OpenAI will begin adding an invisible watermark to texts generated and processed by ChatGPT and Codex in the European Union. The introduction will be phased in over the coming weeks and will cover users of all plans.
The company introduced the technology as part of fulfilling the requirements of the European Artificial Intelligence Act, which stipulates that content created by generative systems can be recognized in a machine-readable way.
How the tagging will work
The technology is called textGrain. It does not place a visible mark in the text and does not add hidden characters, extra spaces or unusual punctuation. Instead, it changes the statistical model by which the system chooses between possible words and parts of words. A special detector can analyze this pattern and check if the text contains a signal from OpenAI.
“The watermark does not add hidden characters, extra spaces, or unusual punctuation“, OpenAI explained.
The company points out that the tagging is part of a broader approach to tracing the origin of content. For images and audio, OpenAI uses other technologies and metadata, but for text, it chooses a separate method that is consistent with the way language models generate responses.
Limitations remain
OpenAI will not release the watermark detection tool publicly at launch. Access will be granted to approved researchers and specialized organizations, who will evaluate the accuracy of the system and work on improving it.
The reason is the limitations of the technology. Short texts are harder to recognize, and editing, translation, or significant processing can weaken the signal. Both false positives and missed detections are possible.
The watermark itself does not indicate who used ChatGPT, how much of the text was created by the model, or how much human input was contributed. It does not prove authorship, ownership, or legal responsibility. Similarly, the absence of a signal does not prove that the text was entirely written by a human.
Separate mode for API clients
Starting October 5, OpenAI API clients around the world will be able to optionally enable text markup for specific models. The feature remains off by default, meaning that developers and companies will have to decide for themselves whether to enable it in their applications.
OpenAI plans to extend watermarking to services that use its models through cloud partners. European input for ChatGPT and Codex will remain regional, with no automatic global activation of the feature.