OpenAI's ChatGPT Watermark in the EU: Will It Show Up in a Turnitin Report?

Nothing published so far says it will. OpenAI announced on 5 October 2026 that it will add an invisible watermark to "eligible ChatGPT and Codex text output in the European Union". The announcement says this happens over the coming weeks; OpenAI's help centre, updated about a day later, already says in the present tense that EU ChatGPT text includes it. Neither page gives a start date. The detector that reads the mark is not public: OpenAI is taking applications from research and academic organisations and approving them case by case. Turnitin's help centre and its two changelog pages, checked on 7 October, don't mention the watermark at all. The AI percentage in a Turnitin report comes from Turnitin's own classifier, which, as Turnitin describes it, scores the wording itself.

So as far as the watermark goes, the report your instructor opens this week works the way it did before the announcement. (Turnitin did make one unrelated change on 6 October: at institutions that license it and have it switched on, AI detection now also covers German, Turkish and Dutch submissions.) What the watermark does change is narrower, and the rest of this piece goes through it: which text carries it, who can check for it, and what a check could and couldn't show.

OpenAI's post is called "Our approach to EU text provenance rules" and is dated 5 October 2026. Our scripted requests to it were refused, so we worked from the copy the Internet Archive saved on 6 October at 13:42 UTC, and checked it against the live page in a browser on 7 October; the text matched. A moderator's summary on OpenAI's developer forum, posted the same day, gives the same timeline. The post ends by pointing to a help centre article on provenance signals, which carries an "Updated" label from about a day later. The post sets out three steps, and they don't all start on the same day.

Two things in that table matter more than they look.

The two OpenAI pages don't line up on timing. In the post, the ChatGPT step is written as "we will add". The help centre's entry for text is in the present tense: "In the EU, ChatGPT-generated text includes an invisible change to the randomness used in the model's word choices to comply with the EU AI Act." Neither page gives a start date, so text you produce with ChatGPT in the EU from this week on may already carry the mark, and nothing on either page lets you tell. Text from September came out before this EU rollout was announced. The help centre does mention "Prior tests in ChatGPT", without saying when they ran or for whom.

And "eligible" is never defined in the post. It doesn't name the ChatGPT models that will carry the mark, and neither page says how OpenAI decides that text was produced in the EU, whether by account country, by where you are when you type, or something else. We searched the text of the post for location, travel, VPN and IP address and got no hits (the word watermark comes up 42 times in the same text, so the search was reading it), and the help centre page doesn't mention location or country either. If you study in Lyon on an account opened in Toronto, or the other way round, neither page tells you which side of the line you're on.

There's one more route that has nothing to do with ChatGPT. The API option is worldwide, so a writing app built on OpenAI's models could turn watermarking on for its users wherever they are. The post doesn't list which customers have opted in, or say whether such an app would tell you.

The watermark is called textGrain. It's a statistical pattern in which words the model picks, so there's nothing to see on the page. OpenAI's technical report, released the same day, is specific about what reading it takes: "The detector requires only the generated text and the secret key". OpenAI calls the detector "our text watermark detector", and anyone without that key has nothing to check against.

OpenAI isn't handing that check out widely yet. The post says access "will initially be limited to approved researchers and expert organizations that can help us evaluate and improve the technology", granted "on a case-by-case basis", and gives its reason plainly: "Given the risk of missed watermarks and false positives, we are not making it publicly available at launch."

Could a university get access? It can apply, but for a stated purpose that isn't marking. OpenAI's help centre says "Text detector access is currently limited to approved research and academic organizations working to improve how reliably text watermarking works, how easy it is to use, and how clearly the technology and its results can be explained." It also names some of the researchers it works with now, including people at Cornell University, ETH Zurich and the Kempelen Institute of Intelligent Technologies. Neither page mentions students, teachers or Turnitin. The post does describe what the tool sends back: "whether it detects an OpenAI watermark, without identifying the user or revealing their prompts or conversations."

The post also says OpenAI plans "to make the technology available in open source". That sentence is about the method. It doesn't say that would give anyone a way to check text from ChatGPT, and the report is clear that the check needs the secret key.

Nor does anything in the announcement say ChatGPT itself can check text for the mark. If someone shows you a chatbot's opinion of your essay, that's a separate thing again; Is ChatGPT an AI detector? What OpenAI published about its own classifier covers it.

Not anywhere we looked. On 7 October, between 11:57 and 11:58 UTC, we searched the places where Turnitin would be most likely to say so. Each count is a case-insensitive match on the page text, with a control word from the same page to show the search was working.

The same FAQ does explain how Turnitin's detector works. Sentences are split into overlapping segments, and "Each segment is classified by the AI detection model and given a value between 0 and 1, denoting the probability of the text being likely human or AI-generated." A few paragraphs later it says its classifiers "are trained to detect these differences in word probability". That's a judgment made from the writing. Nothing in that description relies on a mark placed in the text, or on anyone's key.

The count only shows one thing: Turnitin hasn't said it reads the watermark. It doesn't show that Turnitin couldn't, or that it hasn't applied for access. A vendor that had done either quietly would look exactly the same from outside. If Turnitin does announce something, the product updates page is where it dated its last change to AI detection, so that's the page to watch.

Not much, and OpenAI's own post is the clearest source on that. It has a whole section headed "What a text watermark doesn't tell you", and three of its points bear directly on coursework:

• "A watermark does not measure human contribution. It can indicate that an OpenAI system generated or processed part of a passage, but not how much human judgment, editing, or creativity went into it."
• "A watermark does not identify the user. It does not associate a person, organization, account, prompt, or conversation with the text."
• "The absence of a detected watermark does not prove human authorship." Among the reasons OpenAI gives, text may "come from an unsupported model, predate watermarking, or have been generated by another company's tools."

The technical report describes the mark the same way, as "a binary provenance signal". It answers a single yes-or-no question about whether an OpenAI system was involved.

So a detection couldn't tell anyone whether you wrote the essay and had ChatGPT tidy two paragraphs, or the other way round. A clean result wouldn't clear you either. Neither outcome answers what your course actually cares about, which is whether the way you used AI was within its rules.

1. Go by your course's AI rules, not by the watermark. The legal duty behind all this sits with the AI companies. OpenAI's post puts it this way: "The EU AI Act requires generative AI providers to make generated text identifiable in a machine-readable way." That's an obligation on OpenAI. What you're allowed to do is set by your course and your institution, and none of that changed this week.
2. If your course asks you to declare AI use, declare it the way it asks. A watermark doesn't make a declaration more or less necessary. Which Writing Tools Count as "AI Assistance" You Have to Disclose? A Checklist goes through it tool by tool.
3. If a Turnitin AI score is raised with you, deal with it as a Turnitin score. It comes from the classifier described above, not from a watermark. Ask to see the full report and the highlighted passages. What your instructor is told to do when your AI percentage is high shows the guidance Turnitin gives the other side of the table.
4. If someone tells you a watermark was found, ask which tool found it and who ran it. At launch OpenAI's detector is only open to approved research and academic organisations, and OpenAI describes that access as being for work on how well watermarking works; a result from a tool without that access isn't a check for this mark. Even a real detection, by OpenAI's own account, says an OpenAI system was involved somewhere, not how much or by whom. If you did use ChatGPT in a way your course allows, say so and show how.
5. Keep your drafts. A version history shows how the work was built up over days or weeks, which neither a watermark nor a detector score can show.

I'm in the EU and used ChatGPT for an essay in September. Is that text watermarked? Not as part of the EU rollout, which OpenAI announced on 5 October. OpenAI lists text that may "predate watermarking" among the reasons a detector finds nothing. How Turnitin's classifier scores that essay is a separate question.

Can my teacher check my essay for the watermark? Not with any public tool. OpenAI says it isn't making the detector publicly available at launch. Access is by application, for approved research and academic organisations studying the technology.

Does it apply if I'm outside the EU? For ChatGPT, OpenAI describes it as EU only and says: "We are not making text watermarking a global default at launch." Separately, API customers anywhere can opt in for select models.

Will my Turnitin AI percentage change because of it? None of the Turnitin pages we checked on 7 October connect the AI score to a watermark. The score comes from Turnitin's own classifier.

How does OpenAI decide whether text was generated in the EU? Neither the post nor the help centre says.

• OpenAI, "Our approach to EU text provenance rules", 5 October 2026: live page, read in a browser 7 October 2026, and the Internet Archive copy of 6 October 2026, 13:42 UTC.
• OpenAI Help Center, article on provenance signals in OpenAI-generated content, labelled as updated about 6 October 2026; read in a browser 7 October 2026.
• OpenAI, textGrain: Entropy-Calibrated Watermarking for Language Model Text, technical report dated 5 October 2026; read 7 October 2026.
• OpenAI Developer Community, OpenAI's approach to EU text provenance rules, moderator post of 5 October 2026; read 7 October 2026.
• Turnitin, Turnitin's AI writing detection capabilities FAQs, Turnitin product updates and Turnitin release notes; read 7 October 2026.