What Is the Claude Watermark?
On August 14, 2026, Anthropic announced that future Claude models will embed an invisible watermark in the text they write. The company says the change is meant to comply with the EU AI Act, and it links the change to the EU's transparency Code of Practice for marking AI-generated content. Models launched on or after August 2, 2026 are meant to carry the mark from launch, while earlier models get a transition period and are being brought in over the coming months.
The mark is not a visible label and not a string of special characters. Whenever Claude has several equally good words to choose from, a keyed method makes the choice instead of a plain random draw. Readers see nothing unusual, the method adds no extra tokens or cost, and Anthropic says it carries no information about the person or organisation using Claude. Anthropic also notes that other major providers who signed the same Code are adding watermarks of their own. The best primary source is Anthropic's announcement on the Claude text watermark.
There is no public text detector. In its September 2026 update, Anthropic described a detection API in private preview for eligible organisations such as regulators, law enforcement, media and fact-checkers, researchers, educators and EU civil society groups, plus enterprises that must verify marks for their own compliance. Anthropic plans to widen access over time, so check the announcement for the latest status.
Does Claude Insert Zero-Width or Invisible Characters?
Anthropic describes the watermark as word choice, not as hidden characters or metadata tags, so it does not depend on invisible Unicode. Many websites claim that AI watermarks are zero-width spaces. For Claude's text watermark, that is not how Anthropic says it works.
If you find invisible characters in a piece of AI text, they most likely came from somewhere else: a web page, a PDF, a word processor, a CMS or another tool in your workflow. They are still worth cleaning, which is what this page is built for. BleepingComputer's look at the watermark-remover market draws the same line: removing hidden characters works and can be checked, but it is not the watermark.
What This Claude Watermark Remover Actually Does
The tool is an invisible character checker and cleaner. It scans pasted text and handles the following:
- Zero-width and format characters: zero-width space (U+200B), word joiner (U+2060), byte order mark (U+FEFF), soft hyphen (U+00AD), filler characters such as the Hangul fillers and the Braille blank, and deprecated format characters.
- Joiners and direction marks: zero-width joiner and non-joiner, left-to-right and right-to-left marks and bidirectional control characters. In Safe mode the ones that are legitimate where they sit stay.
- Hidden Unicode: invisible math operators, variation selectors, Unicode tag characters and other unassigned ignorable characters.
- Unusual spaces: non-breaking space, thin space, narrow no-break space, ideographic space and other lookalikes, which become a normal space. Line and paragraph separators become a normal line break.
- Control characters: stray control codes. Tabs and line breaks are never touched.
- Em dashes: optional replace or remove, with spacing handled for you.
It does not rewrite your sentences, translate anything or change your meaning. That is also why it cannot remove a statistical watermark, which is explained further below. For single-purpose versions of the tool, see the Invisible Character Remover, the Zero-Width Space Remover and the Em Dash Remover.
How to Use the Claude Watermark Remover
- Paste your text into the input box at the top of this page.
- Choose your options. Leave Safe mode on for emoji and non-Latin languages, pick an em dash style if you want one, and decide whether to normalise unusual spaces.
- Click "Scan" if you only want a report. You will see each hidden character on the character map, a count by code point and what would happen to each one.
- Click "Clean Text" to remove the characters. The cleaned version appears in the second box along with a short summary.
- Copy or download the result and paste it into your document, CMS or code editor.
Why Hidden Characters Show Up in AI Text
Invisible characters are common in modern writing workflows, and they cause real problems. They break exact-match search, confuse spell checkers, produce odd line breaks, corrupt URLs and slugs, and can cause bugs when pasted into code or spreadsheets. The usual sources are:
- Copy and paste from web pages, chat interfaces, PDFs and word processors, which often add non-breaking spaces, soft hyphens and direction marks.
- Editors and CMS tools that insert special spaces or joiners while formatting.
- Multilingual text. The zero-width non-joiner and joiner are legitimate and necessary in Urdu, Persian, Hindi and other scripts, which is why Safe mode exists.
- Deliberate hiding. Security researchers have shown that variation selectors and tag characters can be used to smuggle hidden data inside ordinary-looking text. Scanning for them is a sensible precaution before you paste text into an important system.
Can Any Tool Remove Claude's Real Watermark?
Anthropic says light editing probably will not remove its text watermark, while rewriting every word will. In other words, the only way to change the pattern is to change the words themselves.
That has two practical consequences. First, character cleaners, including this one, cannot touch the statistical watermark. BleepingComputer reports that most watermark-remover claims cannot be independently verified, because there is no public detector, and that some tools only strip metadata. Be cautious about any site that promises to remove all of it with one click. Second, rewriting tools change your wording, which can alter meaning or introduce errors, so any output should be read carefully before you use it.
For publishers, the more useful question is quality. Google has said it rewards helpful, people-first content however it is produced, while its spam policies target low-quality mass-produced pages. Editing AI drafts with real expertise, facts and original insight is worth more than any cleaning step.
What a Watermark Does and Does Not Prove
Anthropic's own documentation is careful about the limits, and they matter if you are trying to judge a piece of text:
- A detected mark suggests Claude processed the text. It does not show Claude wrote it from scratch, because Claude may only have proofread, translated or summarised text from another source.
- A missing mark does not prove the text is human-written. It can be absent because the model predates marking, the passage is very short, the text was heavily edited or paraphrased, or the platform does not support it.
- Light proofreading of a person's own text leaves little for the mark to act on, so it may not be detectable.
- Code and strictly factual passages have fewer arbitrary word choices, so the mark is sparser there. Translations done by Claude are marked, because Claude chooses every word.
- The mark stays with the text when you copy it, and it is applied at the model level across Claude products and cloud partners.
The details are in Anthropic's support article, How Claude marks AI-generated content.
Files and Content Credentials (C2PA)
Files use a different mechanism from text. Claude attaches signed C2PA Content Credentials, which are metadata, to supported file types such as PNG and JPEG images, where Claude's file generation features are available. Because it is metadata, it can be lost when a file is converted, re-saved or screenshotted.
Anthropic offers a free Claude file checker that looks for these credentials in images, video and audio files. It does not check text. This page works only with plain text, not files.
Em Dashes in AI-Generated Writing
Many readers associate em dashes with AI writing. An em dash is not a watermark, just a punctuation style that some models use often and plenty of human writers use too. If you prefer a different rhythm, the tool can replace each em dash with a spaced hyphen or a comma, or turn it into a plain space. Dashes at the very start or end of a line are dropped cleanly. Because punctuation changes how a sentence reads, give the result a quick read-through. There is a dedicated Em Dash Remover with examples of each option.
Using AI Text Responsibly
Cleaning formatting artifacts is a normal editing step. It is not a licence to misrepresent who wrote something. Follow the rules of your school, employer, publisher or platform, and disclose AI assistance wherever it is required. This tool is meant for tidying text, not for hiding authorship where honesty matters.