THE WATERMARK
A statistical signal in word choices
Anthropic uses a version of Google DeepMind's SynthID Text. A private key and the preceding words nudge which word Claude picks, and only a holder of the key can test for the pattern.
REMOVAL GUIDE, UPDATED OCTOBER 2026
How to remove Claude watermark, and what the numbers say about each method
Claude's watermark is a statistical pattern in its word choices, not a hidden symbol. That decides which removal methods work. Below: eight steps, six methods compared, and the published test numbers behind each one.
Scan my text free Updated
claude.ai
Today
Draft a two line reply to the client.
Message Claude
MacBook Pro
9:41
ClaudeCleanLocal scanPASTED FROM CLAUDE
SCANNING
…hidden characters
U+200Bzero-width space1
U+00ADsoft hyphen1
U+00A0no-break space1
Step 1 done. Step 2 is the rewrite with a non-Claude model.
9:41
ClaudeCleanLocal scanPASTED FROM CLAUDE
SCANNING
…hidden characters
U+200Bzero-width space1
U+00ADsoft hyphen1
U+00A0no-break space1
Step 1 done. Step 2 is the rewrite with a non-Claude model.
Short answer
You remove the Claude watermark by replacing Claude's word choices, not by deleting characters. Anthropic says light editing probably won't remove it completely, while a complete rewrite where every word is replaced will. Rewrite it yourself, or with a different model, and check the meaning afterwards.
THREE LAYERS
What is the Claude watermark, and what can you remove?
People mean three different things by Claude watermark. Only one of them is the real watermark, and each needs a different fix.
NOT THE WATERMARK
Hidden characters from copy and paste
Zero-width spaces, soft hyphens and look-alike letters can ride along when text is copied between apps. Anthropic states nothing is added to Claude's text, so these come from the pipeline, and a scan removes them.
NOT THE WATERMARK
Style habits readers notice
Em dashes, tidy summary closers and 'not X but Y' lines make text feel machine written. They are what AI classifiers and people react to, and an edit fixes them without touching the statistical signal.
TRY IT: WHAT EACH FIX REACHES
One draft, three layers, four ways to handle it
Pick a method. The panel shows which layer it changes.
Pasted from Claude
ThanksforthequickturnaroundZWSPonthecontract.SHYIt'snotjustadeal,it'sapartnership.WewillsignandreturnitbyMonday.
Claude word choiceZWSP hidden characterabc style tell
Hidden characters2 present
Style tells1 line remains
Statistical watermarkUntouched
Every tinted word is a Claude word choice. That is where the statistical watermark lives.
PUBLISHED TEST NUMBERS
How much editing does it take to remove a text watermark?
Anthropic has not published removal numbers for Claude. These are the closest published measurements, each on its own watermark and test setup.
17%
OpenAI's detection rate on its own watermark after 25% of words were swapped for synonyms, down from about 92% (400-token passages)
66%
The same OpenAI test with only 10% of words swapped: the watermark was still caught two times in three
5.4%
SynthID-watermarked text still caught after a DIPPER paraphrase, down from 87.6% (true positives at a 5% false positive rate)
57.2%
Watermark detection on GPT2-XL text after an 11B-parameter paraphraser, down from a perfect score (1% false positive rate)
DRAG THE SLIDER
How much of a watermark survives each kind of edit
Left: OpenAI's synonym test on its own watermark. Right: two paraphrase studies, before and after.
OPENAI TEXTGRAIN, 400-TOKEN PASSAGES
17%still detected
SynthID text, DIPPER paraphrase
DAMAGE, 2025
Before0.0%
After paraphrase0.0%
GPT2-XL watermark, 11B paraphraser
DIPPER, 2023
BeforePerfect score
After paraphrase0.0%
Different watermarks, models and setups. Compare the direction, not the decimals.
Read these numbers with care
OpenAI's figures describe its textGrain watermark, not Claude's. The DAMAGE study comes from Pangram Labs, a detector company, and ran on a small Gemma model. The direction is consistent: small edits leave most of the signal, deep paraphrasing by a different model removes most of it.
START WITH THE PART YOU CAN CHECK
Clean the characters, then rewrite the words.
ClaudeClean scans hidden Unicode in your browser, then offers a Light, Medium or Strong rewrite through a non-Claude model with a before and after diff. Rewriting is best effort. No detector result is promised.
- 300 rewrite credits a month
- 20,000 characters per request
- .txt, .docx and .pdf files
STEP BY STEP
How to remove Claude watermark from text in 8 steps
The order matters: decide whether you may, clean the characters, then change the words that carry the signal.
THE FLOW, END TO END
From a Claude reply to text you can send
Steps 2 to 8 below in motion: copy, clean the characters, rewrite the words, then read it back.
Copy from Claude
Scan characters
Rewrite the words
Read back, copy
claude.ai
Draft a two line reply to the client.
Thanks for the quick turnaround on the contract. We will sign and return it by Monday.
Copy
Reply to Claude
MacBook Pro
9:41
ClaudeCleanLocal scanPASTE TEXT OR UPLOAD A FILE
Thanks for the quick turnaroundZWSP on the contract.SHY We will sign and return it by MondayNBSP.
READY TO SCAN
0found
U+200Bzero-width space1
U+00ADsoft hyphen1
U+00A0no-break space1
Copy from Claude
Scan characters
Rewrite the words
Read back, copy
9:41
ClaudeCleanLocal scanPASTE TEXT OR UPLOAD A FILE
Thanks for the quick turnaroundZWSP on the contract.SHY We will sign and return it by MondayNBSP.
READY TO SCAN
0found
U+200Bzero-width space1
U+00ADsoft hyphen1
U+00A0no-break space1
- 01
Check whether you may remove it
Read the policy that governs the text first: course rules, a publisher's AI policy or a client contract. If AI use must be declared, declare it, because a rewrite changes the words, not how the draft was made. - 02
Scan for hidden characters
Run a character scan to catch zero-width spaces, soft hyphens, odd non-breaking spaces and look-alike letters picked up while copying. This is cleanup, and it leaves the watermark itself untouched. - 03
Find the passages Claude wrote
Mark which sentences are Claude's words and which are yours. Anthropic says the watermark only applies to words Claude chooses, so your own sentences carry no signal and need no work. - 04
Rewrite from the idea, not the sentence
Close the draft, write each point again in your own words, then compare. Starting from the meaning replaces far more words than editing line by line, which is the depth Anthropic says it takes. - 05
Use a different model if you rewrite with AI
Never ask Claude to paraphrase itself, because its new words are watermarked too. The DIPPER researchers note the paraphraser must be a different model from the one that wrote the text. - 06
Push past light synonym swaps
Swapping a word here and there leaves most scored words in place. OpenAI's own test on its watermark needed 25% of words replaced to cut detection to 17%, and Anthropic sets the bar at every word. - 07
Add what only you know
Replace generic lines with your own examples, numbers and sources. New material changes more of the text, and it is also the part a reader or marker values most. - 08
Read it back for meaning
Heavy paraphrasing drifts. Check every fact, name, number and quote against the source, and fix sentences that now say something different from what you meant.
SIX METHODS SIDE BY SIDE
Which Claude watermark remover method works?
What each method does to the three layers, with the evidence behind the watermark column. Signal numbers come from different watermarks and setups, so compare direction, not decimals.
| Method | Hidden characters | Statistical watermark | Style tells | Cost and effort |
|---|---|---|---|---|
| Character cleaning | Removes zero-width, soft hyphen and look-alike characters | No effect: Anthropic says no hidden characters are added | Unchanged, except optional em dash swaps | Free, seconds, runs in the browser |
| Light manual edit | Only if you retype the edited lines | Anthropic: probably won't remove it completely | Some tells remain | Low effort, minutes |
| Synonym spinner | Depends on the tool | Partial: OpenAI saw 92% to 66% at 10% swaps, 17% at 25% | Often clumsy word choices | Free to cheap, seconds |
| Rewrite with a non-Claude model | Output is newly generated text | Strong drop in tests, e.g. 87.6% to 5.4% (DAMAGE) | Depends on the prompt and model | Paid credits, minutes plus a review |
| Translation round trip | Output is newly generated text | Weakens it, unless Claude translates (then it carries over) | Meaning and idiom drift | Free, but needs careful checking |
| Rewriting it yourselfMost reliable | New keystrokes, nothing carried over | Anthropic: a full rewrite of every word will remove it | Your own voice | Highest effort, the most reliable |
Swipe to compare →
Sources: Anthropic, OpenAI's figures via BleepingComputer, DAMAGE (Pangram Labs) and Google's SynthID docs on translation.
HOW CLAUDECLEAN HANDLES IT
Character cleanup in the browser, word changes on request
The scan and clean never leave your device. The rewrite is a separate, opt-in step you can inspect line by line.
01
> source text
Paste the Claude draft
Paste text or upload a .txt, .docx or .pdf. Anonymous use covers 2,000 characters, enough to test one section.
02
scan report
3 zero-width1 soft hyphencleanable signals2Strip what copy and paste left behind
The scan flags zero-width characters, soft hyphens, odd spaces and look-alike letters, then cleans them locally. The statistical watermark is not a character, so this step leaves it alone.
03
before → after
clearusefulanswerfocusedhelpfulresponse
Copied ✓Rewrite with a non-Claude model
Light targets about 35% of words reworded, Medium about 60%, Strong 85% or more, in up to 3 passes. You review the diff and fix anything that drifted.
PRICING
Clean the text free. Rewrite more with Pro.
The scan and hidden-character cleanup are free. Pro adds room for rewrites and files.
PRO
CLAUDECLEAN PRO
ClaudeClean Pro
$5/ month
Cancel anytime.
- 300 credits every month (≈300,000 characters of rewriting)
- 20,000 characters per request
- Clean .txt, .docx and .pdf files
- Light / medium / strong intensity
- Dashboard with history
Checking account…
Free: 10 credits a month with an account, or try 2,000 characters right here without one. Strong rewrites cost 2× credits.
EDGE CASES
Does translating Claude text or using another AI remove the watermark?
The questions people ask after the obvious ones, answered from Anthropic, Google and OpenAI's own statements.
FIVE ROUTES, FIVE OUTCOMES
What happens to the watermark on each route
Tap a route. The verdicts come from Anthropic, Google and OpenAI statements quoted below.
Claude draft Watermark in the word choices
Does translating it remove the watermark?
Does running it through ChatGPT remove the Claude watermark?
Words ChatGPT picks are not Claude's choices, so the Claude signal fades as more words change. Since October 5, 2026, eligible ChatGPT text in the EU can carry OpenAI's own watermark, so you may swap one mark for another. API watermarking is opt-in and off by default.
Is a short paragraph detectable at all?
Less so. Anthropic says detection is weak on small samples and confidence rises with length. OpenAI reported about 80% detection at 200 tokens and 95% at 400 for its own mark, so short text gives any detector less to work with.
Does the watermark apply to code?
Less. Anthropic says code carries less watermarking, since there are fewer free word choices in code. The prose around it, such as comments and docs, is marked like any other text.
Can any remover prove that it worked?
No. The detector needs Anthropic's key and sits in private preview. BleepingComputer found that remover sites flooding search cannot prove their claims, and that heavy rewriting is the only known way to remove the mark.
HABITS THAT HELP AND HURT
What should you avoid when removing an AI watermark?
Most failed attempts make one of the mistakes on the right.
THE MOST COMMON MISTAKE
Why asking Claude to reword itself does not help
Same sentence, two rewrites. Tinted words are Claude word choices, the place the watermark lives.
Original Claude sentence
Thanksforthequickturnaroundonthecontract.
Ask Claude to paraphrase it
Claude words: 0 of 0Watermarked again
Rewrite it from your notes
Claude words: 0 of 0No Claude word choices
Do this
- Rewrite from your notes. Starting from the idea replaces more words than polishing Claude's sentences.
- Use a different model for AI rewrites. A model that is not Claude chooses words with no Claude key behind them.
- Change structure, not just vocabulary. Merging, splitting and reordering sentences changes far more of the scored words.
- Keep a copy of the original. You need it to check that facts, quotes and numbers survived the rewrite.
- Scan characters before you paste anywhere. It removes copy and paste debris that can break formatting or trip filters.
Avoid this
- Asking Claude to paraphrase itself. Its rewrite is new Claude text, watermarked again.
- Trusting a character stripper alone. It removes debris, not the statistical pattern.
- Stopping at a few synonyms. OpenAI's 10% swap test still caught its mark 66% of the time.
- Paying for a remover that promises a clean score. With the detector private, nobody can verify that promise.
- Skipping the read-back. Spinners and translations change meaning in ways a quick glance misses.
RULES BEFORE TOOLS
Is it legal or allowed to remove the Claude watermark?
Removing the signal is a technical act. Whether it is acceptable depends on who receives the text and what they asked for.
Academic integrity: the watermark is not the rule
A school's policy is about who did the work, not about a statistical pattern. If your course bans or limits AI help, rewriting Claude's draft to hide it breaks that policy whether a detector fires or not. Where AI help is allowed with disclosure, disclose it.
SCHOOLS
Course policy decides
Many courses allow AI for brainstorming but not for submitted prose. Read the syllabus and ask your instructor when it is unclear.
EU AI ACT
Article 50 targets providers and publishers
Providers must mark AI output in a machine-readable way. A Cloud Security Alliance note puts the deadline for systems already on the market at December 2, 2026, with fines up to 15 million euros or 3% of turnover.
WORK AND CLIENTS
Contracts can ask for disclosure
Publishers, agencies and clients increasingly write AI use into briefs and contracts. Rewriting a draft does not change what you agreed to tell them.
HOW WE GOT HERE
When did Claude start watermarking text?
The dates that shaped what a watermark is and who can read it.
OCTOBER 23, 2024
SynthID Text published in Nature
Google DeepMind described its text watermark in Nature, tested live on nearly 20 million Gemini responses, and released it open source. Claude's watermark is a version of it.AUGUST 2, 2026
EU AI Act Article 50 applies
Transparency rules for AI-generated content start to apply. Anthropic marks models launched from this date at launch, and older models move over in a transition period.AUGUST 13, 2026
Remover tools questioned
BleepingComputer reported a wave of watermark remover sites and found almost none could prove their claims.AUGUST 14, 2026
Anthropic announces the Claude text watermark
Anthropic confirmed the statistical watermark and a detection API in private preview for eligible organizations, with access expanding over time.OCTOBER 5, 2026
OpenAI adds textGrain in the EU
OpenAI began marking eligible ChatGPT and Codex text in the EU, rolling out over the following weeks, with opt-in API marking worldwide.DECEMBER 2, 2026
Grace period ends for older systems
Per the Cloud Security Alliance reading of the Act, systems already on the market before August 2 must meet the machine-readable marking duty by this date.
WHO CAN READ IT
Can anyone check if the Claude watermark is still there?
This is why every removal claim online is unverifiable for now.
ANTHROPIC ONLY
The key stays private
Detection needs Anthropic's key. The API is in private preview for groups such as regulators, researchers, media and fact checkers, via an interest form.
LIMITS
It cannot name the author
Per Anthropic, a positive result means Claude was likely involved. It cannot tell Claude writing from Claude heavily editing, and a negative does not prove a human wrote it.
FAILURE MODES
Edits, translation and mixing
Claude's help center says detection can fail on text that is heavily edited, paraphrased, translated or mixed into other writing.
PUBLIC DETECTORS
Turnitin and GPTZero can't read it
Without the key, third-party tools fall back on their own style classifiers. Their scores say nothing about whether the watermark is present.
WATCH
A short video on why the watermark is not permanent
WHERE THE FACTS COME FROM
Sources for removing the Claude watermark
Primary statements from Anthropic, OpenAI and Google first, then the research and reporting behind the numbers.
ANTHROPIC
Claude's text watermark
Light edits probably won't remove it, a full rewrite will. Translation by Claude keeps it.
View sourceCLAUDE HELP CENTER
How Claude marks AI-generated content
Detection can fail on heavily edited, paraphrased, translated or mixed text.
View sourceOPENAI
Approach to EU text provenance
textGrain on eligible EU ChatGPT and Codex text; API marking opt-in worldwide.
View sourceBLEEPINGCOMPUTER
OpenAI adds invisible watermarks in the EU
OpenAI's synonym test: 10% swaps to 66% detection, 25% swaps to 17%.
View sourceBLEEPINGCOMPUTER
AI watermark removers can't prove they work
Character stripping is measurable; heavy rewriting is the only known removal.
View sourceNATURE
Scalable watermarking for LLM outputs
The SynthID Text paper Claude's watermark builds on, Dathathri et al., 2024.
View sourceGOOGLE AI FOR DEVELOPERS
SynthID text documentation
Robust to mild paraphrasing, weakened by thorough rewrites and translation.
View sourceARXIV, PANGRAM LABS
DAMAGE: detecting adversarially modified AI text
SynthID detection fell from 87.6% to 5.4% after DIPPER paraphrasing.
View sourceARXIV, KRISHNA ET AL.
Paraphrasing evades detectors of AI text
DIPPER paraphrases cut watermark detection to 57.2% on GPT2-XL.
View sourceFAQ
Claude watermark removal FAQ
Yes, with enough rewriting. Anthropic says light editing probably won't remove it completely, while a complete rewrite where every word is replaced will. There is no button that strips it, because it lives in word choices, not in characters.
A tool can only work by rewriting the words. Character cleaners fix hidden Unicode but leave the statistical signal untouched. Since Anthropic's detector is private, no remover can prove a result. No detector result is promised.
Not if Claude does the translating. Anthropic says a translation written by Claude carries the watermark. Translating with a different system weakens it, and Google says SynthID confidence drops sharply after translation.
No. The watermark applies to every word Claude chooses, so a Claude rewrite is new watermarked text. Rewriting yourself or with a different model is what changes the word choices that carry the signal.
No. Reading the watermark needs Anthropic's private key, and the detection API is in private preview for eligible organizations. Turnitin and GPTZero run their own classifiers, which judge writing style, not the watermark.
It depends on where the text goes. Removing it does not change who wrote the words. Schools, publishers and clients set their own disclosure rules, and the EU AI Act asks for disclosure on some AI text published for the public.