs choices to bias, and a detector needs those choices to count. A tweet, a subject line or a two-sentence answer gives a handful of positions, and a handful is not enough to separate a real mark from luck. Under roughly 150 words the honest answer is "not enough text", and the detector says exactly that instead of inventing a number. Anthropic makes the same point: the mark is only weakly detectable on very short passages. Between 150 and 400 words the band is wide; above that it narrows. Code, lists of figures and highly formulaic sentences carry little mark by construction, because there was little choice to bias, and they are reported as such.

False positives: proofreading and translation

The mark is added by the model to the words it produced. If you wrote a text yourself and asked Claude to proofread it, translate it or tighten it, the passages Claude touched come back marked, and a detector will report a likely watermark on your own work. Anthropic itself notes this. So a "likely" result means "a marked Claude model was involved somewhere", not "Claude wrote this" and certainly not "the ideas are not yours". Our detector shows which paragraphs drive the score, which is usually enough to see whether the whole text or one edited section carries the mark.

Claude AI detector vs Claude watermark detector: not the same test

Stylometric detectors (GPTZero, Turnitin, Originality and others) look at how a text is written: sentence-length variance, predictability of each word, favourite connectors. They answer "does this read like a language model wrote it", need no key, work on any text and have well-documented false positives on non-native writers and formulaic genres. Watermark detectors look for a specific keyed bias a vendor deliberately added. They answer "did model X touch this text". They say nothing about a text from an unmarked model, and an "unlikely" result is what human writing and unmarked-model writing both look like. Neither instrument judges whether the ideas are yours. If what you need is the first kind of answer for Claude specifically, the Claude AI detector page explains why the watermark check is the better instrument and where it stops. If your text came from ChatGPT, there is no statistical mark to look for in August 2026: the ChatGPT watermark remover covers the invisible-character case instead.

When to use it

• Before publishing a draft Claude helped with, to know whether a client's or platform's check would flag it.
• After proofreading or translating your own text with Claude, since that alone can add the mark.
• To check the output of any remover, including ours: the after score is the same instrument.
• To settle a false-positive dispute: a stylometric flag with an "unlikely" watermark score is a useful data point, not a proof.

Limits, stated plainly

The detector estimates; it does not certify. Only Anthropic, holding the key, can certify. Short texts give a shrug. Keys and schemes can change, and we track public information. If a text is scored "likely" and you have the right to publish it as your own, the Claude watermark remover rewrites it under a meaning lock and shows the after score. If the trouble is invisible characters rather than a statistical mark, the AI text cleaner is the right tool. For Gemini text, see the SynthID text detector and remover; for any vendor, the AI watermark detector runs the same family of checks.

Availability

The detector opens with the rest of Unmarker, free at the entry tier: three texts a day, up to 1,500 words each. The waitlist gets access first, and one email at launch, nothing else.

Questions people ask