OpenAI is implementing a phased approach to text provenance in response to the EU AI Act, introducing invisible watermarks to eligible ChatGPT and Codex outputs in the European Union while offering an opt-in option for global API customers. The company released its textGrain technology, which adds statistical signals to word choices, alongside limited access to a detector for approved researchers.
- Text watermarking is off by default globally but will be added invisibly to EU ChatGPT and Codex users in the coming weeks.
- Global API customers can opt in to watermarked outputs for select models starting today.
- The textGrain detector is available initially only to approved researchers and expert organizations to evaluate reliability.
- Evaluations show detection rates drop significantly with editing, such as synonym replacement reducing detection from 92% to 66% for 10% word changes.
- Watermarks do not measure human contribution, establish ownership, identify users, or verify accuracy.
OpenAI emphasizes that current watermarking technology has limitations regarding editing and translation, necessitating a cautious rollout focused on transparency about what the signals can and cannot prove.