OpenAI is preparing to add an invisible watermark to text generated by ChatGPT and Codex for users in the European Union. The change will roll out over the coming weeks and is part of OpenAI’s response to the EU AI Act, which requires AI-generated text to be identifiable in a machine-readable way.
There will be no visible “AI-generated” label sitting underneath a ChatGPT response. Instead, OpenAI will alter the statistical pattern of the model’s word choices so that a separate detector can identify a signal in the finished text.
OpenAI calls the technology textGrain. It is designed to work without inserting hidden characters, unusual punctuation or invisible spaces into the response.
What Exactly is OpenAI Watermarking?
The watermark is built into the way ChatGPT chooses words.
When a language model generates a sentence, it calculates the likelihood of many possible next words or tokens. textGrain subtly adjusts those choices using a secret pattern. Over a longer passage, those small changes create a statistical signature that a detector can look for.
For someone reading the text, nothing should look different.
OpenAI says the watermark also remains when users copy and paste the text because the signal is contained in the wording itself rather than in a separate file or piece of metadata.
That makes it different from many conventional AI detectors. A third-party detector generally tries to guess whether writing looks AI-generated after it has already been produced. textGrain is an embedded provenance signal created when the model generates the text.
What Does OpenAI Say About the Rollout?
OpenAI says the EU rollout is deliberately phased because text watermarking is still an emerging technology.
“Text watermarking and detection remain early technologies with significant limitations,” the company said in its announcement. OpenAI added that its approach is intended to meet the EU requirements while being transparent about what the technology can and cannot establish.
Eligible ChatGPT and Codex users across all plans in the EU will receive the watermarking system as the rollout expands. OpenAI is not making it a global default at launch.
There is a separate option for developers. API customers anywhere in the world can opt in to text watermarking for selected models, although it will remain switched off by default.
Can the Watermark Tell If a Person Wrote the Text?
No. This is one of the most important details.
A detected watermark can indicate that an OpenAI model generated or processed part of a passage. It cannot tell how much of the work came from a human, whether the user substantially edited it, or who should be considered the author.
OpenAI itself says a watermark does not establish authorship, ownership or legal responsibility. It also cannot determine whether AI produced an entire document or helped edit material that a person had already written.
The opposite is also true. If a detector does not find a watermark, that does not prove a human wrote the text.
The passage may have been heavily edited, translated, generated by another AI system, or simply be too short for reliable detection.
That makes textGrain a provenance tool rather than a universal “AI detector.”
How Reliable is the ChatGPT Watermark?
OpenAI’s tests show that the system can work well on longer passages, but editing can weaken the signal.
In one test cited by the company, detection was around 92% before changes were made. Replacing 10% of the words with synonyms reduced detection to about 66%. Replacing 25% of the words brought it down to about 17%.
Shorter passages create another problem because there are fewer word choices from which the detector can identify the statistical pattern. Mathematical writing can also be harder to detect because there is less flexibility in choosing words.
Language is another variable. OpenAI’s testing across the EU’s 24 official languages found differences in detection performance, with Spanish performing better than Romanian in the company’s initial evaluation at a 1% false-positive rate.
These limitations explain why OpenAI is not immediately giving everyone access to the detector.
Who Will Be Able to Check for the Watermark?
OpenAI is initially limiting access to approved researchers and expert organisations.
The company has opened applications for groups studying text provenance, detection reliability and responsible uses of the technology. OpenAI says the restricted access will allow it to gather real-world feedback before making broader decisions about the detector.
So, for now, an ordinary ChatGPT user in Europe will receive watermarked text without necessarily having a public tool that can independently check every response.
Why is the EU Requiring this?
The move is tied to the transparency requirements of the EU AI Act.
The European Commission’s AI-generated-content rules require certain AI-generated or manipulated content to be identifiable. The transparency obligations began applying from August 2, 2026, including requirements around AI-generated or manipulated text published on matters of public interest.
OpenAI has also signed the EU Code of Practice on transparency of AI-generated content and says text watermarking is part of its effort to meet those commitments.
The regulation is therefore pushing AI companies toward systems that can provide provenance information without relying entirely on visible labels.
What Does this Mean for People Using ChatGPT?
For most users, the immediate experience should not change.
Someone can still ask ChatGPT to write an email, article, report or other piece of text. The difference is that eligible AI-generated text in the EU will carry a signal that software may be able to identify later.
For publishers and businesses, that could eventually provide another way to establish whether OpenAI-generated material was involved in a workflow. For researchers and educators, it could offer an additional provenance signal when examining AI-assisted work.
But it should not be treated as a final answer to the question, “Did AI write this?”
A watermark cannot measure human contribution, and editing can weaken detection. OpenAI acknowledges those limitations itself.
OpenAI plans to open-source textGrain
OpenAI says it plans to make textGrain available as open-source technology so researchers and developers can examine the approach and build on it. The company also says its testing showed textGrain matched or exceeded the performance of other approaches it evaluated, including Google’s SynthID for text.
That could become an important part of the rollout.
Rather than keeping the entire system behind a proprietary wall, OpenAI wants outside researchers to examine how text watermarking performs in different languages, formats and real-world situations.
For now, though, the practical change is straightforward: ChatGPT and Codex text generated for users in the EU will begin carrying an invisible OpenAI signal, while developers worldwide can choose to turn the technology on through the API.
The watermark may help establish that OpenAI was involved in generating a passage. It will not, by itself, tell the whole story of who wrote it.
