Texts from ChatGPT and Codex will soon receive a tag in the EU that is invisible to everyone but technically detectable. OpenAI is thus responding to the labeling requirement of the AI Act – while simultaneously acknowledging how easily the signal can be weakened.
OpenAI, the AI research and practical application company, has outlined in a statement on EU rules for text origin how it will label AI-generated texts in the future. This is in response to Article 50 of the AI Act, whose transparency obligations have been in effect since August 2, 2026.
Afterwards, providers of generative AI must mark their output in a machine-readable way so that it is recognizable as artificially generated. For text, OpenAI uses a watermark embedded in the word choice itself. This will apply to all ChatGPT plans, including the free plan.
Key Facts at a Glance
- In the coming weeks, texts from ChatGPT and Codex in the EU will receive an invisible watermark, across all tariffs.
- Outside the EU, ChatGPT remains without a watermark; this also applies to Switzerland.
- Customers worldwide can now voluntarily activate a watermark in the API; it remains off by default.
- Initially, only selected researchers and professional organizations will receive the detector that recognizes the watermark.
This is how textGrain works
The method is called textGrain. It inserts an invisible statistical signal into the model's word choice: where several formulations fit equally well, the selection follows a hidden pattern. A detector searches a text for precisely this pattern.
According to OpenAI, this has no significant impact on the quality of the answers. In eight benchmarks of the current Astra model, the results with and without watermarking are very close, sometimes slightly higher, sometimes slightly lower. OpenAI also intends to release the technology as open source; a technical report already describes it.
How reliable is the detection?
OpenAI outlines the limitations of its method, both internally and numerically. With a false positive rate of one percent, the detector identified the watermark in approximately 80 percent of passages containing 200 tokens and in approximately 95 percent of passages containing 400 tokens, based on texts related to psychology. In mathematics, where word choice allows for less leeway, the figures were significantly lower.
Post-processing has an even stronger effect:
| Replaced words (400 tokens) | Recognition rate |
|---|---|
| none | approximately 92 % |
| 10 % by synonyms | approximately 66 % |
| 25 % by synonyms | approximately 17 % |
Furthermore, a watermark says nothing about how much a person contributed to the text, who owns it, who created it, or whether it is accurate. Conversely, its absence does not prove that a person wrote the text.
Not the first provider
With this move, OpenAI is following a similar approach to Anthropic. The provider of Claude had announced in the summer that it would mark texts using slightly modified word choices.
This does not create a shared detector. OpenAI's tool only checks whether a text contains an OpenAI watermark – it does not recognize texts from other models as AI-generated.
A short-range signal
I consider the watermark in its current form primarily a legal obligation, not a tool for reliably detecting AI-generated texts in everyday life. Replacing a quarter of the words is enough to reduce the detection rate to less than a fifth, and the detector isn't public anyway.
For users in Germany and Austria, nothing visibly changes in ChatGPT; the watermark is solely due to the wording. In Switzerland, which is not part of the EU, texts remain without a watermark.
Do you think a watermark in AI-generated texts makes sense if it can be circumvented simply by changing a few words? Let us know in the comments where you would draw the line.
