OpenAI will add invisible watermarks to ChatGPT text in the EU to meet AI Act rules
The watermark, called textGrain, hides a statistical pattern in the model’s word choices. It will roll out to ChatGPT and Codex users in the European Union over the coming weeks; API developers worldwide can switch it on, but it is off by default.
The announcement: OpenAI said Monday it will begin adding an invisible watermark to text generated by ChatGPT and Codex in the European Union to comply with the EU AI Act, TechCrunch reported. The act’s transparency rules, which took effect Aug. 2, require AI companies to mark AI-generated content in a way other systems can identify. The rollout covers eligible users on all plans in the EU only; OpenAI said it is not making text watermarking a global default.
How it works: There is no visible symbol. The system nudges which words the model picks, leaving a pattern people can’t see but a detector holding a secret key can find, according to an OpenAI technical report co-written with researchers from the University of Pennsylvania and Yale. Because the signal is in the words themselves, it survives copy and paste. OpenAI said the watermark does not identify the user and that it saw no meaningful change in model performance with it on.
Its limits: Editing weakens it. In OpenAI’s tests, swapping 10% of words for synonyms cut detection on 400-token passages from about 92% to 66%, and replacing a quarter of the words pushed it below 20%, The Decoder reported. Short passages, math answers and translated text are also harder to detect. OpenAI cautioned that a missing watermark “does not prove human authorship,” and that a detected one says nothing about how much human editing or judgment went into a passage.
Who can check: For now, detector access goes only to approved researchers and expert organizations, which OpenAI said it will grant case by case. The tool will report only whether an OpenAI watermark was found, The Decoder reported.
Context: Anthropic said two months ago it would watermark Claude-generated text worldwide, a move that drew pushback from some users, TechCrunch noted. Unlike Anthropic’s mandatory approach, OpenAI’s API watermarking is opt-in, The Decoder reported.