OpenAI Watermarks ChatGPT EU Outputs by Default

The EU AI Act Sets the Default

The decision responds directly to the EU AI Act, which took effect in August. The law requires that AI-generated content be detectable by other tools, and OpenAI’s regional default is the result.

Rivals moved earlier. Anthropic introduced text watermarking for its models in August and switched it on globally, while OpenAI is limiting its default to the EU for the ChatGPT and Codex apps. In OpenAI’s API, watermarking stays off by default but can be enabled.

TextGrain Works Best on Untouched Text

OpenAI’s method is called textGrain, and the company has published a technical paper describing how it works. Like other LLM watermarking tools, it embeds patterns in word choices that humans cannot notice but a specialized key can detect.

Detector access will start with a limited group of researchers and organizations, with an approval process for later requests. OpenAI’s own tests show how quickly performance drops once text is changed:

• 92 percent detection under ideal conditions
• Altering just 10 percent of the text cuts success by almost 30 percent
• Changing 20 percent of the text lowers detection by nearly 75 percent
• Shorter or translated texts also show lower success rates

Existing standards such as SynthID and the C2PA project are also relatively easy to circumvent, and textGrain faces the same basic limits. There is still no completely reliable way to mark AI content.

Hashlytics Take: Compliance Is Not the Same as Detection

The 92 percent figure describes text nobody has touched, which is rarely how AI writing circulates. A light rewrite, a translation, or a few swapped sentences is enough to push detection toward coin-flip territory on OpenAI’s own numbers. Add a detector that only a small circle of approved researchers can use, and the public has no easy way to check anything. The EU rule gets satisfied on paper, but the watermark is better read as a regulatory checkbox than a tool for spotting AI text in the wild. Whether it becomes more than that depends on how robust the next version of the method turns out to be.