AI & Tech · National
OpenAI will watermark ChatGPT text by default in the EU, while the feature stays off by default in other regions
OpenAI says it will automatically watermark ChatGPT text in the European Union to meet the EU AI Act and offer the feature, off by default, elsewhere. Ars Technica notes such marks are often easy to get around.
Listen to this brief · 4 min
Read aloud by an AI voice from the text below. Download MP3
What happened
OpenAI has announced that it will begin automatically watermarking text produced by ChatGPT in the European Union, Ars Technica reported. The company will also make the watermarking feature available in other regions, but there it will be switched off by default.
The difference comes down to regulation. According to Ars Technica, the European rollout is driven by the need to comply with the EU AI Act, which took effect in August. As a result, ChatGPT will have different default settings depending on where it is used. More coverage of these issues is in our AI and technology section.
Why Europe gets it first
What the EU AI Act requires
Ars Technica reported that the law requires content produced by AI models to be marked in a way that another tool can detect. In other words, the mark has to be something that detection software can find, not just a label a person reads.
Off by default elsewhere
Outside the EU, OpenAI will offer the same watermarking feature, but it will not be active unless it is turned on. Ars Technica's report does not describe a regional law driving the feature outside Europe.
How textGrain works
A proprietary method
OpenAI calls its approach textGrain, and the method is proprietary, according to Ars Technica. The company has published a technical paper explaining how the system works.
Hidden patterns in word choice
Ars Technica said textGrain generally works like other watermarking tools for large language models. It places patterns in word choices that a human reader would not notice and that do not meaningfully change the overall quality of the output. Someone with a key can then use a specialized detector to find those patterns.
Who gets the detector
OpenAI says it will give a limited number of researchers and organizations access to the detector. The company also plans a request-for-approval process so that others can be added over time.
The limits of AI watermarks
Ars Technica noted that there is still no completely effective and reliable way to mark AI-generated content. A few standards already exist, including SynthID and the C2PA project. The outlet described them as relatively easy to get around for anyone with basic know-how.
The outlet said the same is likely true of OpenAI's watermark. That assessment comes from Ars Technica, not from OpenAI, whose published technical paper gives the company's own explanation of how textGrain operates.
What it means for ChatGPT users
For people using ChatGPT in the EU, text generated by the service will carry the watermark automatically. In other regions, the feature will be available but inactive by default.
Because the patterns are designed to go unnoticed by human readers, the watermark is not something a person would spot on the page. Finding it requires the key and the specialized detector, which OpenAI is limiting to approved researchers and organizations for now.
What to watch next
- The start of OpenAI's automatic watermarking of ChatGPT text in the European Union.
- How OpenAI runs its request-for-approval process for adding researchers and organizations to detector access over time.
- The rollout of the off-by-default watermark option in regions outside the EU.
