Simple Unmark guide
Does ChatGPT Watermark Text? OpenAI's textGrain Explained
- Published
- Last updated
- Sources checked
- Primary references: 3
Short answer
Yes, for eligible ChatGPT and Codex text in the European Union, rolling out from 5 October 2026. OpenAI calls the method textGrain: an invisible statistical signal in word choices, with no hidden characters. It is not a global default, and API watermarking is opt-in. OpenAI's own test shows changing words weakens it. There is no public checker.
On 5 October 2026 OpenAI published how it will mark text under the EU AI Act. The answer is no longer "text is not watermarked." It is also not "every ChatGPT reply in the world now carries a mark."
Choose the privacy boundary before every rewrite.
Private is the lowest-cost path and stores no text in our application database. Confidential verifies Google Confidential Space in your browser and encrypts to a key the workload created for that one request, so our web backend receives no text — though DeepInfra still performs the rewrite. Confidential AI, with the model inside a confidential GPU, is coming soon.
Compare the three modes and their limitsWhat OpenAI announced
Starting 5 October 2026, API customers worldwide can opt in to text watermarking for select models. It stays off by default in the API. Over the coming weeks, an invisible watermark is added to eligible ChatGPT and Codex text in the European Union, on every plan. OpenAI is not making text watermarking a global default at launch.
Images and audio stay on the older stack: C2PA Content Credentials plus SynthID, with public verification. That stack is not the text watermark.
Is ChatGPT text watermarked outside the EU?
OpenAI’s announcement specifies eligible ChatGPT and Codex text in the European Union; it does not announce a worldwide default. Check the dated provider status tracker for the scope we last verified.
Text from apps built on OpenAI's API is a separate case. Developers anywhere can opt in to watermarking for select models, and it stays off unless they do, so you usually cannot tell from the outside whether a given app has it on.
textGrain is word choice, not characters
OpenAI names the method textGrain. It adds an invisible statistical signal to the model's word choices. A detector looks for that signal and needs a secret key. Nothing is inserted as a hidden character.
Weakening it therefore means changing the words, whether by editing or rewriting. OpenAI's own numbers show how quickly synonym replacement lowers detection, and a character scan does nothing to it.
Editing can weaken the watermark. In an evaluation of 400-token passages, replacing 10% of words with synonyms reduced detection from about 92% to 66%. Replacing 25% of words reduced it to 17%.
Invisible characters are still a different thing
Copied text still picks up zero-width characters from pages, editors and chat clients. Those are not textGrain. Finding one does not mean ChatGPT watermarked the passage, and deleting them does not remove the word-choice signal.
The same goes for claims that ChatGPT marks its text with special spaces or em dashes. OpenAI's documented text watermark is textGrain, and OpenAI does not describe any character as part of it. An em dash is visible punctuation and tells you nothing about a watermark.
Hidden characters are still worth removing when they break search, diffs, validation or pasted code. The invisible character remover does that locally and reports each code point by name.
What a detection does not prove
OpenAI is direct about the limits. A watermark can indicate that an OpenAI system generated or processed part of a passage. It does not say who used the system, who owns the text, or whether the content is accurate. If no watermark is detected, that does not prove a person wrote it. The text may be short, edited, translated, from an unsupported model, or from before the rollout.
There is no public checker. Detector access opens first to approved researchers and expert organizations. Until that changes, nobody outside that group can measure whether a specific passage still carries textGrain.
Sources
- Our approach to EU text provenance rules
OpenAI (5 October 2026) · First-party source
- textGrain: Entropy-Calibrated Watermarking for Language Model Text
OpenAI technical report (5 October 2026) · First-party source
- UTR #36: Unicode Security Considerations
Unicode Consortium · First-party source
Published by Simple Unmark. Our Methodology explains how we source claims, what we test, and what we do not claim.
Apply this to your own text.
Guests get three free cleans without an account. Choose Private or browser-verified Confidential processing before each rewrite. Deterministic character inspection is free and runs only in your browser.
Keep reading
- How AI Text Watermarks WorkHow hidden-character marks and statistical token watermarks differ, how detection works, and why removal claims should always be bounded.
- What Is SynthID Text?SynthID Text explained: how Google's text watermark is applied during generation, how detection works, what weakens it, and how it differs from image watermarking.
- Hidden Unicode vs Statistical Text WatermarksThe two mechanisms confused most often: invisible Unicode characters versus keyed token-choice watermarks. What each one is, what removes it, and what neither proves.
- AI Watermarks vs AI Detectors vs Provenance MetadataThree different things routinely confused: keyed watermarks added at generation, statistical AI detectors that guess, and C2PA metadata attached to files.
