OpenAI textGrain Watermarks EU Text, Opens API Opt-In

textGrain is a pragmatic compliance feature, but its signal should be treated as provenance evidence—not definitive authorship proof.

• –EU ChatGPT and Codex users will receive watermarking across eligible plans over the coming weeks; API watermarking remains opt-in and off by default.
• –Detection weakens with short, constrained, translated, or heavily edited text, making false positives and false negatives unavoidable.
• –The watermark does not identify users, measure human contribution, establish ownership, or prove that unmarked text was written by a person.
• –Developers should combine watermark results with generation logs, disclosure policies, and other provenance systems rather than treating textGrain as a standalone audit mechanism.
• –OpenAI plans to open-source the technology, which could help establish broader interoperability for AI-text provenance.