Claude Watermark Remover
Paste your text to reveal and strip every invisible character in it — free, instant, nothing uploaded.
What the scanner finds
Text picked up from a PDF, a word processor, a rendered web page or a chat interface routinely carries characters that render as nothing at all. Most are harmless residue. Some are not.
Unicode tag characters are the ones worth knowing about. They mirror the ASCII range and display as absolutely nothing, which makes them the standard vehicle for hiding instructions inside text that someone will paste into an AI system. The scanner decodes them rather than just counting them, so you can read what a hidden payload actually says. Bidirectional controls reorder how text displays without changing what it contains — in source code that is the Trojan Source attack, where a reviewer and a compiler read different programs.
The rest are mundane and far more common: zero-width characters, non-breaking and narrow spaces, soft hyphens, and Cyrillic or Greek letters standing in for Latin ones. They matter because two strings that look identical stop being equal — CSV imports silently duplicate records, lookup keys never match, search returns nothing for a term plainly on the page, and diffs mark lines as changed when nothing visible changed.
The full reference covers every block, its code point range, and when finding it is genuinely a problem.
C2PA inspector for images
Images use a different mechanism entirely, and this one is readable. Drop a JPEG, PNG or WebP to see its signed provenance manifest: what produced it, what was done to it, and who signed the claim.
Drop a JPEG, PNG or WebP here
The file is parsed in this tab. Nothing is uploaded.
Frequently asked questions
- Can you remove Claude's watermark?
- Not with any tool available today, including this one. Claude's text watermark is a statistical bias in which words the model chooses, not a character or a marker embedded in the file. There is nothing in the text to locate and delete. Tools that advertise removal strip invisible Unicode characters, which are unrelated to the watermark and were never carrying it.
- Does Claude watermark its text?
- Yes. Since 2 August 2026, Anthropic embeds an imperceptible watermark in text generated by Claude models launched on or after that date. It works by subtly biasing word choice, so the signal is spread across the vocabulary of a passage rather than stored in any single place. It survives copy and paste because it is the words themselves.
- How do Claude watermark removers actually work?
- Almost all of them remove zero-width Unicode characters, narrow spaces and homoglyphs, then return the text. That operation is real, but it has nothing to do with Claude's watermark. The text comes back looking processed while the statistical signature is completely untouched.
- Can I check whether text carries Claude's watermark?
- No. Anthropic has not published the detection method, so no third party can compute it. Any site claiming to detect Claude's watermark in text is guessing, usually from surface style cues like em-dash frequency. This also means removal tools cannot verify their own results.
- Will removing invisible characters remove the watermark?
- No. It removes invisible characters, which is worth doing for its own reasons: they break CSV imports, database keys, deduplication, search and diffs, and Unicode tag characters can hide instructions aimed at AI systems. None of that touches the watermark.
- Does Claude watermark images too?
- Files use a completely different and much more visible mechanism. Supported file types get a C2PA manifest: signed provenance metadata following an open standard. Unlike the text watermark, anyone can read it, and the inspector on this page does. It is also fragile — a screenshot or a format conversion destroys it entirely.
- Does paraphrasing remove the watermark?
- Anthropic documents that heavily paraphrased text may carry no detectable signal, alongside other limitations: no signal on models predating the rollout, and none on passages too short to carry one. In practice that means a detection result is weak evidence in both directions, which matters most for the people being assessed by it.
- Is the watermark in Claude Code and the API as well?
- Yes. Marking applies across the API, Claude, Claude Code, Claude Cowork and Claude Tag, and through cloud distribution on AWS, Google Cloud and Microsoft Foundry. It is applied worldwide rather than only in the EU.
- Why does this watermark exist at all?
- The EU AI Act requires providers of generative systems to mark synthetic output in a machine-readable way. Anthropic applied it globally rather than maintaining a separate European build. It is a transparency obligation on the provider, not an enforcement tool designed for schools or employers.
- I was accused of using AI but I didn't. What should I do?
- Establish first which of three very different things flagged you: a commercial AI detector, an actual provider watermark check, or somebody's judgement. Commercial detectors have significant false positive rates, particularly for people writing in a second language. Drafting history, version timelines and process artefacts are the evidence that holds up. Stripping invisible characters does nothing and looks bad.
Read further
How to remove the Claude watermark
The direct answer, mechanism by mechanism — including the one case where removal genuinely is possible.
What a mark actually means
A detection says content was processed by a model. That is a much weaker claim than “written by AI”, and the gap is where people get hurt.
Accused of using AI when you didn’t
What to do if a detector flagged your work. Which evidence holds up, and which arguments waste your time.
How Claude marks its output
The text watermark, C2PA file signing, covered surfaces and documented limits — from Anthropic’s own material.