OpenAI shared implementation plans for EU AI Act text provenance requirements.
Starting October 5, API customers worldwide can opt in to text watermarking for select models. Over coming weeks, invisible watermarking rolls out to eligible ChatGPT and Codex users in the EU only. Watermarking will not become a global default at launch.
The technology, textGrain, adds a statistical signal through word choices. A detector looks for this signal, but short passages and later edits reduce reliability. Detector access initially limited to approved researchers and expert organizations.
The watermark can indicate OpenAI-generated content but cannot identify users, measure human contribution, determine ownership, or verify accuracy. No detected watermark does not prove human authorship.
OpenAI published a technical report and plans open-source release so others can build on the technology.