OpenAI is preparing to mark ChatGPT-generated text automatically for users in the European Union, a regional shift shaped by the EU AI Act and the wider pressure to make AI output detectable. The same watermarking feature will be available outside the EU, but it will not be enabled by default there.
The change puts a spotlight on a hard problem: labeling AI-generated text in a way that machines can later recognize, while keeping the writing normal for human readers. OpenAI’s answer is a proprietary method called textGrain, but the company’s own testing shows that this kind of watermark is useful without being definitive.
What OpenAI is changing
OpenAI will begin automatically watermarking text generated with ChatGPT in the European Union. Outside the EU, the feature will be offered, but it will stay off unless enabled.
The company is taking that different approach because of the EU AI Act, which took effect in August. The source describes the law as requiring AI-produced content to be marked in a way that another tool can detect.
That makes the EU the place where default watermarking becomes part of the standard ChatGPT experience first. In other regions, OpenAI is treating the same capability as optional rather than automatic.
The rollout will apply at least in the ChatGPT and Codex apps. For the API, the watermarking feature appears set to remain off by default while still being available as an option.
How textGrain is meant to work
OpenAI calls its watermarking method textGrain. The system is proprietary, and the company has published a technical paper explaining it.
At a high level, the approach follows the pattern used by other large language model watermarking tools. Instead of adding a visible label, it places signals inside word choices. Those signals are not meant to stand out to a human reader, and they are not supposed to meaningfully reduce the general quality of the output.
The watermark is intended for a specialized detector. Someone with access to the right key and detection tool can look for the pattern and decide whether the text likely came from a watermarked model output.
OpenAI says detector access will initially go to a limited number of researchers and organizations. Others will be able to seek access through a request-for-approval process over time.
Why detection is still fragile
The most important limit is that watermarking text is not the same as proving origin with certainty. The source notes that there is still no completely effective and reliable method for this problem.
OpenAI’s own tests show why. In the best case, textGrain reached a 92 percent successful detection rate. That is meaningful, but it is not perfect, even before anyone tries to alter the text.
The performance drops when the generated text is changed. OpenAI’s tests show that altering just 10 percent of an output reduces successful detection by almost 30 percent. Changing 20 percent of the text can lower the success rate by almost 75 percent.
Shorter and translated text also tend to be harder to detect than longer text. That matters because many real uses of ChatGPT involve brief passages, summaries, edits, or text that moves across languages.
The practical takeaway is straightforward: textGrain can make AI-generated text easier to identify in some cases, but it should not be treated as a guaranteed signal. A watermark can help with detection, but it can also be weakened by ordinary editing.
How this compares with other watermark efforts
OpenAI is not alone in trying to watermark AI-generated content. The source names SynthID and the C2PA project as existing standards, while also noting that such systems can be relatively easy to circumvent for anyone with basic know-how.
The same broad concern applies to textGrain. Its signals live inside wording choices, so changes to the wording can reduce the detector’s ability to recognize the pattern.
Anthropic, a competitor to OpenAI, introduced watermarking for text generated by its models in August. The difference is in deployment: Anthropic enabled watermarking globally, while OpenAI is making it the default only where regulators require it, at least for now.
That contrast shows two paths for AI watermarking. One approach treats watermarking as a global baseline. The other ties default use to regulatory requirements while leaving other regions and the API with more optional control.
What happens next
OpenAI says the EU rollout will happen “in the coming weeks,” which means users in the European Union should see the default change first.
For users outside the EU, the key difference is choice. The watermarking feature will exist, but it will not automatically be turned on by default.
For anyone relying on AI detection, the limits matter as much as the rollout. Watermarking may become more common, especially where rules require machine-detectable labels, but OpenAI’s numbers show that edited, shorter, or translated text can still complicate detection.
That makes ChatGPT watermarking a useful compliance tool, not a final answer to the broader question of how to identify AI-generated writing reliably.