Zuro

Claude watermark remover — free, and no signup to start

Paste Claude, ChatGPT or Gemini output. See the AI-likeness score, then re-generate every token so the watermark has nothing left to attach to.

Source text

Short answer

A Claude watermark is a statistical bias in which near-equivalent words the model picked, not hidden characters. Copy-paste and light edits do not remove it. Re-generating every token does. Zuro does that in one click and shows the score before and after.

What the Claude watermark actually is

Anthropic applies its text watermark at the sampling layer, below the model, using a version of Google DeepMind's SynthID-Text approach. Nothing is inserted into the text — there is no hidden Unicode to highlight and delete. Instead, at every token the model nudges its choice within a shortlist of words that were already plausible. The result reads exactly like normal Claude output to a person, and only Anthropic's key can score the text against that nudge later.

This matters because most advice about removing AI markers is wrong. Adding a few synonyms, an em dash and a contraction changes how the text reads. It does not change which tokens were sampled — and the watermark rides the tokens. Anthropic says as much: light editing probably will not remove it, but a complete rewrite where every word is replaced will. That is the operation this page automates.

Before and after

Before · score 78 (AI-like)

It is important to note that leveraging a robust content strategy is crucial in today's landscape. Moreover, teams that foster a seamless workflow will navigate the complexities of publishing with greater confidence.

After · score 24 (human-like)

A solid content plan is what makes the difference. Teams that get their workflow into shape stop firefighting every release, and publishing starts to feel routine instead of risky.

Scores are a stylometric heuristic over sentence rhythm, lexical variety, em-dash rate, AI transition phrases and paragraph regularity. They are not Anthropic's keyed read, so treat them as a pressure gauge rather than a certificate. The test report publishes the sample sizes and the cases where removal still fails.

What this does not do

It does not make text "undetectable" forever, because detectors keep changing and none of them read the watermark either. It does not handle images inline — those need C2PA metadata stripping and inpainting, which the image page covers. And it will not rescue a twelve-word fragment: below roughly fifty words there is not enough sample for any of this to mean much.

When to use back-translate

Paraphrase is the default because it keeps quality high. Back-translation pushes the passage through a pivot language and back, then smooths it. It is heavier and it can mangle idiom, so reach for it only when a paraphrase still scores high — usually on very short or very formulaic text. Whatever mode you use, re-read the output before you publish it.

Price, in plain terms

The score check is free and needs no account. Your first removal is anonymous too. After that a free account holds three removals, a 20-removal pack is $5 and never expires, and unlimited use is $3.99 a month or $29 a year. Teams are $19 a month for five seats, and the API is a cent per thousand words.

Where to go next

Model-specific pages explain each vendor's mark and its quirks: Claude, ChatGPT, Gemini. Format guides cover Google Docs, PDFs, WordPress and Notion. If you only want a verdict, use the free detector. If you are building, start with the API or the MCP server, and see the full index.

What people use it for

Handed in an essay that came back clean on the two detectors my course uses. The before-and-after score was the part that convinced me it had done anything.
Mara T. — graduate student
We paste client copy through it before it goes near the CMS. The revision log stops filling up with AI drafts.
Dan R. — agency editor
Back-translate saved a very formulaic product description that a normal paraphrase could not move. Re-read it after, as they say.
Sofia K. — ecommerce ops

Reviews are collected from users and shown unedited. Ratings are self-reported; the aggregate figure above is not independently audited.

Questions

Does it cost anything?

The check is free and needs no account, and the first removal is anonymous. Free accounts get 3 removals; a $5 pack adds 20 that never expire; unlimited is $3.99/month or $29/year.

Can any tool read the watermark directly?

Not without the producer's key. Anthropic's and Google's marks are readable only by them, so everything public — including Zuro — measures the statistical footprint instead. A score is evidence, not proof.

Will light editing do it?

No. Only replacing every token does, which is what re-generation is. Synonym swappers leave most of the bias in place.

Is my text stored?

It is processed in memory. Logged-in sessions are stored so they appear in your dashboard, and you can log out at any time.

What about images?

Images need C2PA metadata stripping and inpainting, which the image page walks through. The engine here is text.