Free ChatGPT Watermark Remover & Checker for Text & Code
Paste ChatGPT-generated text or code, see every hidden watermark character exposed in an x-ray view, and copy a clean version in one click. Free, online, no sign-up, and nothing you paste ever leaves your browser.
More: what was found, x-ray view and options
X-ray view
Your text with every hidden character exposed as a labelled chip. Hover a chip for its Unicode name.
What was found
| Code point | Character | Count | Action |
|---|
Does ChatGPT put a watermark in its text?
Not yet. OpenAI hasn't shipped a statistical text watermark for ChatGPT as of this page's verification date, even though it signed the EU transparency code of practice that requires one, so this is a "when," not an "if." What copied ChatGPT text usually does carry is invisible Unicode debris left over from the chat interface's own formatting, unrelated to watermarking but just as capable of breaking a CSV import or a code diff, and that's exactly what the checker below finds.
Last verified: 11 August 2026
ChatGPT is the vendor where the gap between what people believe and what is deployed is widest. There is a persistent belief that ChatGPT hides zero-width characters in its answers as a covert signature. It does not, and the checker above will demonstrate that on your own text in a couple of seconds.
Current status, plainly
Three separate things are often collapsed into one claim. Keeping them apart is the whole answer:
| Mechanism | Status for OpenAI |
|---|---|
| Statistical watermark in generated text | Not deployed at the verification date above |
| Hidden Unicode characters as a signature | Not used, and never has been |
| C2PA provenance metadata on generated images | Applied, and carried in the image file |
| Commitment to future text marking | Signed the EU transparency code of practice |
OpenAI has discussed text watermarking publicly for years, and research into token-level schemes predates the current regulatory push by a long way. The reasons for not shipping one are not secret and are mostly about consequences rather than feasibility: a watermark that only one major provider applies pushes users toward providers that do not apply one, and it puts non-native English speakers, who lean on assistive tools more heavily, at greater risk of being flagged.
The trajectory
OpenAI signed the EU code of practice attached to Article 50(2) of the AI Act, the same instrument Google satisfies with SynthID and under which Anthropic confirmed watermarking for newer Claude models in August 2026. That is a public commitment to machine-readable marking of synthetic output, which makes some form of text marking a reasonable expectation over the coming period rather than speculation.
What it is not is a description of today. If you are trying to work out whether a specific piece of text came out of ChatGPT, there is currently no provider-side signal in it to find. Style-based AI detectors are not a substitute: they produce false positives on ordinary human writing at rates that make them unsafe for any consequential decision.
Images are the exception. C2PA metadata on a generated image is real, checkable provenance, and it travels in the file. It is also fragile in ordinary use: screenshotting, re-encoding, or uploading through a platform that strips metadata removes it. Absence of C2PA proves nothing about origin.
Why people find hidden characters in ChatGPT output anyway
They do find them, and the characters are real. The cause is the copy, not the model.
The chat interface renders a response as HTML with proper typography. That rendering can include no-break spaces to keep units attached to numbers, narrow no-break spaces in certain locales, and thin spaces around punctuation. When you select the rendered text and copy it, you copy the typography with it. Paste into a plain text editor and those characters are still there, invisible, and now sitting in your document.
Markdown makes a second contribution. Copying a formatted answer often brings across non-breaking spaces used for indentation in nested lists, and occasionally a zero-width space where the renderer needed a break opportunity.
Neither is a watermark. Both are worth cleaning, because both break string matching and data imports. That is what the tool above is for, and the hidden characters guide covers where else they come from.
What this tool can and cannot do for ChatGPT text
Can
- Prove, on your own text, that there is no invisible signature hiding in it, the x-ray view shows every code point that is not an ordinary character.
- Strip the typographic debris that copying from the chat interface leaves behind.
- Normalise spaces and line separators so the text behaves predictably in code, CSV files and forms.
Cannot
- Tell you whether text was written by ChatGPT. No provider-side signal exists in the text to detect, and we do not guess from writing style.
- Read C2PA metadata from an image file. This tool works on pasted text only.
- Predict when or whether OpenAI ships a text watermark. When that changes, this page changes, and the verification date moves with it.
Questions about ChatGPT and watermarking
Does ChatGPT hide invisible characters in its responses?
No. It does not embed zero-width characters as a signature. Invisible characters sometimes appear in text copied out of the interface because the page renders typographic spaces and the copy takes them along, but that happens with any well-typeset web page.
Can anyone prove a piece of text came from ChatGPT?
Not from the text alone. There is no deployed watermark to detect, and style-based detectors are unreliable enough that they should not be used for consequential decisions. Provenance for images is a different matter, because C2PA metadata is checkable when it survives.
Why do people say OpenAI has a watermark ready but has not released it?
Because research into token-level text watermarking is well established and OpenAI has discussed it publicly. Having a viable method is not the same as deploying it, and the trade-offs, competitive pressure and the risk of false accusations against non-native speakers, are why deployment has lagged the research.
What is the difference between C2PA and a watermark?
C2PA is signed metadata attached to a file describing how it was made; it is removed when the file is re-encoded or the content is copied out. A watermark is embedded in the content itself and survives ordinary copying. OpenAI applies the former to images and, at the verification date, neither to text.
If OpenAI adds a text watermark later, will this tool remove it?
No. Any realistic text watermark will be statistical, embedded in token selection, and a character cleaner cannot touch that class of signal. We will update this page to describe what was deployed and what it means in practice.
Primary sources
- OpenAI on provenance and C2PA - the company's own statement of what it marks and how.
- The C2PA specification body - the standard behind image provenance metadata.
- The general-purpose AI code of practice at the European Commission.
Related reading
- Claude's confirmed watermark - how the same rumour resolved into a real deployment next door.
- Microsoft Copilot and what its output carries - a product built partly on OpenAI models.
- Hidden characters in text - the real reason invisible characters show up in your paste.
- C2PA explained: what content credentials do and do not prove.