By the Deslop desk · Published 16 Aug 2026
On 11 August 2026, Anthropic confirmed something that will land on a lot of desks this school year. Claude models launched from 2 August onward weave an invisible watermark into the text they generate. Worldwide, not just in Europe. If you use Claude, your output now carries a mark you cannot see.
It is not invisible ink between the letters. When Claude writes, it constantly chooses between near-equal words: start or begin, big or large. The watermark rigs those choices into a secret pattern, spread across the whole passage. Anthropic has said the mechanism is a version of Google DeepMind's SynthID-Text. The words stay ordinary. The fingerprint lives in the statistics of how they were picked.
That design has one huge consequence: reading the mark requires the key. Without Anthropic's own tools, the pattern is indistinguishable from normal writing. Any website claiming today that it detects the Claude watermark is guessing or lying. Anthropic has committed to publishing detection tools and an API. When those ship, checking becomes possible. Not before.
Picture two students. One asked Claude to write her essay from scratch. The other wrote every word herself, then pasted it into Claude and asked for a grammar check. Both essays can now carry the mark. The watermark cannot tell those two students apart, and Anthropic says so.
This matters because the last technology that promised to catch AI writing, the detector industry, produced years of false accusations. Students were failed on the word of tools with double-digit error rates, and in January 2026 a New York court ordered a university to expunge a wrongly accused freshman's record. Watermarks are better technology than detectors. They will still be misread the same way if institutions treat "marked" as "cheated."
If you write with Claude's help and your work may be scrutinized, keep your drafts. Version history in a document editor remains the strongest proof of authorship there is. If you assess other people's writing, do not treat any single signal, mark or score, as a verdict. And if a website offers to detect or remove AI watermarks for money today, close the tab. The removal tools are laundering, and the detection tools do not have the key.
What you can honestly measure today is the writing itself. Our free editor counts AI-associated tells per 1,000 words against published benchmarks, measured with the same open lexicon: AI marketing copy scores about 222, Lincoln 3.7, Darwin 0. It also x-rays for hidden characters that ride along in copy-pasted AI text. Neither is proof of anything. Both are evidence you can see, which beats a verdict you cannot check.
Picture a pair of dice. Every time Claude writes, it rolls them to pick between near-equal words: begin or start, big or large. A watermark secretly weights those dice in a pattern set by a private key that only Anthropic holds. The words that come out are ordinary, so the text reads normally. The mark is not in any single word. It lives in the pattern across hundreds of rolls.
To read that pattern, you need the key that weighted the dice. Without it, the rolls look random, because they were designed to. Anthropic tuned the system so the bias is invisible to readers and to any tool that lacks the key. Human-rater studies and live A/B tests confirmed the marked text reads the same as unmarked text. That is the whole point of the design.
So this is not a puzzle waiting to be cracked by a clever website. It is sealed by math. Anyone who tells you their tool detects the Claude watermark today is guessing, and a guess about someone's honesty is exactly how the false-accusation problem started.
There is a difference between "did Claude write this" and "does this read like Claude." We cannot answer the first. Nobody can, yet. We can help with the second, and we are careful to call it what it is.
Our editor measures writing habits: the tell-words, the rule-of-three sentences, the flat rhythm that machine drafts tend toward. It reports them per 1,000 words against real benchmarks, so you see a number in context rather than a verdict pulled from nowhere. It also x-rays for hidden characters, the invisible marks and disguised letters that ride along in copy-pasted AI text. Both are evidence you can see and check. Neither is proof of authorship, and we never dress them up as proof.
Anthropic has said it will release tools, and an API, that let anyone check text for its watermark. When that arrives, we will build a watermark check powered by the official tool, labeled as exactly that, with the processing-not-authorship caveat printed beside every result. That is the honest way to answer the question, and it becomes possible the day the keyholder opens the door. Not before.
Yes. Claude models launched on or after 2 August 2026 embed an invisible watermark in generated text, applied worldwide. Anthropic confirmed it on its Help Center on 11 August 2026 and is working to extend it to older models.
No. It is not a visible mark or a hidden character. It is a statistical pattern in which words the model chose. The text reads normally because it is normal text.
Right now, you cannot, and neither can any website. Reading the mark requires Anthropic's key. Anthropic has promised public detection tools and an API; until those ship, every "Claude watermark detector" you find online is guessing. When the official tool exists, we will host a checker built on it.
No. Detectors guess from writing style; they do not hold Anthropic's key. A detector may still flag Claude text for sounding machine-drafted, but that is the same style-guessing they have always done, with the same false-positive record.
No, and this is the most important answer on this page. Anthropic states the mark shows content was processed by Claude, not authored by it. Paste your own essay in for a grammar check and the output can carry the mark. Marked does not mean cheated.
Anthropic itself says heavy editing, translation, or retyping can strip the mark, while light edits may leave it intact. That cuts both ways: absence of a mark proves nothing about origin. We will not help anyone strip marks to deceive; we simply report what the technology can and cannot hold.
Not because of this. If your work may be scrutinized, keep your drafts and version history; that remains the strongest proof of authorship there is. Disclosure policies at your school or workplace matter more than the mark does.
The day Anthropic releases its public detection tools, we will add a watermark check powered by the official tool, labeled exactly as what it is, with the processing-not-authorship caveat printed beside the result. Until then, the honest answer to "can you check for the Claude watermark" is no, and nobody else can either.
Claude text now carries an invisible statistical watermark. Only Anthropic can read it for now, it proves processing rather than authorship, and editing can strip it. Distrust anyone selling certainty in either direction. Measure what can be measured, keep your drafts, and check the writing itself.
Measure your text against the benchmarks →Sources: Anthropic Help Center, "How Claude marks AI-generated content" (updated 11 Aug 2026). TechCrunch, 11 Aug 2026. Forbes, 13 Aug 2026, on the processing-versus-authorship distinction. The Decoder, 11 Aug 2026, on marks stripping under heavy editing. Matter of Newby v. Adelphi Univ., 2026 NY Slip Op 26021. Related: the AI words list with our measured benchmarks · what does deslop mean.