OpenAI Details Its EU Text-Provenance Approach, Rolling Out Invisible 'textGrain' Watermarking
OpenAI detailed its approach to complying with EU AI Act text-provenance requirements, centered on a new invisible text watermarking technology called “textGrain.”
Details
- Phased rollout: API customers worldwide can opt into watermarking for select models starting now; over the coming weeks, ChatGPT and Codex will add watermarks to eligible text for users in the EU specifically, not globally
- Detector access: approved researchers and expert organizations can apply for limited access to a watermark-detection tool
- Disclosed limitations: detection accuracy drops sharply on short text (about 80% for 200-token passages vs. 95% for 400-token passages); heavy editing significantly weakens the watermark, with replacing 25% of words cutting detection to roughly 17%; mathematical content is harder to watermark reliably than flexible prose
- Explicit caveats: OpenAI states watermarks cannot measure human contribution, establish ownership, identify individual users, or verify factual accuracy — “the absence of a detected watermark does not prove human authorship”
- Framing: positioned as one layer within a broader content-provenance strategy rather than a complete solution
What happened next
The release extends OpenAI’s pattern of pairing new compliance-driven technical capabilities with candid disclosure of their limitations, similar in tone to its safety cases framework and incident postmortems — and gives EU regulators and researchers a concrete, if imperfect, tool for text provenance ahead of stricter AI Act enforcement deadlines.