A native macOS app that removes AI watermarks and provenance metadata from your text and files — for privacy and hygiene on content you own.
Paste text from Claude, ChatGPT, or Gemini, or drop in a document. Dewatermark strips the hidden signals those tools leave behind.
Modern AI tools mark their output in a few different ways. Dewatermark handles each:
| What | How it's marked | What Dewatermark does |
|---|---|---|
| Invisible characters | Zero-width spaces, hidden Unicode, exotic spaces | Clean — removes them, lossless |
| Statistical watermarks | The word choices themselves (how Claude/Gemini actually mark text) | Rewrite — rewords with a local model to reduce them |
| File metadata | C2PA "Content Credentials," AI tags in PDFs, Word docs, images | Clean — strips the metadata |
The one thing to understand: Inspect only checks for invisible characters. Claude and Gemini embed their watermark in the wording itself — invisible to a character scan. If your text came from an LLM, use Rewrite to reduce those statistical marks. That's the main event for AI text.
- Download the latest
Dewatermark-x.y.z.dmgfrom Releases and drag Dewatermark to Applications. - First launch: right-click the app → Open (it's unsigned, so macOS asks once). Or run:
xattr -d com.apple.quarantine /Applications/Dewatermark.app
- That's it — no Python install, no dependencies. Everything needed is bundled.
Paste text into the Text tab → Clean to strip invisible characters → Copy or Save As.
Paste text → Rewrite Text. Uses a local AI model (via Ollama) to reword the text and reduce statistical watermarks. The result is cleaned again afterwards.
To enable Rewrite:
# Install Ollama, then pull a model:
ollama pull llama3.2Dewatermark detects Ollama automatically once it's running. You can also point it at any OpenAI-compatible endpoint in Settings.
Drop a file into the File tab (or Choose File) → Clean. Works with PDF, Word, Markdown, HTML, SVG, PNG, JPEG, and text/code files. Writes a filename.cleaned.ext next to the original — your original is never touched.
- macOS 13 or later
- Nothing else for basic cleaning (Python 3 and all tools are bundled / auto-detected)
- Ollama only if you want the Rewrite feature
Bundled tools (included, no install needed): c2patool (C2PA manifest inspection) and exiftool (thorough PDF/document metadata stripping).
Text watermarks live in the wording itself, spread across every sentence. Two honest consequences:
- Removing them means rewording, not tidying. A light edit barely moves the signal. The Rewrite feature rewords substantially, which is what actually reduces it.
- Rewording changes the copy. The rewritten text is only as good as the local model doing the rewriting. For polished prose, that's a real trade-off.
So: use Clean (invisible characters, file metadata) when you want a lossless, verifiable result. Use Rewrite when the statistical mark is the concern and you accept a rewording pass.
What Dewatermark cannot do (no tool can, honestly):
- Remove pixel-level watermarks baked into AI-generated images (e.g. SynthID-media)
- Defeat "soft-bound" C2PA that re-links to an online manifest
- Guarantee any specific detector will pass — there's no public universal detector to test against
For your own content, this is a privacy and hygiene tool — not a way to pass AI-text off as human-written for fraud.
Built on the excellent watermarks-remover engine (vendored, pinned snapshot in Engine/).
swift build # build
swift test # test suite (requires Xcode)
python3 -m pytest # engine parity tests
scripts/fetch-tools.sh # fetch bundled tools (c2patool/exiftool)
scripts/vendor-engine.sh main # update the engine from upstream
scripts/build-app.sh # build the .app
scripts/make-dmg.sh 0.2.0 # package a DMGRelease a new version:
git tag v0.2.0 && git push origin v0.2.0 # CI builds and publishes the DMG + zipMIT. Engine and bundled tools retain their respective licenses (see Engine/watermarks-remover/LICENSE).