Can You Remove the Claude Watermark?

The one method that works

Claude's watermark is a keyed bias in word choice. When several words would do, Claude picks the one its key prefers, and the pattern of those picks is what Anthropic's detector measures. Anthropic's own description, from its September 1 announcement, calls it "a numerical way of determining the likelihood that Claude was involved in writing a piece of text."

That tells you exactly how to remove it. Replace the words and the mark goes with them. Anthropic said as much on August 14: light editing "probably won't remove the watermark completely," and a complete rewrite where every word is replaced will. Scott Aaronson, whose 2022 proposal is the root of this family of schemes, named the practical defeat path on August 22: translation, emoji interleaving, or paraphrase through an open model.

The research agrees, and has for years. Krishna and colleagues showed in 2023 that paraphrasing evades detectors of AI-generated text. Zhang and colleagues proved in "Watermarks in the Sand" that no strong watermark can survive an attacker who can make quality-preserving edits, which is what a paraphrase is. A statistical watermark is a property of one specific word sequence. Change the sequence and you lose the property.

How to do it

1. Run the Claude output through a model that is not Claude. A local open model is the obvious choice: Gemma, Llama or Qwen through Ollama or LM Studio keeps the text on your machine. Any hosted non-Anthropic model works too.
2. Ask for a full rewrite, not a polish. Something like: rewrite this in your own words, sentence by sentence, keeping every fact, the structure and the tone, and do not summarize or add anything.
3. Read the result against the original. Rewriters drop qualifiers, flatten numbers and occasionally invent. The facts are your responsibility.
4. Do not use Claude for the rewrite. Anthropic states that every word Claude chooses carries the mark, so a Claude paraphrase or a Claude translation keeps it.

Translation through a non-Claude route works for the same reason: the word choices in the new language were made by something else.

What it costs

The signal is proportional to how much of the original wording survives. Sebastian Raschka's August 22 explainer describes a curve, not a switch: replace half the words and roughly half the signal remains, and a long enough passage can still score. A light rewrite is therefore worse than useless. It costs quality without clearing the mark. Either rewrite all of it or leave it alone.

A full rewrite also spends the reason you used Claude. Tone shifts, precision drops, and a smaller model will get some things wrong. If the text mattered enough to worry about the mark, it matters enough to proofread.

What does not work

The first row is what most "Claude watermark removers" sell. BleepingComputer put it plainly on August 13: "The hard part is the watermark itself. It does not live in hidden characters." A green result from a Unicode cleaner means no hidden characters were present. It says nothing about the mark.

Can you check that it worked?

No, and neither can anyone else outside Anthropic. As of September 8, 2026, Anthropic's detection API is in private preview for eligible organisations (regulators, media, fact-checkers, researchers, and enterprises with their own AI Act obligations) by access request. There is no public detector. An August test of five removers concluded there was "no clean getaway," and the reason is the same: nothing could be verified against the real detector.

That cuts both ways. Nobody can prove your rewritten text is clean, and nobody can prove the original was marked. An honest tool reports the Claude row as Unknown in both cases. If you rewrote thoroughly, the published mechanics say the mark is gone. That is a well-founded expectation, not a verified result.

What DetectMark does here

The text tool removes what can be removed and verified in the browser: hidden Unicode, clipboard HTML fragments, and stray spacing characters, with a diff of every change and a re-scan of the result. It labels the statistical row Unknown, with the reason attached.

Below that result it offers the rewrite described above, through OpenAI, as a separate step: a dialog states the exact size of the text, nothing is sent until you confirm, and the rewrite appears beside the original with a count of how much wording changed. It carries no verdict, because nothing can verify the outcome. If you would rather keep the text on your machine, the local-model route in the steps above does the same job.

If you need the mark gone, the method above is the whole answer. What you do with it is covered by the terms: the mark exists so that Claude's involvement can be established, and removing it to misrepresent authorship is your decision and your liability.