According to AI Weekly, on October 5, 2026 OpenAI started rolling out an invisible watermark, called textGrain, in text generated by ChatGPT and Codex for users in the European Union. OpenAI's official page, "Our approach to EU text provenance rules", instead speaks of a rollout in the coming weeks. If you use ChatGPT to write the text for your website, your social posts or your emails to customers, this affects you, even though you won't see anything different on screen.
What we know, and what we don't
We know how it works, according to OpenAI's page. The watermark adds an invisible statistical signal to the model's word choices, which a detector can then recognize. ActuIA specifies that no hidden characters or invisible spaces end up in the text: the signal lives in the words chosen. For API customers, worldwide, the watermark is off by default and can be turned on. ActuIA and AI Weekly cite Article 50 of the EU AI Act, which since August 2, 2026 requires AI-generated content to be marked in a machine-readable way.
We also know how reliable it is, at least in OpenAI's tests. On psychology texts, at a 1% false positive rate, the detector recognizes about 80% of 200-token passages and about 95% of 400-token ones. A token is a small piece of text. On math texts the result is much lower, because there are few words to choose from. If 10% of the words are replaced, detection falls to about 66%, and with 25% to about 17%. For now, the detector is available only to researchers and approved organizations.
OpenAI also puts the limits in writing: the watermark does not measure human contribution, does not establish who owns or is responsible for a text, does not identify who produced it and does not verify that it is accurate.
We don't know when it will reach every account, because AI Weekly's date and OpenAI's wording don't match. We don't know how it works in Italian: ActuIA reports lower percentages in languages other than English, but Italian is not among those I read. We don't know whether a search engine will ever use this signal: no source says so.
What it means for a small business
This is our own reading of the sources. If you run a small business, little changes in your daily work today. The detector is not public, so nobody can check your texts with one click, and no source links the watermark to Google rankings. The news reminds you of something more practical: a text that came out of ChatGPT carries a trace, and the responsibility for what you publish remains yours anyway.
Rewriting a text to hide where it comes from is wasted time, and OpenAI itself says the watermark does not measure human contribution. Your time is better spent making the text right: checked facts, real prices, a tone that sounds like you.
What to do
- Decide where you use AI to write. List the texts that start from ChatGPT: website pages, posts, emails, product descriptions.
- Check every fact before publishing. Numbers, opening hours, prices, names and rules should be compared with a source you can open.
- Add what the AI doesn't know. A real case, a photo of your own, the true price, the questions your customers actually ask.
- Consider a line of transparency, where it makes sense: "draft written with the help of AI and reviewed by a person" is enough.
- If you have an API or connected tools, ask whoever manages them whether the watermark is on and what effect it has on your texts.
What to avoid
- Believing the watermark certifies quality. OpenAI says it does not verify the accuracy of a text.
- Buying tools that promise to "erase the AI signature". No source shows they help your site, and whoever guarantees it is guessing.
- Assuming it applies to every tool. The sources talk about ChatGPT and Codex, and about anything else we don't know.
A hypothetical example
Hypothetical example: a craftsman who sells windows uses ChatGPT to write the page "How to choose a window". Before publishing, he compares every measurement and every warranty with the manufacturers' data sheets and finds a duration written wrong. He adds two photos of his own job sites and the questions customers ask most. At the bottom of the page he puts a line: "Draft written with the help of AI, reviewed by Marco". The watermark, if there is one, changes none of this.
The bottom line
OpenAI is rolling out an invisible watermark in ChatGPT and Codex text for European users, with a detector closed to the public and limits stated by the company itself. For a small business the useful move is the usual one: check the facts, add your own material and sign what you publish. If you use ChatGPT for your website text, or have already wondered how the rules are changing, write to us at info@sarabi.cloud and tell us your case.
Sources: OpenAI, Our approach to EU text provenance rules, ActuIA and AI Weekly.
Have a similar case or a doubt? Write to us at info@sarabi.cloud.