Remove the invisible watermark from Claude text.
Since August 2026, every text Claude writes carries a statistical watermark that detectors can read. Unmarker rewrites your text so it reads as yours, and keeps exactly what you meant.
Founder pricing for the first 500. No spam, one email at launch.
As covered by TechCrunch, Forbes and Axios (Aug 11 to 13, 2026): Anthropic now watermarks Claude output. Read what it means.
Our onboarding guide walks new customers through setup in three short steps, so nobody has to open the manual on day one.
Our onboarding guide takes new customers through setup in three short steps, so nobody needs the manual on their first day.
Illustrative values. Meaning check: passed. Formatting: kept.
Claude output is now watermarked. Here is what that means.
Three facts, sourced from Anthropic's help center and the press coverage of August 11 to 13, 2026.
On August 11, 2026 Anthropic said it watermarks text generated by Claude. The mark is statistical (the same family as Google's SynthID-Text): it lives in the choice of words, not in hidden characters, so you cannot see it or strip it with find and replace.
How the watermark worksIt applies worldwide to Claude models released since August 2, 2026, in the app and through the API. It survives copy-paste and light edits, and is only weakly detectable on very short texts.
Which texts are affectedThe timing follows the EU AI Act transparency obligations (article 50) that started to apply in August 2026. A side effect: a human text you proofread or translate with Claude comes back watermarked and can be flagged as AI-written.
Why false positives happenPaste. Rewrite. Verify.
Three steps, no account gymnastics. Bulk and API do the same thing at volume.
Paste
Drop in your text, Markdown or HTML. Unmarker estimates the watermark score of the input so you know where you start.
Rewrite
The text is rewritten at the word-choice level, where the watermark lives, then checked sentence by sentence against the original for meaning.
Verify
You see the estimated after score, a diff of every change and your formatting intact. Copy, download or push through the API.
Built like an instrument, not a magic button.
What is planned for launch. Everything below ships to early access first.
Meaning lock
Every sentence is checked against the original for meaning. If a rewrite drifts, it is flagged and redone before you see it.
Formatting kept
Headings, lists, links, code blocks and tables come out exactly where they went in. Markdown and HTML both supported.
Tone control
Keep the original voice, or shift it: more direct, more formal, plainer English. One setting, applied consistently.
Before / after detection score
See an estimated watermark score for the input and the output, side by side. You decide when it is good enough.
Bulk and API
Drop in a folder of files or call a REST endpoint. Same engine, same meaning checks, at any volume.
Privacy: deleted in 24 h
Texts are processed to produce your rewrite and are deleted from our servers within 24 hours. Never used for training.
People who write with Claude and deliver to someone else.
Content and SEO teams
You draft with Claude, edit heavily, and still get flagged by a client's checker. Ship copy that reads as your team's.
Freelance writers
Clients now run detectors on deliverables. Keep using your tools without arguing about a watermark score.
Translators and proofreaders
You paste a human text into Claude for a polish and it comes back marked. Unmarker returns it to a clean state.
Developers
Your product generates text with Claude. Add one API call before delivery so your output is not silently labelled.
Planned launch pricing.
Founder pricing is locked for the first 500 waitlist members. Prices are targets and may change before launch.
- 3 texts per day
- Up to 1,500 words per text
- Before / after score
- Meaning check and diff
Founder pricing for the first 500. No spam, one email at launch.
- Unlimited texts
- Up to 10,000 words per text
- Tone control
- Bulk upload, history, priority queue
Founder pricing for the first 500. No spam, one email at launch.
- REST endpoint, JSON in and out
- Same meaning checks as Pro
- Webhooks for bulk jobs
- Volume pricing on request
Founder pricing for the first 500. No spam, one email at launch.
Understand the watermark before you remove it.
Short, sourced explainers. No hype, links to the primary sources.
Claude's text watermark, explained
On August 11, 2026 Anthropic said every text Claude writes now carries an invisible statistical watermark. Here is what it is, which texts it covers, what survives editing, and why a proofread human text can be flagged.
Read the guideHow-toHow to remove the Claude watermark
Find-and-replace, spell checkers and light edits do nothing against a statistical watermark. Here is what actually neutralises it, what it costs you in meaning, and how Unmarker approaches the problem.
Read the guideComparisonAI text watermarks compared
Which AI vendors actually watermark the text they generate in August 2026, how each mark works, who can detect it, and what that means for anyone who publishes AI-assisted text.
Read the guideStraight answers.
Is it legal to remove the watermark?
Editing or rewriting a text you generated is legal in the jurisdictions we operate in; the watermark is a provenance signal, not a copy-protection or licence mechanism. Rules about disclosing AI-generated content can still apply to you (for example the EU AI Act, an employer's policy or a publisher's editorial rules). Unmarker does not change what you are required to disclose, and we do not provide advice on circumventing academic or exam rules.
Does it work on ChatGPT or Gemini text?
Unmarker is built for Claude's watermark, which is the one Anthropic announced in August 2026. Gemini text can carry Google's SynthID-Text signal, and OpenAI has not deployed a comparable watermark on ChatGPT output at the time of writing. Support for other watermarks will be added when they are actually shipped and measurable.
Will it change what my text says?
No. The rewrite is checked sentence by sentence against your original for meaning, and any drift is flagged and redone. Names, numbers, quotes, links, code and formatting are kept as they are. You always see a diff before you copy.
How fast is it?
A 1,000-word text takes a few seconds. Bulk jobs run in parallel and you get an email when they finish. The API returns synchronously for texts under about 3,000 words.
Does it work against GPTZero, Turnitin or Originality?
Those tools mostly estimate whether a text looks AI-written from its style, which is a different problem from reading Anthropic's watermark. Unmarker targets the watermark and shows you an estimated before / after score for it. It does not promise any result on third-party style detectors, and we make no absolute claims about any detector.
What about short texts?
Anthropic's own note says the watermark is only weakly detectable on very short outputs. Below roughly a hundred words there is often little to remove, and Unmarker will tell you so instead of rewriting for nothing.
What happens to my data?
Your text is sent over TLS, processed to produce the rewrite, and deleted from our servers within 24 hours. It is never used to train models. Waitlist emails are stored only to send you the launch email; see the privacy page.
When does it launch?
Early access opens in batches to the waitlist, first come first served, starting with the first 500 signups who get founder pricing. Everyone on the list receives exactly one email when their batch opens.
Be first in line when Unmarker opens.
Founder pricing for the first 500. One email at launch, nothing else.
Founder pricing for the first 500. No spam, one email at launch.