ChatGPT will soon sign all the texts it writes on your behalf.

On Monday, October 5, OpenAI announced the introduction of an invisible watermark in texts generated by ChatGPT and Codex within the European Union.“In the coming weeks,” OpenAI will incorporate “an invisible watermark into eligible ChatGPT and Codex text in the European Union”, the company wrote in a blog post.

This significant new feature will apply to all subscriptions, whether free or paid, but will apply only in the European Union. OpenAI clarifies that the text watermark is not intended as “a global default setting” at this stage. According to the American startup, this approach “allows room to learn about real-world uses and feedback”. Before introducing the “watermark” worldwide, the AI giant wants to first see how it goes in Europe.

We’re expanding our approach to content provenance to include text in response to EU regulatory requirements, while recognizing the significant limitations of current text watermarking technology.

Our tools already help verify whether an image or audio file was created with our models.This work builds on those efforts to help people better understand when content may have been generated or edited with an OpenAI model.

In the EU, we’ll start watermarking eligible text from ChatGPT and Codex over the coming weeks to comply with the EU AI Act.

Customers using our API can turn on text watermarking for select models worldwide today.

— OpenAI (@OpenAI) October 5, 2026

A hidden signature in AI-generated texts

Named textGrain, the technology used by OpenAI “adds an invisible statistical signal to the model’s word choices,” explains the company. A detector then only needs to search for the statistical signal to assess whether a passage bears ChatGPT’s signature or comes instead from a human mind. There will therefore be no hidden code or symbol in the texts.

When ChatGPT writes a sentence on behalf of its users, it does not choose a word at random. At each step, it weighs several suitable words and chooses the most likely one. With the watermark, OpenAI adds a secret rule to the way ChatGPT chooses the words that will make up its sentences. This secret key, held solely by OpenAI, prompts the AI to choose certain words over others, without changing the meaning or quality of the text. This rule ultimately creates a kind of invisible signature that a detector can identify. The startup finds “no significant difference” in quality in texts marked with the watermark.

AI watermarking that is far from infallible

OpenAI claims that textGrain has “matched or surpassed” other approaches for labeling AI-generated content, including SynthID, Google’s solution. The company hastens to add that “good performance under ideal conditions does not guarantee reliable detection in everyday use”.

OpenAI also admits that the system is not infallible. The shorter the text, the more difficult detection becomes. OpenAI designed its detector to be very rarely wrong, but once in every hundred cases, it falsely claims that human-written text was generated by ChatGPT. On a passage of approximately 150 words, the detector picks up the watermark 80% of the time. On a passage twice as long, approximately 300 words, the detection rate rises to 95%.

Not surprisingly, the slightest edit made by a human makes watermark detection more difficult. In fact, “replacing 10% of words with synonyms reduced detection from approximately 92% to 66%”, OpenAI acknowledges. Translating the source text also makes AI detection more difficult. Circumvention tools already promise to erase watermarks that work in the same way.

In practice, “the absence of any detected watermark does not prove that a text was written by a human”, she warns.The text may indeed be too short, too heavily edited by a human, translated, or may have come from another AI. For reference, online AI detectors, such as the controversial Pangram, are also far from infallible. As for privacy, OpenAI assures that the watermark “does not identify the user” and does not associate any account, query, or conversation with the text generated by ChatGPT.

Also read: On LinkedIn, you’ll finally be able to report all those unbearable AI posts

A detector reserved for researchers

At first, the watermark detector will be reserved for researchers and specialized organizations. They will have to help OpenAI assess the reliability of the system, most likely with a view to a public rollout. Developers, for their part, can activate the watermark now on certain AI models, anywhere in the world. It will, however, remain disabled by default.

OpenAI is not the first company to take the step. Two months ago, Anthropic had already announced that it was going to watermark all texts generated by Claude. Mirroring its rival, Anthropic specified that the tool has its “limits” and refused to speak of irrefutable proof in the event of a positive result.

A requirement of the AI Act

By watermarking all texts generated by ChatGPT, OpenAI complies with the requirements of the AI Act, the European regulation on AI. Since August 2, Article 50 requires providers of generative AI to make all their texts identifiable by a machine.OpenAI, Anthropic, Microsoft, Google and Meta have until December 2, 2026 to comply with regulations and avoid a fine.

The arrival of textGrain comes as the rise of artificial intelligence makes waves throughout society, including in French literary circles. After receiving several prestigious awards for his novel “It Was That or Die,” writer Thélyson Orélien was accused of having extensively used artificial intelligence. Despite his denials, the Académie Goncourt removed the book from its initial selection “to preserve the integrity of the prize”.

👉🏻 Follow tech news in real time: add 01net to your sources on Google, and subscribe to our WhatsApp channel.