WMWATERMARK / CLEANER
BY @IVANAINAI ↗Read the guide ↓

Free, local, honest about the limits

Clean the marks
that can actually
be cleaned.

Find invisible characters in text and remove ordinary metadata from images. Your content stays inside your browser. Embedded SynthID and statistical text watermarks are a different problem, so this tool never pretends otherwise.

◆ No upload◆ No account◆ Open in your browser
BUILT BY @IVANAINAI · MORE FREE TOOLS SOON ↗

01 / CLEANER

Use it now.

CLEAN RESULTCLEAR
Your cleaned text appears here.

This removes hidden Unicode and formatting noise. It cannot remove a watermark encoded in word choice or token probabilities.

02 / FIELD GUIDE · UPDATED AUGUST 13, 2026

How AI watermarking actually works across Claude, ChatGPT and Gemini

The word “watermark” is covering three different technologies. Treating them as one is why so many removal guides are confidently wrong.

TYPE 01

Hidden characters

Zero-width characters and directional formatting live inside copied text. They are easy to inspect and remove, but they are not the same thing as a statistical model watermark.

EASY TO CLEAN
TYPE 02

File metadata

EXIF, XMP and signed C2PA provenance travel with a file. Re-exporting pixels can remove them, along with useful authorship and editing history.

REMOVABLE
TYPE 03

Embedded signals

SynthID can live in pixels, audio or token choices. It is designed to survive ordinary edits. A metadata cleaner cannot touch it.

NOT A FILE TAG

WHAT EACH PROVIDER MARKS

Based on current official documentation
PROVIDERTEXTIMAGES / FILESWHAT THIS TOOL DOES
CLAUDESupported models released from Aug 2, 2026 embed a watermark in generated text.Supported files can carry signed C2PA provenance metadata.Removes hidden Unicode and re-renders raster files. It cannot verify Claude’s statistical text mark.
CHATGPTOpenAI’s current provenance table does not list text. It says text support is a future goal.Supported images include C2PA metadata and SynthID.Removes the metadata layer from raster files, not pixel-level SynthID.
GEMINIGemini app and web text can carry SynthID through altered token probabilities.SynthID is embedded directly into generated images, video and audio.Cleans hidden characters and metadata only. It does not neutralize SynthID.

THE PRACTICAL WORKFLOW

Use the smallest intervention that solves the real problem.

If you are cleaning copy for a CMS, removing odd formatting may be enough. If you are protecting image privacy, re-exporting can remove location, device and provenance metadata. If the goal is to defeat embedded detection, a browser metadata tool cannot honestly promise that.

  1. 1
    Inspect before changing anything

    Paste the text or scan the image. Keep the original if provenance matters.

  2. 2
    Clean the removable layer

    Strip hidden formatting from text or re-render a raster image locally.

  3. 3
    Know what remains

    Statistical text marks and pixel-level SynthID are not ordinary metadata.

  4. 4
    Verify where a detector exists

    Use OpenAI Verify for supported OpenAI media or ask Gemini about supported Google media.

CLAUDE

Heavy rewriting can weaken detection. Nobody can prove a “one-click removal” yet.

Anthropic says supported Claude text watermarks may stop being detectable after heavy editing, paraphrasing, translation or mixing. It also says public detection details are still forthcoming. That makes any current “100% Claude watermark remover” claim impossible to verify independently.

OPENAI + GEMINI

A screenshot can remove metadata and still leave an embedded signal.

OpenAI says SynthID may remain when C2PA metadata is gone. Google says its image watermark is designed to survive cropping, filters and lossy compression. Re-saving a file and removing its pixel watermark are two separate claims.

PLEASE DON’T SKIP THIS

Cleaning metadata is not proof that content is human-made.

This tool is useful for privacy, broken formatting, file compatibility and understanding provenance layers. Removing machine-readable history can also erase useful attribution. Keep disclosure requirements, platform rules and local law in mind, especially for commercial, political, journalistic or regulated content.

PRIMARY SOURCES

01Anthropic

How Claude marks AI-generated content

↗02OpenAI

Provenance signals in OpenAI-generated content

↗03Google DeepMind

How SynthID works across text and media

↗