ChatGPT Text Watermark in the EU: Why Teachers Can't Check It

Updated October 16, 2026: added who can and can't use OpenAI's text detector (including why schools can't use it to check essays), how API customers switch the watermark on, new details from the textGrain technical report and fresh coverage from Ars Technica, The Register and TechRadar.

ChatGPT and Codex text in the European Union is getting an invisible watermark, but almost nobody will be able to check for it. OpenAI announced the change on October 5, 2026, to meet the EU AI Act's transparency rules. Eleven days later, its updated help center makes one point especially clear: the text detector is limited to approved research groups, is free for them, and specifically does not cover academic integrity work such as checking student essays. Here is what the watermark is, who it affects and what it can and cannot prove.

What happened

In a blog post titled "Our approach to EU text provenance rules," published October 5, 2026, OpenAI laid out a phased plan. From that day, API customers anywhere in the world could opt in to text watermarking for select models, though it stays off by default. Over the following weeks, the company said, it would add an invisible watermark to eligible ChatGPT and Codex text output for users on all plans, but only in the EU. TechCrunch reports that OpenAI is not making text watermarking a global default at launch, and TechRadar (October 13) notes that US and UK ChatGPT users won't be affected.

The third step is the detector. OpenAI opened applications for access to its text watermark detector, but access is limited at first to approved researchers and expert organizations that can help evaluate and improve the technology. The company said that, given the risk of missed watermarks and false positives, it is not making the detector publicly available at launch. Its existing tools for checking images and audio, including the openai.com/verify web tool and its Content Provenance API, remain open to the public.

Key details

How textGrain works

OpenAI calls its method textGrain. According to the company's help center, it subtly adjusts how the model randomly chooses between possible words or word pieces as it writes. A secret key decides which adjusted choices are used, and over a long enough passage those choices form a pattern that a detector holding the same key can look for. OpenAI says the watermark does not add hidden characters, invisible spaces or unusual punctuation, and it is not visible to readers. The accompanying technical report, "textGrain: Entropy-Calibrated Watermarking for Language Model Text," lists authors from the University of Pennsylvania, Yale University and OpenAI. It says the detector needs only the generated text and the secret key, and does not need to know how strongly the watermark was set during writing. OpenAI says it built its own method, rather than using SynthID or TextSeal, to get more control over the trade-off between how detectable the watermark is and how varied answers to the same prompt can be. It still plans to release textGrain as open-source technology, but it has not yet linked a code repository or given a release date.

What OpenAI's own tests show

OpenAI published several test results, and they show both promise and weak spots:

Length matters. At a target false positive rate of 1%, the detector found watermarks in about 80% of 200-token passages and about 95% of 400-token passages for content such as psychology answers.
Math is harder. Detection rates were substantially lower for content such as mathematics, where there is less freedom in word choice.
Editing weakens it. In 400-token passages, swapping 10% of the words for synonyms cut detection from about 92% to 66%. Swapping 25% cut it to 17%.
Language matters too. The help center says that, in tests across all 24 official EU languages using translated prompts, Spanish had the highest detection rate (69.0%) and Romanian the lowest (42.2%). OpenAI says it can turn up the watermark's strength for languages where the signal is weaker.
No quality hit, it says. On the benchmarks it uses for its Astra model, OpenAI reports no meaningful performance differences with and without watermarking, and it says the speed impact is negligible. The help center adds that earlier tests in ChatGPT showed no change in thumbs-down rates or other product measures.

OpenAI also says that, in its evaluations, textGrain matched or beat the other approaches it tested, including Google DeepMind's SynthID for text. Even so, the company warned that "strong performance under ideal conditions does not guarantee reliable detection in everyday use." Outside reviewers were blunter. The Register called it "weak sauce" watermarking, pointing to how easily synonym swaps cut detection, and Ars Technica said the results are roughly in line with similar watermarking systems seen before.

What a watermark does not tell you

The company listed several things a detection result cannot establish. A watermark does not measure how much a human contributed, does not settle who owns the text or who is responsible for it, does not identify the user, and does not verify whether the text is accurate. OpenAI also said the absence of a watermark "does not prove human authorship," because text may be too short, edited, translated, produced by an unsupported model, or generated by another company's tools. Other tools take a different route. Wikipedia's new LLM Paste Check, which we covered in our report on rogue OpenAI agents and Wikipedia's May outage log, warns editors when pasted text looks copied from one of five AI chatbots, and Wikimedia says it looks for markers in the copied HTML rather than using watermark detection.

Who can check for the ChatGPT watermark

OpenAI's help center page on provenance signals, updated in the past week, spells out the detector rules in more detail than the original announcement:

Researchers only, for now. Qualifying use cases are academic and research organizations studying text provenance, detection reliability or how people understand provenance results. OpenAI says it reviews applications case by case.
Not for checking student work. The help center says academic integrity uses, "such as checking student work for AI use (e.g. in writing essays or theses)," are currently outside the program's scope.
Not for business screening either. General business uses such as commercial fraud prevention, authentication, verification or reselling detection as a service are also currently out of scope.
Free for approved groups. OpenAI says text detector checks are currently free.
Named partners. OpenAI says it is working with a limited group that includes John Thickstun of Cornell University, Martin Vechev of ETH Zurich and INSAIT, and researchers at the Kempelen Institute of Intelligent Technologies (KInIT).

The help center also stresses that the watermark contains no details about the user, organization or prompt, and that the detector only reports whether it found an OpenAI watermark.

How API customers turn it on

For developers, the help center now gives the exact settings. An organization-wide default lives under Organization settings, then Data controls, then Text provenance, and a single project can override it under Project Settings, then Text provenance. Customers switch on "Allow text watermarking," pick which models should be covered and save. OpenAI says API outputs are covered except for models planned for imminent deprecation, and that it will extend coverage to all legacy models over the coming weeks. Turning on watermarking does not give a customer access to the text detector.

Why it matters

It is driven by EU law. According to the European Commission, Article 50 of the AI Act has applied since August 2, 2026, and requires providers of generative AI to mark AI-generated content in a machine-readable, detectable way. AI systems already on the market before that date have until December 2, 2026 to meet the marking requirement.
Short answers and code are mostly outside it. OpenAI notes that the EU's Code of Practice on transparency of AI-generated content does not require watermarks in outputs shorter than 200 tokens, about 150 English words, or in code snippets. TechRepublic points out that Codex being included does not mean every code snippet it writes will carry a detectable watermark.
A reversal of sorts. TechCrunch notes that The Wall Street Journal reported in 2024 that OpenAI had built a text watermark but held off on releasing it, partly over concerns that users would switch to rivals. Search Engine Journal adds that a company survey at the time found almost 30% of ChatGPT users said they would use the service less if watermarking were introduced.
Rivals are moving too. TechCrunch reports the announcement comes two months after Anthropic said it would watermark text generated by Claude worldwide, a decision that drew backlash from some users. Ars Technica notes the difference: Anthropic enabled its watermark globally, while OpenAI makes it the default only where regulators require it. Search Engine Journal reports that Anthropic uses a version of SynthID-Text. Anthropic, Google, Meta, Microsoft and OpenAI are among the companies that have committed to follow the EU's code of practice, according to TechCrunch.
Schools and workplaces should be careful. Because a watermark cannot show how much of a passage a person wrote, and because OpenAI's detector is not offered for checking student work, TechRepublic's advice matters more. It advises users under AI disclosure policies to keep their own records of AI use, and warns organizations against treating a detector match as proof of authorship. For everyday habits that get better results from the chatbot, see our 10 ChatGPT tips for better answers in 2026.

What's next

As of October 16, 2026, OpenAI has not announced that the EU ChatGPT rollout is complete, though its help center now describes EU ChatGPT text as carrying textGrain watermarks. Because the change is invisible, EU users may not notice anything at all. OpenAI says it is working with cloud partners to bring watermarking to its models' outputs on their platforms in the coming weeks, and that it will extend API coverage to all legacy models in that period. Its technical report is due to be updated with more details, and the company says it will expand detector access only when it believes "results can be interpreted responsibly."

Some questions remain open. Search Engine Journal noted that OpenAI's post did not say whether EU ChatGPT users can switch the watermark off, and the help center still does not offer a user setting for it. (API customers can now see the list of supported models in their settings.) It is also unclear when, or whether, the watermark will reach users outside the EU. For now, OpenAI says it expects to revisit each part of its approach as the technology, standards and evidence evolve. The December 2 deadline gives regulators a clear date to watch for how other AI providers comply. The watermark is not the only ChatGPT change this month: on October 15, OpenAI began rolling out GPT-Live-1.1, which brings new voices to ChatGPT Voice, according to its announcement on X.

FAQ

Will ChatGPT text be watermarked in the US or UK?

Not by default for now. OpenAI says the ChatGPT and Codex watermark is rolling out in the EU only. Developers anywhere can opt in to watermarking for select models through the API, where it is off by default.

Can I check if text has OpenAI's watermark?

Not yet as an ordinary user. OpenAI's text detector is limited to approved research and academic organizations who apply for access. The company's public verify tool works for supported images and audio, not text.

Can teachers use OpenAI's detector to check essays?

No, not for now. OpenAI's help center says academic integrity uses, such as checking student essays or theses for AI use, are currently outside the scope of its text detector program.

How do I turn on text watermarks in the OpenAI API?

Go to Organization settings, then Data controls, then Text provenance for an organization default, or Project Settings, then Text provenance for one project. Turn on "Allow text watermarking," choose your models and save.

Can the ChatGPT watermark be removed?

OpenAI's own tests suggest heavy editing can defeat it. Replacing a quarter of the words with synonyms dropped detection to 17% in one test, and translation and substantial paraphrasing can also make the watermark undetectable, according to the company.

Search This Blog

The Frontier Brief