OpenAI will add an invisible watermark to some texts produced by ChatGPT and Codex in the European Union in the coming weeks. Named textGrain, this technique changes very slightly the way the model chooses its words so that an authorized detector can then locate a statistical signal.
This mechanism can help to establish that a text probably comes from a compatible model. It does not, however, identify the author, measure the share of human work or verify the accuracy of the content. It therefore does not transform an AI detector into automatic proof of fraud, plagiarism or lack of human intervention.
Information verified on 6 October 2026. OpenAI distinguishes three situations: the option is already available to API clients worldwide for some models, it remains disabled by default in the API, and its application to eligible outputs of ChatGPT and Codex in the European Union is expected to be phased in over the next few weeks.
What OpenAI really deploys in Europe
OpenAI announced textGrain on 5 October 2026. For European users of ChatGPT and Codex, the change must concern eligible texts on all packages. This is a regional deployment announced, not a global and immediate activation on each response produced by OpenAI services.
On the API side, organizations around the world can request textGrain activation now on a selection of templates. However, the option remains disabled by default. A company using the API should therefore not assume that all its contents already carry this signal.
The detector is also not open to the general public. OpenAI is currently accepting applications from researchers and expert organisations and provides for initial access to approved actors. This restriction is intended, in particular, to limit misuse, but it means that a creator or customer will not necessarily be able to verify a text from the outset of deployment.
This initiative is part of the European framework for the transparency of synthetic content. The article 50 of the European Regulation on Artificial Intelligence requires suppliers of systems generating synthetic content to mark outputs in a machine-readable format, where technically possible. The same article provides separate rules for certain deployers who publish content in the public interest.
How textGrain works without adding any visible mention
A text watermark is not a label displayed under the paragraph and is not a metadata file attached to the document. textGrain acts during the generation. At each step, a language model chooses the following word or fragment of word from several possibilities. The system discreetly favours certain variants according to a secret rule, while trying to preserve the meaning and fluidity of the text.
A word alone proves nothing. But on a long enough passage, the repetition of these choices can form a statistical reason that the detector seeks. OpenAI explains that it designed the system to adjust the signal strength according to the uncertainty of the model: when several formulations are plausible, the watermark may be more marked; When the text is very constrained, it must preserve the correctness of the answer.
That difference counts. A narrative explanation of several hundred words offers more lexical possibilities than an equation, a line of code or a very short factual answer. The performance of the detector cannot therefore be summarized by a single rate valid for all situations.
In evaluations published by OpenAI with a threshold targeting around 1 % false positives, detection reaches approximately 80 % on psychology passages of 200 tokens and 95 % on passages of 400 tokens. Results are weaker on highly constrained content, including some mathematical answers. These figures describe controlled tests, not an individual guarantee for every text encountered online.
Why a detector cannot prove the author of a text
The most important point is within the limits recognized by OpenAI. textGrain looks for the likely presence of a signal produced by a model. He did not know who had requested the text, why it had been generated, how many corrections had been made or who took responsibility for the published version.
VIFLY’s perspective: a machine signal can document the intervention of a tool, but it never replaces proof of editorial responsibility. The best proof of human contribution is not the lack of watermark: it is traceable work, with sources, drafts, corrections and a person who responds to the result.
A positive text on the detector may have been thoroughly verified, completed and assumed by a specialist. Conversely, a text without detectable signal can be content generated by another model, too short output, or text whose watermark has been weakened by transformations.
OpenAI also specifies that textGrain does not allow to identify the user at the origin of the generation. It does not establish intellectual property, plagiarism or the veracity of claims. Presenting a detection result as definitive proof of authorship would therefore go far beyond what technology actually measures.
The modifications also weaken the signal. In published trials, the replacement of 10 % words with synonyms greatly reduces detection, and a transformation of 25 % words can make it very weak. It should not be concluded that it would be sufficient to rewrite to circumvent an obligation. This shows above all why the absence of a signal is not a reliable proof of human writing.
Two practical scenarios for creators and freelancers
Consultant prepares newsletter with ChatGPT
A consultant uses ChatGPT to propose a first plan, then adds its anonymous customer data, checks the numbers, replaces several examples and rewrites the conclusion. The final version may retain a textGrain signal, have weakened or not entered the eligible perimeter. None of these cases adequately measure their contribution.
If a customer disputes the authenticity of the content, the best folder is not a screenshot of a consumer detector. These are the sources consulted, successive versions, revision comments and final validation. The consultant may then publish the resource on the blog VIFLY and consolidate its expertise, offers and evidence into a LinkHub VIFLY. Confidence arises from the consistency of this set, not from an isolated score.
A trainer publishes an article on a topic of public interest
A cybersecurity trainer asks Codex or ChatGPT to help structure an article on a new regulatory obligation. It rereads the text, goes back to official documents, corrects errors and assumes publication under its name. The technical watermark and the possible obligation to inform the reader then answer two different questions.
The first helps a system to recognize a probable technical origin. The second concerns transparency towards those exposed to the content. In particular, the European Regulation provides for an exception to certain disclosure obligations for public interest texts where they have been subject to human or editorial control and where a person or organisation is responsible for their publication. This exception must be assessed according to the context; It does not mean that any quick rereading is enough.
To then turn this expertise into an opportunity, the trainer can propose a clear step, for example a diagnosis or an exchange via VIFLY Booking. textGrain does not change the fundamental logic: visibility, understanding, trust, action, opportunity.
Technical watermark and visible mention: two different subjects
The most likely confusion is that the invisible watermark of OpenAI automatically replaces a visible indication of the type "content generated with an AI". That is not what the European regulation says.
The article 50 separates several responsibilities. Generating system providers must make certain outputs detectable in a machine-readable format, as far as technically feasible. Separate obligations may apply to organisations or persons broadcasting hyperstructures, or to certain texts generated or manipulated by AI and published to inform the public on matters of general interest.
The presence of textGrain can therefore help OpenAI fulfil a technical obligation without producing the clear information that certain uses require from the public. Conversely, displaying a visible indication does not guarantee that a content bears a detectable watermark.
For the independents, the right question is not only "is the text marked?". We must ask: who provides the system, who uses the content, what type of publication, what level of human transformation exists and who assumes editorial responsibility?
Nuances and points to bear in mind
An organisational drift must also be avoided: automatically punish a student, employee, freelancer or author on the basis of a detector. OpenAI itself requests that the results be interpreted with caution and placed in a set of indices.
What to do next? The VIFLY action plan
The right strategy is not to try to produce an "undetectable" text. It consists of using the AI in an assumed way, protecting the quality of the information and being able to show how an idea has become published content.
TextGrain and ChatGPT Watermark FAQ
Are all ChatGPT texts now watermarked?
No. OpenAI announces a gradual deployment on the eligible outputs of ChatGPT and Codex in the European Union. In the API, textGrain is available as an option for some models and remains disabled by default.
Is the watermark visible in the text?
No. textGrain creates a statistical pattern through the word choices of the model. This is not a statement displayed to the reader.
Can a detector prove that someone cheated?
No. It may indicate the probable presence of a watermark, but it does not know the identity of the user, the context, or the amount of human work. There must be one more clue.
Does a rewrite remove the watermark?
Changes can weaken the signal, sometimes strongly. The result depends on the length, type of text and extent of the transformations. This does not remove any transparency obligations.
does textGrain verify that a text is true?
No. The watermark concerns the probable technical provenance, not the accuracy of the information. The facts must always be verified from reliable sources.
Should a visible mention be added to each IA-assisted content?
Not necessarily. Obligations vary according to role, type of content, broadcast context and level of editorial control. A clear internal policy and a concrete case analysis are still needed.