OpenAI will begin adding an invisible statistical watermark to eligible text generated by ChatGPT and Codex in the European Union, turning AI provenance from a research challenge into a feature that could affect millions of everyday users.
The change is being introduced as the European Union begins enforcing transparency obligations under the AI Act. But the most important detail is what the watermark does not do: it is not a visible label, it does not identify an individual ChatGPT user, and its absence cannot prove that a piece of text was written entirely by a human.
According to OpenAI’s official announcement, the technology is called textGrain and is designed to leave a statistical signal inside AI-generated text.
Key Takeaways:
- OpenAI says eligible ChatGPT and Codex text in the EU will receive an invisible watermark.
- The system, called textGrain, changes statistical patterns in token selection rather than inserting visible labels or hidden characters.
- API developers can opt into watermarking on supported models, while it is not being switched on globally by default.
- Editing, paraphrasing, translation and short passages can significantly weaken detection.
- OpenAI is initially restricting detector access to approved researchers and expert organizations.
ChatGPT Watermarking Is Coming to the EU
OpenAI says text watermarking will roll out to eligible ChatGPT and Codex users across the European Union.
For developers, the situation is different. API customers can opt into watermarked output on supported models, meaning the technology is not becoming a universal requirement for every response generated through OpenAI’s developer platform.
That distinction matters.
The announcement does not mean that every sentence produced by an OpenAI model anywhere in the world will suddenly carry a detectable watermark. The consumer rollout is focused on Europe, while developers have more flexibility.
The move comes as the European Union implements Article 50 of the AI Act, which introduces transparency obligations for providers of generative AI systems.
The European Commission’s guidance on AI-generated content says providers should make synthetic text, audio, images and video detectable in a machine-readable form when technically feasible.
For readers tracking how regulation is reshaping major AI products, this is another example of legal requirements beginning to influence product design. More coverage of artificial intelligence, software and emerging technology is available on TechNewsHome.
How OpenAI’s textGrain Watermark Works
The word “watermark” may suggest a hidden code attached to a document.
textGrain works differently.
Large language models such as ChatGPT generate text by repeatedly choosing the next likely token. A token can represent a whole word, part of a word or punctuation.
OpenAI says textGrain subtly changes the probability of certain token choices while the model generates a response. Those choices create a statistical pattern that a compatible detector can later analyze.
In other words, the watermark is embedded in the way the text is constructed rather than being added afterward as metadata.
The underlying research is described in OpenAI’s textGrain technical paper.
That approach offers one major advantage: copying the text from ChatGPT into a website, document, email or social network does not automatically strip away the signal.
But there is also an obvious weakness.
Change enough of the wording and the statistical pattern becomes harder to detect.
Detection Works, but It Is Not Foolproof
OpenAI is clear that textGrain should not be treated as a perfect AI detector.
In company evaluations using a target false-positive rate of 1%, OpenAI reported detecting the watermark in roughly 80% of certain 200-token passages and around 95% of comparable 400-token passages.
Detection becomes more difficult in highly constrained content, including mathematical material, where the model has fewer natural wording choices.
Editing has an even greater effect.
OpenAI reported that in one test involving 400-token English passages, replacing 10% of the words with synonyms reduced detection from about 92% to 66%. Replacing 25% brought detection down to 17%.
Those numbers require context.
They come from OpenAI’s controlled evaluations and should not be interpreted as universal detection rates for every real-world document.
Results will depend on the length of the passage, the type of content, how heavily it has been edited and how the detector is configured.
This leads to one of the most important corrections to exaggerated interpretations of the announcement:
ChatGPT-generated text is not becoming permanently or infallibly traceable.
A heavily rewritten, translated or shortened response may no longer produce a strong enough watermark signal for reliable detection.
A Watermark Cannot Prove Who Wrote a Document
This distinction is particularly important for schools, publishers, companies and employers.
A positive textGrain result may indicate that an OpenAI system generated or processed part of a passage.
It does not establish who submitted the document.
The watermark does not reveal a ChatGPT account, conversation history, prompt or individual user. It also cannot determine how much of a final document was written or edited by a person.
Likewise, failure to detect the watermark does not prove human authorship.
There are many reasons why AI-generated text might not produce a positive result.
The passage could have been heavily edited. It might have been translated. It could have been generated before watermarking was enabled. It could come from an unsupported model or from another AI provider entirely.
That means textGrain is better understood as a provenance signal than a definitive AI-authorship test.
For academic integrity investigations or workplace disputes, relying on a watermark detector alone would therefore be risky.
Why OpenAI Is Restricting Access to the Detector
OpenAI is not giving unrestricted public access to its textGrain detector at launch.
Instead, the company says approved researchers and expert organizations will initially receive access on a case-by-case basis.
That cautious approach is notable.
AI detectors can produce both false positives and false negatives, and the consequences of misinterpreting a result can be serious.
In universities, a false accusation could affect a student’s academic record. In publishing, it could wrongly challenge an author’s work. In professional environments, it could lead to disputes over whether employees used generative AI.
Restricting access may also make it harder for users to systematically test ways of defeating the watermark.
The European Commission’s approach similarly focuses on practical, technically feasible methods rather than demanding a system that can identify every piece of AI-generated content with absolute certainty.
AI Text Provenance Is Becoming Part of the Infrastructure
The most significant part of OpenAI’s announcement is not whether textGrain can catch every ChatGPT-generated paragraph.
It cannot.
What is changing is the expectation that major generative AI providers should supply technical mechanisms that help identify synthetic content.
OpenAI already supports provenance mechanisms for some other types of media. Text is considerably more difficult because wording can be changed quickly while the underlying meaning remains almost identical.
That makes robust text watermarking one of the harder technical problems in AI provenance.
The EU AI Act is effectively pushing that problem from research papers into real-world products used by millions of people.
Over time, watermarking could become another signal used by publishers, social networks, regulators and researchers when assessing the origin of online content.
But users should understand its limitations from the beginning.
textGrain is not a digital fingerprint proving that a specific person used ChatGPT. It is not a universal AI detector. And it cannot guarantee that rewritten AI content will remain detectable.
Its real value is narrower but still important: providing additional evidence about how a piece of text may have been produced.
As AI-assisted writing becomes increasingly routine, that distinction between provenance and proof of authorship may become more important than the watermark itself.
FAQ:
Does all ChatGPT text now contain a watermark?
No. OpenAI says watermarking is being introduced for eligible ChatGPT and Codex text in the European Union. API developers can opt into the technology on supported models.
Can users see the ChatGPT watermark?
No. textGrain is an invisible statistical signal created through the model’s token-selection process. It is not a visible label or hidden character added to the text.
Can editing remove the ChatGPT watermark?
Editing can significantly weaken its detectability. OpenAI’s own testing shows that paraphrasing or replacing words can substantially reduce the detector’s ability to identify the watermark.