OpenAI starts adding an invisible watermark to ChatGPT text in the EU

On October 5, OpenAI announced that it is adding an invisible watermark to ChatGPT and Codex text. According to TechCrunch and BleepingComputer, it starts only in the European Union, for eligible users on all plans, rolling out over several weeks. Developers can turn it on themselves through the API in any country; it is off by default. As TechCrunch reports, the reason is the EU AI Act, which requires AI-generated content to be marked.

How it works

The system is called textGrain. These are not hidden characters: the model slightly favors certain words, which leaves a statistical pattern in the text. You can't see it, but a detector with the key can find it. The mark survives copy and paste, but it does not reveal the author, the account or the prompt.

Where it weakens

OpenAI describes the limits itself. In its test on 400-token passages, replacing 10% of words with synonyms cut detection from 92% to 66%, and replacing 25% cut it to 17%. Length matters: at a 1% false-positive target, the mark was found in about 80% of 200-token responses and about 95% of 400-token ones. Math, translated and heavily edited text are the hardest cases.

Also, a missing mark does not prove a human wrote the text. And the detector is currently open only to approved researchers and expert organizations, not to the public.

What it means

For an ordinary writer the practical impact is small. If you edit the text and write in your own voice, the mark weakens, and there is hardly anyone to check it anyway. It is more a step toward regulators than a tool for catching AI. But the direction is clear: per TechCrunch, Anthropic announced the same for Claude two months earlier, so marking is becoming the norm. A sensible stance: use AI as an assistant, and stay responsible for the meaning and the edit.

Sources: TechCrunch, BleepingComputer.