OpenAI Rolls Out EU Text Watermarking, Opens Opt-In API Access for Select Models
Updated
Updated · OpenAI · Oct 5
OpenAI Rolls Out EU Text Watermarking, Opens Opt-In API Access for Select Models
3 articles · Updated · OpenAI · Oct 5
Summary
OpenAI said eligible ChatGPT and Codex text in the EU will gain invisible watermarks over the coming weeks, while API customers worldwide can opt in starting today for select models.
The move answers the EU AI Act’s machine-readable provenance rule, but OpenAI kept watermarking off by default and limited detector access to approved researchers and expert groups.
OpenAI’s textGrain detector found about 80% of 200-token passages and about 95% of 400-token passages at a 1% false-positive target, with weaker results for constrained text such as mathematics.
Editing sharply reduced detection: in 400-token tests, replacing 10% of words cut detection to 66% from about 92%, and replacing 25% dropped it to 17%.
OpenAI said watermarking did not meaningfully affect Astra benchmark performance and plans to open-source the technology as part of a broader provenance system spanning text, images and audio.