ChatGPT Watermark — What Can Actually Be Detected and Removed

A free, browser-based scanner for anyone who copies text out of ChatGPT — writers, editors, students, and developers — that checks the pasted content for hidden characters with no sign-up and no upload. It also explains honestly what OpenAI's text watermark is and why a character scanner cannot detect it.

Free · No Account · Browser-Based · Nothing Uploaded

Scan ChatGPT Text for Hidden Characters

Drop a plain-text file here to check and clean, or click to browse

.txt, .md, .csv, .json, .html, .css, .js, .ts, .py, .java, .go, .xml, .log

▶ Scan rules (4 groups enabled)

Copy text straight from your favorite AI chat window and paste it above to scan it for free.

What This Scanner Can Detect in ChatGPT Text

When text travels out of ChatGPT — copied from the web interface, exported, or pasted through an editor — it can carry invisible characters that no one typed. This scanner looks for those, character by character:

  • Zero-width characters — zero-width space (U+200B), non-joiner (U+200C), joiner (U+200D), and word joiner (U+2060). Invisible, but they break search matching, string comparison, and form validation.
  • Byte-order mark (U+FEFF) — a leftover marker at the start of some copied or exported text. It can trip up parsers, CSV imports, and command-line tools.
  • Non-standard spaces — non-breaking, narrow no-break, and ideographic spaces that look identical to a normal space but are different code points.
  • Soft hyphens (U+00AD) — invisible break hints that silently change how words are indexed and compared.
  • Bidirectional control characters — U+202A–U+202E and U+2066–U+2069, which can reverse visual order without changing the stored text.

About interface markup: attributes in ChatGPT's page source, such as data-start, are HTML-level artifacts of the interface — they do not survive a plain-text copy. Once text is copied as plain text, what remains are character-level artifacts like the ones listed above.

What Is the ChatGPT Text Watermark (textGrain)?

OpenAI has described a text watermarking system referred to as textGrain. It works very differently from the characters above: instead of inserting marks into the text, it biases which tokens the model chooses, so the statistical pattern of the words carries a signal.

  • It leaves no characters behind. A watermarked paragraph looks like an ordinary paragraph. There is nothing to select, delete, or highlight.
  • It needs OpenAI's own detector. Verification relies on a detection key and OpenAI's tooling — a third-party scanner cannot check a textGrain mark by looking at the text alone.
  • It is affected by editing. Paraphrasing, translating, or heavily rewriting the text changes the token pattern that carries the signal.

Straight answer: this scanner cannot detect textGrain, and neither can any other character-level tool. Detection of a statistical watermark requires the watermark's own detector and key. This page will not claim otherwise.

How to Clean Text Copied From ChatGPT

  1. Paste the text into the scanner above, or upload the exported file.
  2. Run the scan. Each finding is listed by category with a severity level and its exact position, so you can see whether a problem character sits inside a word, a URL, or a data cell.
  3. Choose the rule groups to apply and copy the cleaned result. For bulk cleanup of documents, the dedicated text cleaner follows the same rules.

Everything runs in your browser. Your pasted text is never uploaded, logged, or stored, and there is no account and no scan limit.

Why a Clean Result Is Not Proof of Authorship

Removing invisible characters makes text behave predictably in editors, databases, and search. It does not make the text human-written, and it does not remove a statistical watermark. If your question is "was this written by AI?", a character scanner is the wrong instrument — and a tool that answers "yes" or "no" with confidence is guessing.

FAQ

Can this tool detect a ChatGPT watermark?

No, not if the watermark is OpenAI's textGrain. This scanner finds character-level traces in copied text — zero-width characters, byte-order marks, odd spaces, soft hyphens, and BiDi controls — and it cannot detect a statistical watermark that leaves no characters behind.

What is the ChatGPT text watermark called?

It is referred to as textGrain. OpenAI's text watermarking approach biases which tokens the model chooses, so the signal lives in the statistical pattern of the words rather than in any character you can select or delete.

Can I remove a ChatGPT watermark by editing the text?

You can remove hidden characters, but not a statistical watermark by deleting characters. A cleanup pass strips copy artifacts; a textGrain signal is carried by the word pattern, so heavy rewriting or paraphrasing is what affects it — not character removal.

Why does text copied from ChatGPT contain invisible characters?

Because copying from a rendered web interface can carry formatting artifacts along with the words. Zero-width spaces, non-breaking spaces, byte-order marks, and soft hyphens are the common ones, and none of them are visible in a normal editor.

Is data-start a watermark?

No. data-start is an HTML attribute in ChatGPT's page markup, not a character inside the text, and it does not survive a plain-text copy. What does survive are the character-level artifacts this scanner looks for.

Does the scanner tell me whether an essay was written by AI?

No. A clean or flagged result says nothing reliable about authorship. AI-generated text often contains no hidden characters at all, and human text pasted through a web editor sometimes does, so a character scan is not an AI detector.

Is my pasted ChatGPT text uploaded anywhere?

No. The entire scan runs as client-side JavaScript in your browser tab, so your text and any uploaded file are never sent to a server, never logged, and never stored after you close the page.

Is this scanner free and does it need an account?

It is free and requires no account, no login, and no payment details. There is no scan counter and no daily cap, because the work happens on your own device rather than on a metered server.

Does it work on exported files, not just pasted text?

Yes. You can upload a plain-text file — .txt, .md, .csv, .json, .html, .css, .js, or a log file — and the same character-level scan runs on its contents, locally in the browser.

What should I do if the scan finds nothing?

Nothing needs cleaning. An empty result means none of the character-level signals were present in the text you supplied — it does not mean the text is human-written or watermark-free.

Related Guides

← Back to AI Watermark Detector

Related Tools