warmprose Find the sentences that sound machine-made.Rewrite them warm, in your own voice.

Claude's text watermark: what a remover can and cannot do

Since August 2026, text written by Claude has carried a watermark. Anthropic explained the method in public, and the explanation is worth reading before you look for a tool that strips it out. Two facts come first: the watermark changes no words and adds no hidden characters, and the only way to check for it is an API that Anthropic has not opened to the public. So the practical question is narrower than the search term suggests, and this page works through what a removal tool can and cannot change.

What the mark physically is

Anthropic's post describes the method as a version of SynthID-Text, the approach Google DeepMind published in Nature in 2024. It does not insert anything into the text. Instead it changes how the model picks between words that would all work equally well.

The example Anthropic uses is a sentence about the weather. After "The weather today was cold and", the next word will not be "sugary". It could be "overcast" or "grey", and either one reads fine. In normal operation a random number settles that choice. With watermarking, the random number is replaced by something derived from a key and the words just before it. The words stay random from a reader's point of view. They are not random from the key holder's point of view.

"Watermarking carries no identifying information and can't be traced to a specific person, organization, or chat."

Anthropic, How Claude's text watermark works, 14 August 2026

Two more details from that post matter for anyone worried about what the mark reveals. It says the watermark is not specific to Claude, because other large model providers signed the same EU code and are implementing their own versions. And it says the mark applies globally, not only to users in Europe, because Anthropic had no reliable way to scope it by region at launch.

What the mark can and cannot prove

This is where most of the confusion sits, and it is worth being blunt about it. The watermark answers one question: how likely it is that Claude was involved in producing this text. It does not answer the question people usually care about.

The mark can showThe mark cannot show
Claude probably touched this text at some pointThat a human did not write it
The text is consistent with a given keyThat another AI wrote it instead
Something about a long passageAnything reliable about a short one
That processing happenedWhich person, account, or chat did the processing

Compiled from Anthropic's watermark post and the Claude help center article on marking, both read 3 November 2026.

Anthropic draws a distinction that gets lost in most coverage: the mark cannot separate "Claude wrote this" from "Claude edited this heavily". A person who drafted a report themselves and asked Claude to tighten three paragraphs ends up with a text that may carry the mark. The ideas and the structure are the person's. The mark does not know that.

Does editing take it out

Anthropic answered this directly, and the answer is more specific than the removal-tool marketing around it.

"Light editing probably won't remove the watermark completely", while "a complete rewrite where every word is replaced will".

Anthropic, quoted in TechCrunch, 15 August 2026

Anthropic then adds a sentence that the whole removal-tool category tends to skip. In the case of a complete rewrite, the company says, it becomes arguable whether the text can still be described as AI-generated at all.

That sentence undercuts the whole removal category, because it concedes that a successful removal may also remove the grounds for calling the text AI-generated. If every word has been replaced, the output is a new piece of writing. Calling it AI text that has been disguised is a claim about history, not about the words in front of you. Whether that history matters depends on who is asking and why, and the watermark itself cannot settle it.

Anthropic also flags where the mark is weakest. It thins out on factual passages, because a sentence like "Newton's most famous work was called Principia" has one correct next word and no room for a choice. It thins out on short passages, because there are fewer decisions to work with. It thins out on proofreading jobs, because when Claude only fixes grammar, nearly all the words remain the author's and there is little for the mark to attach to. Code carries less of it for the same reason.

Who can run the check

Here is the fact that changes the question most: detection is not a public service. Anthropic released a detection API in private preview, and the access list is narrow: regulators, law enforcement, media organisations, fact-checkers, independent researchers, educational organisations, and EU civil society groups, plus enterprises that carry their own legal obligation to verify.

There is a registration form for access. There is no consumer tool that takes a paragraph of text and returns a watermark verdict, and nothing in Anthropic's published material suggests one is coming to the general public.

The practical effect is narrower than the search term suggests. If you are an individual writer, a student, or a marketer, or if you run a small business, you are not on the list of parties who can check a Claude watermark. The scenario where a check on your text changes your life mostly applies to institutions checking at scale, not to a person checking a single document.

This does not make the mark irrelevant. It does mean that a lot of the fear around it is aimed at the wrong target. The thing that can read a Claude watermark is not the thing sitting in a teacher's browser tab or a client's inbox.

So what does a removal service do

There are now sites built around the search term, and the honest description of what they offer is rewriting. One common version pipes your text through a paraphraser and hands it back. Another runs it through a model with an instruction to vary the wording. A third wires both together and sells the output as clean.

Their claim rests on the sentence from Anthropic quoted above, that a complete rewrite where every word is replaced removes the mark. That is a fair reading of what Anthropic said. Where the category gets loose is in the implication that the result is undetectable in general. The mark is one signal. The style of writing is another, and rewriting changes both, which is why anyone judging a text on its prose rather than its watermark will reach a different verdict.

Two practical points apply to any of these tools. First, rewriting a long document costs real money per pass, and replacing every word in a three-thousand-word report usually needs more than one pass. Second, output quality drops as rewriting gets more aggressive, because the tool is being pushed to change choices that carried meaning.

Watermarks and AI detectors are not the same thing

These two ideas get merged constantly and they work in opposite directions. Anthropic made the distinction itself, noting that watermark checking is fundamentally different from the pattern-based detection that tools like Pangram use.

WatermarkStyle detector
What it readsA keyed pattern in word choicesHabits in the prose
Who can checkHolders of the key, in a closed previewAnyone with the website
What it needsEnough text for many word decisionsEnough text to see a pattern of habits
Effect of rewritingA full rewrite removes itDepends on how the rewrite reads
False positivesNot the main risk, since access is limitedA known and documented problem

Compiled from Anthropic's watermark post and the TechCrunch follow-up, 15 August 2026.

The difference matters because the two produce different kinds of worry. A watermark is close to binary and backed by a key. A style detector produces a probability estimate with a wide error bar, which is why arguments about it are louder. Rewriting addresses both, but it addresses them unevenly, and the second one is the one most people run into.

If Claude only polished a draft you wrote

This is probably the largest group of people searching for a remover, and it is the group with the least to worry about. You wrote the piece. You asked Claude to fix the grammar, tighten a few sentences, and cut a paragraph. What comes back is yours with edits.

Anthropic's own description says the mark has almost nothing to attach to in this case, because nearly all the words are yours. Whether enough of it survives to register depends on how long the text is and how heavily Claude changed it. On a light edit of a short document, the answer tends toward no.

If the text is long and the edits were heavy, the mark may survive, and the reasonable response is not to look for a stripping service. It is to finish the work yourself. Read the edited sections and rewrite the sentences that no longer sound like you. This is normal editing practice. It improves the document, and as a side effect rather than a deliberate goal, it moves the text away from a machine's word choices.

Where a rewriter fits, and where it does not

Our rewriter is a style tool, not a watermark remover. When you paste in text, it flags the habits that make English read like a model wrote it and offers plainer alternatives. The point of it is readability, and the measure is whether a person can read the result without hearing a machine. The checker on this page runs on the same analysis, and its free tier covers a short sample without an account.

That overlaps with watermark removal, since the same operation changes word choices, but it is not the same goal and we do not sell it as a stripping service. If your reason for rewriting is that the text sounds stiff, the tool does what you need. If your reason is that a specific institution will run a keyed check, rewriting is not a guarantee, because a full rewrite is a large ask and a partial one leaves parts in place.

What we will not do is promise a clean result we cannot verify. The detection API is closed, so no vendor outside that access list can test their own output against the key. Any site claiming a verified clean pass is describing something they cannot check, and that applies to us as much as to anyone. That is also why the pricing page lists word limits in plain numbers rather than promising outcomes: the limit is a fact, the outcome is not.

The reason this question has a short answer

Putting the threads together: the watermark exists, it is real, and it is switched on globally. It does not identify you. It does not prove a human did not write the text. The only parties who can read it are institutions with access to a closed API. A complete rewrite removes it, according to the company that made it, and the same company points out that after a complete rewrite the label itself becomes arguable.

For most readers, that adds up to a smaller problem than the search term implies. The scenario worth preparing for is not a stranger running a keyed check on your email. It is an institution, an editor, or a client reading the prose and deciding it sounds off. That problem does not need a remover. It needs writing that sounds like a person, which is a different job with a longer history and better tools.

Questions people ask

Does copying and pasting remove the Claude watermark?

No. Anthropic states that the mark is part of the text, so it travels with the text when copied and pasted, and may persist through some editing.

Can I check my own text for a Claude watermark?

Not with a public tool. Detection runs through an API in private preview for regulators, law enforcement, media, fact-checkers, researchers, educational bodies, and EU civil society groups, plus enterprises with their own legal obligation.

Does the watermark mean Claude wrote the whole thing?

No. Anthropic says the mark shows Claude was likely involved at some point and cannot separate "Claude wrote this" from "Claude edited this".

Is the watermark the same as an AI detector?

No. A watermark is a keyed pattern in word choices. A detector reads habits in the prose. Anthropic describes the two as fundamentally different.

Does it apply outside the EU?

Yes. Anthropic applies marking globally at launch rather than trying to scope it by region.

What about images and files?

Claude attaches Content Credentials, a signed C2PA metadata label, to supported file types such as PNG and JPEG. That label sits in the file metadata and is a separate mechanism from the text watermark.

Where to go next

If the thing you are dealing with is prose that sounds like a model, the checker is the faster route and it needs no account for a short sample. If you want the rules behind the flags, the list of habits our checker looks for is public. And if you are comparing tools that promise to clean text, the comparison pages below cover what each one charges and what it actually does.

Run your text through the checker, then read the list of habits that make writing sound machine-generated if you want to fix them yourself. For the wider market, QuillBot and Undetectable AI are covered in detail, and the guides index lists everything else on the site.

The checker's rule list is versioned: every change to the rules behind those flags is dated and listed on What changed in warmprose.

Sources: Anthropic, "How Claude's text watermark works", 14 August 2026; Claude Help Center, "How Claude marks AI-generated content"; TechCrunch, 11 and 15 August 2026; The Verge, August 2026. All read 3 November 2026. Anthropic continues to add marking to models released before 2 August 2026, so the coverage of any specific model should be checked against their published table before relying on it.