Does ChatGPT watermark its text?

Yes, in the European Union, and elsewhere only when a developer switches it on. OpenAI calls it textGrain: an invisible pattern in the words ChatGPT picks. Nobody outside OpenAI’s approved testers can check text for it yet. What anyone can check today is an image, because pictures made in ChatGPT carry signed Content Credentials. Drop one in below.

✓ Free✓ No sign-up🔒 Deleted after readingReads Content Credentials

🔐

Drop or paste an image to read its Content Credentials

Images made in ChatGPT are signed by OpenAI. The reader shows who signed a file, and runs the AI detector when there is no signature.

Choose an image →

This reads image files, not pasted text. Nobody outside the AI companies can check text for their watermarks yet.

Where the text watermark is switched on

ChatGPT in the EUOn for every plan, rolling out over the weeks after OpenAI’s announcement on 5 October 2026. Codex in the EUOn, on the same schedule. The API, worldwideOff by default. Developers can turn it on for some models. ChatGPT elsewhereOff. Text from the US, the UK or anywhere outside the EU carries no watermark unless the app that produced it opted in through the API.

The trigger is the EU AI Act, whose transparency rules have required AI companies since 2 August 2026 to mark AI-generated content in a way machines can detect.

How textGrain works

A language model writes one token at a time, and at each step several next words are plausible. textGrain uses a secret key to steer which of those plausible words gets picked, so the text carries a statistical pattern that reads normally to a person. A detector holding the same key can measure the pattern from the text alone.

There is no hidden character to find and no visible change, and copying or pasting keeps the pattern. OpenAI published a technical report on the method, written with researchers from the University of Pennsylvania and Yale, and says it found no meaningful drop in answer quality.

Who can check ChatGPT text

Not the public, and not us. OpenAI’s detector is open only to approved researchers and expert organisations for now, and its public verification page checks images and audio, not text.

That also means AI text checkers cannot read this watermark. They score writing style, which is a guess, and none of them holds OpenAI’s key. A tool that says it found the ChatGPT watermark in a pasted paragraph is not reading it.

How reliable the watermark is

About 400 tokens (roughly 300 words)About 95% detected, in OpenAI’s tests at a 1% false-positive rate. About 200 tokensAbout 80% detected. 10% of words swapped for synonymsDetection fell from about 92% to 66%. 25% swappedDetection fell to about 17%. Maths, short answers, translationsHarder to detect, because the model has less freedom in its word choices or the words were replaced.

These are OpenAI’s own figures from its October 2026 report.

What a detected mark proves, and what it does not

A detected mark means ChatGPT or Codex probably produced much of the passage. It does not say who asked, which account, what the prompt was, or how much a person rewrote afterwards.
No mark proves nothing. The text may come from outside the EU, from another AI, from an older conversation, or from someone who edited or translated it. A missing watermark is not evidence that a person wrote it.

ChatGPT images are a different story

Pictures generated in ChatGPT carry Content Credentials (C2PA): a record signed by OpenAI and stored in the file. Unlike the text watermark, anyone can read it, and the reader at the top of this page does. A screenshot or a re-upload to most social apps strips it, and then only the pixels are left to judge.

For the visual tells of ChatGPT’s images, see the ChatGPT image detector guide. For how image watermarks and credentials compare across generators, see AI watermarks and Content Credentials.

Frequently asked questions

Does ChatGPT watermark its text?

Yes, in the EU. Since October 2026 OpenAI adds an invisible watermark called textGrain to eligible ChatGPT and Codex text there. Outside the EU it is off, and API developers can choose to turn it on.

Can I check if text has a ChatGPT watermark?

Not yet. OpenAI’s detector is limited to approved researchers and expert organisations, and its public verification page does not check text. AI text checkers guess from writing style and cannot read the watermark.

Does copying and pasting remove the ChatGPT watermark?

No. The pattern is in the choice of words, so it survives copying. Heavy rewording weakens it: in OpenAI’s tests, swapping a quarter of the words for synonyms cut detection to about 17%.

Does the watermark show who wrote the text?

No. It indicates that ChatGPT probably produced the passage. It does not identify the user, the account or the prompt, and it cannot show how much a person edited.

Do ChatGPT images have a watermark?

Images made in ChatGPT carry Content Credentials signed by OpenAI, stored in the file. Upload one on this page to read them. Screenshots and most social apps remove them.