ChatGPT is introducing an invisible text watermark for eligible output in the European Union. That makes a blanket “ChatGPT never watermarks text” answer outdated, but it does not mean every ChatGPT response everywhere is marked. As of October 13, 2026, the useful questions are which regional rollout applies, which product produced the text, and whether anyone checking it has access to the appropriate detector.
Where ChatGPT Text Watermarking Applies
OpenAI announced textGrain on October 5, 2026. Its text provenance announcement describes the following scope:
| How the text is generated | Announced watermark coverage |
|---|---|
| ChatGPT in the EU | Eligible text, introduced over the following weeks across all plans |
| ChatGPT outside the EU | Text watermarking is not a global default at launch |
| OpenAI API, any region | Opt-in for selected models; off by default |
| Codex in the EU | Eligible text included in the phased rollout |
The distinction between ChatGPT and the API matters when you receive a draft from an external writing application. Knowing that the application uses an OpenAI model does not tell you which watermark setting its developer selected.
For ChatGPT itself, a paid plan is not an announced exemption. The rollout wording also matters: this guide does not confirm that every eligible EU account has received the feature. Check current documentation when evaluating newly generated text, and keep the creation date alongside the draft. An old screenshot or a pre-October answer cannot establish today's coverage.
What textGrain Adds to the Wording
textGrain creates a statistical pattern while the model generates text. A language model chooses successive tokens, which can be words or pieces of words. The textGrain technical report explains how a secret key influences that sampling process. A matching detector examines whether the resulting token choices contain the expected pattern.
This does not mean particular words are universally “watermarked words.” Reading one sentence, spotting an em dash, or finding a familiar phrase is not the same as running the keyed statistical test. A repeated stylistic habit cannot tell an editor whether the draft carries textGrain.
OpenAI's provenance help documentation states that textGrain does not insert hidden characters, invisible spaces, unusual punctuation, or extra watermark-only tokens. Its signal is part of the wording.
That distinction changes what document inspection can establish. A character-inspection tool can report the characters present in a file. It does not follow that a clean character report establishes clean provenance. If a tool advertises “ChatGPT watermark removal,” first determine whether it means removing formatting artifacts or evaluating an actual embedded text signal. Those are different jobs, with different evidence required to claim success.
Can You Check a ChatGPT Text Watermark?
There is no generally available OpenAI website where anyone can paste an article and check its textGrain watermark. The public OpenAI verification tool supports images and audio, and explicitly says it does not currently support text verification.
The Content Provenance API documentation confirms that text verification requires organizational approval. OpenAI's help documentation further limits eligibility to provenance research. Enabling API watermarking does not grant detector access; commercial verification and checking student work are outside the current program's scope.
For an editor evaluating a claim about a draft, separate these kinds of evidence:
| Report supplied | Question it addresses |
|---|---|
| Character or formatting inspection | Which characters or formatting artifacts are present? |
| Ordinary AI classifier score | Does the writing resemble patterns the classifier associates with AI? |
| Authorized textGrain detector result | Does the text contain the supported keyed watermark signal? |
OpenAI distinguishes ordinary classifiers from detecting its embedded signal. Ask a supplier which mechanism their report uses, rather than accepting “AI detected” as a complete explanation. Our guide to detecting AI-written content without false certainty covers broader review limitations.
Even a watermark result has boundaries. OpenAI warns that detection can be wrong, does not identify the user, and cannot establish ownership, truth, or how much a person contributed. Missing a mark does not prove human authorship. Treat a result as one piece of provenance evidence, then examine the actual drafting and editing record before making claims about the writer.
Does Copying or Editing Remove the Watermark?
Copying ChatGPT text into a CMS is not, by itself, a watermark-removal method. OpenAI says that preserving the wording is expected to preserve the text signal when the content is reused. Substantial rewriting, paraphrasing, or translation can make detection less reliable, but that is not a guarantee that a particular passage becomes unmarked.
Consider an illustrative publishing workflow: a writer copies a response into a plain-text field, then applies the site's heading styles. If the wording remains unchanged, removing rich formatting has not addressed the word-choice pattern described above. A clean paste and a negative provenance result are separate claims.
Detection is also harder for short passages and text with tightly constrained wording, such as precise mathematical answers. Even unchanged output is not guaranteed to produce a detectable signal.
For a publisher, “we pasted it as plain text” therefore documents an editing step. It does not establish that OpenAI provenance disappeared, that a human wrote the text, or that a detector can reliably settle either question.
What to Record Before Publishing AI-Assisted Text
When you cannot run an authorized watermark check, record what you actually know about the draft. Save the original output and meaningful revisions, along with the generation date, product, and any known model or watermark setting. Mark missing details as unknown instead of inferring them from the finished prose.
For example, an editorial record could look like this. These entries are illustrative, not results from a tested account:
| Field | Example entry |
|---|---|
| Generation context | ChatGPT, October 13, 2026; account region not confirmed |
| Model and watermark status | Model not recorded; no authorized textGrain check performed |
| AI contribution | Initial draft produced with AI assistance |
| Revision evidence | Original response retained; later factual corrections and rewritten sections saved |
| Review status | Automated source checks completed; named human approval pending |
This record gives an editor concrete questions to resolve. It does not certify authenticity, replace review, or prove the absence of a watermark. If a client asks whether the piece was AI-assisted, the drafting history answers that question more directly than a character-cleaning report.
Keep the publishing decision focused on the work itself. Google's guidance on generative AI content emphasizes accuracy, quality, relevance, and added value. A provenance signal is not a content-quality assessment. For the broader search implications, see our guide to AI-generated content and SEO.
Finally, OpenAI says machine-readable provenance does not replace visible notices that may be required. Check the rules that apply to your particular publication; a watermark result cannot settle that obligation for you.
Documentation checked October 13, 2026. This guide was written with AI assistance and automated source and editorial checks, not independent human review. It reports documented behavior, not original detector experiments or confirmation of rollout on individual accounts.