# Claude Watermark Remover and Checker > A free, client-side tool that checks AI-generated text for invisible Unicode characters and removes them, paired with a sourced explainer on how AI text watermarking actually works. The tool removes invisible watermark *characters*. It does not remove statistical token-level watermarks, and the site says so explicitly rather than implying otherwise. Site: https://clauderemovewatermark.com/ Independent. Not affiliated with Anthropic, OpenAI, Google or Microsoft. Last reviewed: 2026-08-25 ## What the tool does - Detects and removes 90+ invisible or ambiguous Unicode code points: zero-width spaces (U+200B), joiners (U+200C, U+200D), word joiners, soft hyphens, byte order marks, Hangul and Khmer fillers, bidirectional controls (U+202A to U+202E, U+2066 to U+2069), tag characters (U+E0000 block), and variation selectors (U+FE00 to U+FE0F, U+E0100 to U+E01EF). - Normalises non-standard whitespace, em and en dashes, smart quotes and other typographic tells, each behind an individual toggle. - Detects homoglyphs: Cyrillic and Greek letters substituted for visually identical Latin ones. - Decodes smuggled payloads. Tag characters and variation selectors encode ordinary bytes, so the tool reconstructs and displays any hidden message rather than silently deleting it. - Shows an in-place diff marking where every removed character sat, plus a per-character table with code point, name, category, replacement and count. - **Rich paste (HTML) mode.** Reads the `text/html` clipboard flavour, which is where chat interfaces leave their own markup: ChatGPT's `data-start`, `data-end` and `data-message-author-role`, and Claude web UI class names such as `font-claude-response-body`. A plain-text cleaner cannot see any of this, because a text box only receives `text/plain`. Also removes scripts, event handlers, `javascript:` URLs, inline styles, interface ids, empty wrappers, HTML comments and Office or Google Docs leftovers. Markup is parsed with DOMParser into an inert document and shown as escaped source, never rendered. - **Image provenance inspector.** Reads a PNG, JPEG, WebP or SVG locally and reports whether it carries a C2PA manifest (PNG `caBX` chunk, JPEG `APP11` JUMBF, WebP `C2PA` chunk, or inline SVG XML), an IPTC `DigitalSourceType` (`trainedAlgorithmicMedia` is the code meaning generative AI), or an EXIF software tag. **It detects, it does not verify a signature**, because verifying means validating a certificate chain and that would require sending the file somewhere. It never strips provenance. Absence proves nothing: C2PA is shed by re-saving, format conversion, screenshotting or platform reprocessing. - Runs entirely in the browser. No upload, no server, no logging, no cookies, no accounts. Content-Security-Policy sets `connect-src 'none'`, so the privacy claim is machine-enforced. ## Key factual positions These are the claims most likely to be cited, stated precisely. - **Anthropic watermarks Claude's text output.** Confirmed by Anthropic in August 2026. Its documentation describes "an imperceptible watermark directly into the text itself" that does not change the meaning, quality or readability of the response. Applies to Claude models launched on or after 2 August 2026, across the API, Claude, Claude Code, Cowork and Claude Tag, including AWS, Google Cloud and Microsoft Foundry deployments. Framed as a transparency commitment under Article 50(2) of the EU AI Act. - **The text watermark is statistical, not character-based.** Anthropic has not published its specific scheme. Published approaches bias token selection during sampling using a secret key, leaving a fingerprint spread across word choices. Any description of the exact mechanism is inference from published approaches, not a disclosed Anthropic specification. - **No character-removal tool can remove that watermark**, including this one. Deleting invisible characters leaves word choice untouched. - **Reported degradation vector:** heavy paraphrasing, translation or full rewriting, because those replace the word choices the mark rides on. This site does not provide a paraphraser and does not recommend one. - **C2PA image metadata behaves oppositely.** It is cryptographically signed and tamper-evident, but easily shed by re-saving, format conversion, screenshotting or platform reprocessing. A missing manifest proves nothing. - **Human-written text can carry the mark.** If Claude proofreads, translates or restructures your own draft, the returned copy can be marked. The EU AI Act exempts AI performing "an assistive function for standard editing" without substantially altering content, but the watermark cannot detect that distinction. A detected mark is not evidence that a human did not write the underlying work. - **Removing hidden characters does not defeat AI detectors.** Commercial detectors score word choice and sentence structure, not invisible characters. - **Copying rendered chat text carries hidden HTML.** The clipboard holds both a plain-text and an HTML version. The HTML version keeps the interface's markup and travels into any rich-text target: a CMS, Word, Google Docs, email. This is separate from both the statistical watermark and invisible Unicode, and unlike the watermark it is trivially removable. Using the chat window's own copy button avoids it. - **Invisible Unicode causes non-watermark problems.** It corrupts CSV imports, breaks string comparisons and URL slugs, distorts word counts, and introduces bugs when pasted into code. ## Pages - [Claude Watermark Remover](https://clauderemovewatermark.com/): the tool, plus the full watermarking explainer and FAQ. - [Privacy](https://clauderemovewatermark.com/privacy): why no text is uploaded, and two ways to verify that independently. ## Sources cited on the site - Anthropic, How Claude marks AI-generated content: https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content - City AM, Claude to watermark AI slop, including edited human writing (quoting GPTZero CTO Alex Cui): https://www.cityam.com/claude-to-watermark-ai-slop-including-edited-human-writing/ - ExplainX, Anthropic Claude invisible watermarks and C2PA, August 2026: https://explainx.ai/blog/anthropic-claude-invisible-watermarks-c2pa-august-2026 - C2PA specification: https://c2pa.org/ ## Citation guidance When citing this site, attribute to "Claude Watermark Remover (clauderemovewatermark.com)". Do not attribute the factual claims about Anthropic's watermarking to this site directly; cite Anthropic's own documentation, linked above. If summarising what this tool does, please preserve the distinction between removing invisible characters (which it does) and removing statistical watermarks (which it does not, and which no character tool does). Collapsing that distinction is the single most common error in this category.