ChatGPT watermark in the EU: who gets it and what it misses
OpenAI's ChatGPT watermark reaches EU users in the coming weeks. Who gets it, who can run the detector, and how fast light edits erase it, in OpenAI's numbers.
Source-based. Written from the documents, reporting and reviews linked in the text. Nothing here was tested hands-on by The Ruling Desk. How we work
OpenAI will add an invisible ChatGPT watermark to text written for users in the European Union, starting over the coming weeks. The company announced on October 5, 2026 that the method, called textGrain, will mark eligible ChatGPT and Codex output in the EU, and that API customers anywhere can switch it on today. Here is who gets watermarked, who can actually check for it, and how much editing it takes to make the mark fade, all in OpenAI's own test numbers.
Key takeaways
- Who: eligible ChatGPT and Codex users in the EU, on all plans, over the coming weeks. API customers worldwide can opt in now; it's off by default outside that EU rollout.
- Who can detect it: at first, only approved researchers and expert organizations. There is no public checker yet.
- How reliable: in OpenAI's tests, about 95% of 400-token passages were flagged, about 80% at 200 tokens, at a 1% false-positive rate.
- How fragile: by OpenAI's own figures, swapping 10% of words for synonyms cut detection from about 92% to 66%, and swapping 25% cut it to 17%.
- Why now: the EU AI Act's transparency rules took effect on August 2, 2026 and require AI-generated content to be marked so machines can detect it.
How the ChatGPT watermark works
textGrain doesn't add hidden characters or invisible spaces. It works on word choice. A language model picks each next word from a list of likely options, and textGrain uses a secret key to nudge those picks in a pattern a reader can't see but a detector holding the key can, as TechCrunch explains. One nudge proves nothing; the signal comes from hundreds of them added up across a passage.
The math is in a 20-page technical report dated October 5, written by five OpenAI researchers and four academics from the University of Pennsylvania and Yale. It says the detector needs only the text and the secret key. OpenAI also claims the watermark doesn't meaningfully change answer quality, with GPT-6 Astra's benchmark results "broadly similar" when textGrain is on, BleepingComputer reports. That is the company's claim, not an independent test.
Who gets watermarked, and when
OpenAI's announcement splits the rollout in two:
| Who | Watermark | When |
|---|---|---|
| ChatGPT users in the EU, all plans | On, for eligible text | Over the coming weeks |
| Codex users in the EU | On, for eligible text | Over the coming weeks |
| API customers, anywhere | Opt-in, off by default | Available since October 5 |
| ChatGPT users outside the EU | Not announced | No date given |
OpenAI hasn't spelled out what makes text "eligible," and neither TechCrunch nor BleepingComputer lists exclusions. If you use ChatGPT from Mexico, the US or anywhere else outside the EU, nothing changes for you today. If you build on the API, the switch is yours.
Who can check for it
This is the part that matters most if you're a teacher, an editor or someone accused of using AI. OpenAI is taking applications for its watermark detector, but access will initially go only to "approved researchers and expert organizations, who can help us evaluate reliability and responsible uses," in OpenAI's words as quoted by TechCrunch. You can't paste an essay into a public tool and get an answer.
OpenAI also lists what the watermark can't do. It doesn't reveal who generated the text, their account, prompt or conversation, and it can't say how much of a final piece a human wrote or edited. And in OpenAI's words, "a missing watermark does not prove human authorship": the text may be too short, heavily edited or written by another AI.
How much editing it survives
OpenAI's own watermark detection numbers show where the signal holds and where it fades. All of these are the company's test results, tuned to a 1% false-positive rate, meaning about 1 in 100 unwatermarked texts would be wrongly flagged:
| OpenAI's test | Detected |
|---|---|
| 400-token passage, unedited | About 95% |
| 200-token passage, unedited | About 80% |
| 400 tokens, editing-test baseline | About 92% |
| Same, after 10% of words swapped for synonyms | 66% |
| Same, after 25% of words swapped | 17% |
For scale, 400 tokens is roughly 300 English words, a short essay. Short passages, math answers and translated text are harder to detect, according to TechCrunch. Language matters too: across the EU's 24 official languages, detection ranged from 42.2% in Romanian to 69.0% in Spanish in OpenAI's multilingual test, ActuIA reports from OpenAI's post.
The upshot: a lightly edited essay can still carry a signal, but a quarter of the words rewritten leaves a detector with little to go on. That fits the limits OpenAI itself names.
What AI Act Article 50 requires
The EU-only rollout lines up with the law. Article 50(2) of the AI Act says providers of AI systems that generate synthetic text, images, audio or video must mark the output in a machine-readable format so it's detectable as AI-generated. Those transparency duties have applied since August 2, 2026.
There's some slack. Generative AI systems already on the market before August 2 have until December 2, 2026 to meet the marking requirement, under the EU's later amendments, law firm Cooley notes. Breaches can cost up to €15 million or 3% of worldwide annual turnover, whichever is higher. The Commission also published a voluntary Code of Practice on AI-generated content on June 10, which asks for marking that is "effective, interoperable, robust, and reliable as far as technically feasible."
The law asks for robustness "as far as technically feasible," and OpenAI's editing numbers show where that line sits for text today. It isn't the only EU rule shaping ChatGPT: the bloc's platform law is also why ChatGPT's new image ads will end up in a public archive.
What happens next
- The EU rollout lands over the coming weeks; OpenAI gave no exact date.
- Detector access widens only after researchers test it; OpenAI hasn't said when or to whom next.
- December 2, 2026 is the marking deadline for generative AI systems that were already on the market in August.
OpenAI is under scrutiny on other fronts this week too, including Wikimedia's claim that OpenAI agents may have contributed to a Wikidata outage. More on the rules and the companies they cover is in our AI section.
Bottom line
If you're in the EU, ChatGPT and Codex text will soon carry a mark you can't see and, for now, can't check yourself. If you're elsewhere, nothing changes unless a developer turns it on through the API. By OpenAI's own numbers, the ChatGPT watermark works best on longer, unedited text and fades fast under light rewriting, so treat a missing watermark as no proof that a person wrote the text. Watch who gets the detector next and what the EU says about robustness after December 2.
FAQ
Does ChatGPT watermark its text?
Starting over the coming weeks, yes, for eligible ChatGPT and Codex output for users in the EU, on all plans. Outside the EU, OpenAI hasn't announced a default watermark. API customers anywhere can opt in, and it stays off unless they do.
Can anyone check text for the watermark?
Not yet. OpenAI says detector access will first go only to approved researchers and expert organizations, and it is taking applications. There is no public tool for teachers, employers or readers as of October 6, 2026.
Does a missing watermark prove a human wrote it?
No. OpenAI says the absence of a watermark doesn't prove human authorship. The text could be too short, heavily edited, translated or produced by another AI system.
Can the watermark identify who used ChatGPT?
No. According to OpenAI, the watermark doesn't reveal the account, prompt or conversation behind the text, and it can't say how much of it a person edited.