Skip to tool

AI watermark remover: see each hidden character before it goes.

Paste text from ChatGPT, Claude, Gemini or anywhere else. You'll see every zero-width space, odd space and tag character it carries, then copy a clean version. It's free, needs no account, and your text stays on your device.

AI text checker

Checked in your browser. Your text isn't sent anywhere.0 characters

The same privacy for your files, 2 GB free.

Most cloud drives can read what you upload. Vaulternal encrypts each file on your device before it leaves, so we can't.

No credit card. Works with Google, Microsoft or email login.

Get 2 GB free

Private by design

What makes this cleaner private

Many watermark removers send your text to a server, and some pass it to another AI model to rewrite. This one runs entirely in your browser.

Your text stays on your device

Detection and cleaning run in JavaScript on this page, with no upload and no server-side processing. You can confirm it in your browser's developer tools: pasting text sends no request.

See every character before it goes

Hidden characters are highlighted in place with their Unicode code and counted by type, so you know exactly what will change. Every removal shows up in the counts.

Keeps the characters your text needs

Emoji sequences, subdivision flags such as England's and Scotland's, and the joiners that Persian and Indic scripts depend on stay intact. Only stray characters are removed.

Clear about its limits

Statistical watermarks can't be removed by deleting characters. The section below explains why, and why no tool can honestly promise text that detectors miss.

What it cleans

What this AI watermark remover finds

Characters that look like nothing on screen but travel with copied text.

Zero-width characters

Zero-width spaces (U+200B), joiners, word joiners and byte-order marks take up no room at all. They split words for search, break code and URLs, and survive copy and paste.

Unusual spaces

No-break spaces (U+00A0) and narrow no-break spaces (U+202F) look like ordinary spaces. In April 2025, users found ChatGPT's o3 and o4-mini models adding narrow no-break spaces to answers. The tool turns them into regular spaces.

Unicode tag characters

Characters from U+E0000 to U+E007F mirror ASCII but render as nothing, so they can carry a whole hidden sentence. That makes them a tool for hiding instructions aimed at AI assistants. The checker decodes and shows that text, and keeps the tag sequences inside subdivision flags such as England's.

Variation selector runs

A single selector after an emoji picks its style and is kept. Long runs of them can encode hidden bytes, a trick known as emoji smuggling, and those are removed.

Text direction controls

Bidirectional overrides and isolates (U+202A to U+202E, U+2066 to U+2069) can make text display in a different order than it's stored. This is the trick behind Trojan Source attacks on code.

Control characters

Stray control codes, interlinear annotation marks and Unicode line separators break forms, spreadsheets and CMS editors. Line separators become normal line breaks.

Know the limits

What no character cleaner can remove

Google's SynthID-Text, Anthropic's watermark for Claude and OpenAI's textGrain don't add characters to your text. As far as their makers have described them, they work through word choice: the model is nudged toward certain words, and the pattern only shows up statistically across a passage.

Deleting characters doesn't touch that pattern, so this tool leaves it exactly as it was. Heavy rewriting or translation generally weakens it, and because the detectors are private, no tool can honestly promise text that passes them. We don't make that promise.

If your school, employer or publisher asks you to label AI-assisted writing, cleaning hidden characters doesn't change that.

By model

ChatGPT, Claude and Gemini: what each one puts in your text

Dated facts, so you can see what this tool helps with and what it leaves alone.

ChatGPT

  • April 2025

    Users found answers from the o3 and o4-mini models containing narrow no-break spaces (U+202F). OpenAI attributed the characters to a side effect of training and said they weren't a watermark.

  • October 2026

    OpenAI announced textGrain, a statistical watermark in word choice, rolling out to ChatGPT and Codex users in the EU, with an opt-in for API customers. Its detector is open only to approved researchers.

This tool removes narrow no-break spaces and other stray characters from ChatGPT text. It can't affect textGrain.

Claude

  • August 2026

    Anthropic announced that supported Claude models watermark the text they write, in line with the EU AI Act's transparency rules. It describes the mark as woven into the wording, and detection is in private preview.

Expect a clean result for Claude text, because the mark lives in the words. Hidden characters you do find usually came from an editor, a web page or a chat app along the way.

Gemini

  • October 2024

    Google DeepMind published SynthID-Text in Nature and released it as open source. It runs in Gemini and works by adjusting word-choice probabilities.

  • Worth knowing

    Searches for a Gemini watermark remover often mean the visible sparkle logo on Gemini images. This tool works on text only.

Hidden characters in Gemini text usually come from copying out of a web page or app. SynthID-Text stays in place.

Vaulternal isn't affiliated with OpenAI, Anthropic or Google. ChatGPT, Claude and Gemini are trademarks of their respective owners.

How it works

Paste, review, copy

1. Paste your text

Paste any text you've copied. The check starts as soon as you paste, on this device.

2. Review what's hidden

Each hidden character is highlighted with its Unicode code and counted by type. Switch off any type you want to keep.

3. Copy the clean version

Copy the cleaned text or download it as a .txt file. What you see in the clean text box is exactly what you get.

Why it matters

When hidden characters cause real problems

Code that won't run

A no-break space or zero-width character copied from an AI answer can break a variable name, a YAML file or a shell command, and the error message rarely points at it.

Publishing and CMS work

Narrow no-break spaces stop lines wrapping where you expect and throw off word counts. Cleaning text before it goes into WordPress, Google Docs or an email saves fixing the layout later.

Text you hand to AI agents

Instructions hidden in tag characters can ride along in a document, web page or email you ask an assistant to summarise. Checking the text first shows you anything planted in it.

Forms, spreadsheets and databases

Invisible characters make two identical-looking values different, so lookups fail, duplicates appear and validation rejects input that looks fine.

FAQ

Questions about AI text watermarks

  • In April 2025, people noticed ChatGPT's o3 and o4-mini models adding narrow no-break spaces, which OpenAI attributed to a side effect of training and said wasn't a watermark. Separately, in October 2026, OpenAI announced textGrain, a statistical watermark in word choice, rolling out to ChatGPT and Codex users in the EU. This tool removes the characters and can't affect textGrain.

From the team behind Vaulternal

Your text is clean. Are your files private?

This page checked your text without sending it anywhere. Most cloud drives work the other way: they hold the keys and can read what you upload. Vaulternal encrypts each file on your device first, so we can't read it. Start with 2 GB of encrypted storage free, and when you need to, let a trusted person reach a specific file under conditions you set in advance, such as a long trip or a hospital stay.

Start free with 2 GB

No credit card. Works with Google, Microsoft or email login.