OpenAI will start watermarking ChatGPT's text in the EU

OpenAI will add an invisible watermark to text generated by ChatGPT and Codex for users in the European Union to comply with the EU AI Act’s transparency requirements. The watermark, implemented as a technique called textGrain and described in a technical report co-authored with researchers from the University of Pennsylvania and Yale, subtly biases model word-choice patterns so a party with a secret key can detect AI-generated passages. The feature will roll out to eligible EU users across plans over several weeks; developers using OpenAI’s API anywhere can opt in for select models today, but the setting is off by default and OpenAI is not enabling it globally at launch. The company reports no meaningful performance impact and emphasizes the watermark does not identify individual users.

Tests show the watermark is detectable in many cases but has clear limits: replacing as little as 10% of words with synonyms reduced detection rates from about 92% to 66%, and short snippets, math answers, and translations are harder to flag. Because the signal is text-based it travels with copy-pasted material, yet a missing watermark does not prove human authorship - text may be heavily edited or generated by another system. Access to the detector is initially restricted to approved researchers and expert organizations. The move follows Anthropic’s earlier, global watermarking of Claude and aligns with several major AI firms’ commitments to the EU code of practice.