ChatGPT watermark in the EU: what it means for your business

What happened

On October 5, OpenAI announced that over the coming weeks it will start watermarking the text that ChatGPT and Codex generate for users in the European Union. The watermark is called textGrain and you can't see it: there is no stamp and no sentence at the end of the text. What changes, very subtly, is which words the model picks, and that leaves a statistical trace that a detector holding the secret key can recognize.

It applies to every ChatGPT plan in the EU, from the free one to the business plans. Outside Europe, not for now.

The reason is Article 50 of the AI Act, which requires AI providers to mark synthetic content in a machine-readable format. For systems that were already on the market before August 2, the deadline is December 2, 2026. Anthropic and Google had already announced their own watermarks for Claude and Gemini, so OpenAI was the missing piece.

The numbers worth knowing

OpenAI has published its own tests, and they are more honest than these announcements usually are:

On 400-token passages, the detector finds the watermark in about 95% of cases. A token is a piece of a word, so we are talking about a few long lines of text.
On 200-token passages, it drops to 80%. The shorter the text, the weaker the trace.
If you swap 10% of the words for synonyms, it falls to 66%. Swap 25% and it falls to 17%.
It works worse on technical content such as maths, because the model has little room to choose words there.
The detector is not public. Only approved researchers and organizations can use it, case by case.
In the API it is optional and off. Anyone integrating OpenAI models into their own product has to switch it on manually in the organization or project settings.

Why it matters, even if you can't see it

The first thing we were asked this week was whether it will now be obvious that a text was written with ChatGPT. The short answer is no, at least for now.

Neither your customers, nor your competitors, nor Google have access to the detector. And the watermark largely disappears as soon as someone edits the text properly. OpenAI says it clearly: the watermark does not measure how much a person contributed, does not prove authorship or ownership, does not identify the user and does not say whether the content is accurate.

What does change is the context. From December, the main assistants on the market will mark what they generate in Europe by default. That opens the door to platforms, media or public bodies asking for detectors, and to telling generated text apart becoming normal in a few years. Today it isn't.

What it means for a small business

It depends on how you use AI. We split it into three cases, because the consequences are very different.

If you use ChatGPT to write your website copy, emails or social posts. That text will carry the watermark if you paste it as is. It isn't illegal and it doesn't create any new obligation for you as a user. What we do recommend is what already made sense before: edit. Text that goes through the hands of someone who knows the business is better text, and it also loses most of the trace.

If you are worried about SEO. Since 2023 Google has said it doesn't penalize content for being made with AI, only for being unhelpful. The watermark doesn't change that. What ranks, both in Google and in the answers of ChatGPT or Perplexity, is still having your own data, real experience and a clear structure. We cover it in detail in how to get ChatGPT to recommend your business.

If you have built a product with the OpenAI API. This is the case we care about most, and the one almost nobody is looking at. A chatbot, a product description generator or a tool that writes reports for your clients can make you the provider of an AI system under the AI Act. And in the API the watermark is off. If Article 50 applies to you and you don't switch it on, the problem is yours, not OpenAI's.

What I would do this week

None of this needs a big project. These are three checks that fit into a morning:

Make an inventory of where you use the API. Not ChatGPT in the browser, but integrations: chatbots, automations in n8n or Make, store plugins, internal tools that generate text for clients.
For every integration that generates content for third parties, decide whether you are the provider. If the answer is yes, or you aren't sure, switch on the watermark in the project settings before December 2. It is a setting, not a development.
Review your content process. If ChatGPT text goes live on your website without anyone editing it, the watermark isn't the problem. The problem is that the text probably adds nothing another site doesn't already say, and that is what really costs you visibility.
Explain it to your team. A five-line internal note stops someone from panicking, or from paying for a tool that promises to remove the watermark.

What we think

We think it's good news, with caveats. It makes sense for generated content to carry a trace, and it is better for providers to add it than for every company to invent its own system.

But it shouldn't be overrated. With 25% of the text edited, detection drops to 17%, so it won't catch anyone who wants to hide. It works as a default signal, and that is already something.

What does worry us is the gap between ChatGPT and the API. The big platforms comply for you. The integrations a small business builds with its trusted provider don't, and that is where December may bring surprises. It is exactly the kind of technical decision we review as an external CTO.

If you want us to look at it with you

We design and maintain AI agents and automations for small businesses, and we work on visibility in search engines and assistants through SEO and GEO. If your team uses AI every day and you want them to do it with good judgment, we also offer team training.

Book a thirty-minute call. We go through your AI integrations with you, which ones could fall under Article 50 and what needs to be switched on before December 2. You leave with a short, concrete list. No commitment.

Sources

This article is based on the following publications, consulted on October 7, 2026:

Frequently asked questions

Yes, in the European Union. On October 5, 2026 OpenAI announced that over the coming weeks it will apply the textGrain watermark to text generated by ChatGPT and Codex for EU users, on every plan. It is invisible: it adds no stamp, just a statistical trace in the choice of words.

Not for the general public, for now. OpenAI's detector is only available to approved researchers and organizations, case by case. Its accuracy also drops sharply with short or edited text: according to OpenAI, it goes from 95% on untouched 400-token passages to 17% if 25% of the words are changed.

Nothing suggests it does. Since 2023 Google has maintained that it values how useful content is, not how it was produced, and it has no access to OpenAI's detector. What hurts rankings is generic content that adds nothing of its own, however it was written.

Only if you switch it on. In the API the watermark is optional and off by default worldwide. It is enabled in the organization or project settings. If you integrate the API into a product that generates content for third parties, it is worth checking whether Article 50 of the AI Act requires you to mark it.

Article 50 of the AI Act has applied since August 2, 2026. For systems already on the market before that date, the obligation to mark content in a machine-readable format applies from December 2, 2026. New systems must comply from the moment they launch.

As a general rule, using ChatGPT to draft your website copy or emails doesn't require a notice. The obligation to mark falls mainly on the provider of the AI system. There are specific cases, such as text published to inform the public on matters of public interest without editorial review, where additional obligations may apply, and those are worth checking.

AI ActChatGPTwatermarkOpenAIAI contentSEO

View all articles

ChatGPT now watermarks its text in Europe: what changes for your business and what doesn't