Our approach to EU text provenance rules

OpenAI has launched textGrain, its text watermarking technology, in response to EU AI Act requirements that generative AI providers make AI-generated text machine-identifiable. Starting immediately, API customers globally can opt in to text watermarking for select models, while ChatGPT and Codex outputs in the EU will receive invisible watermarks by default over the coming weeks. The technology embeds a statistical signal in word-choice patterns, but carries material limitations: at a 1% false positive rate, detection hits roughly 80% for 200-token passages and drops sharply for constrained content like mathematics; replacing 25% of words reduces detection from ~92% to 17%. Detector access is being restricted initially to approved researchers and expert organizations given these reliability constraints, and the watermark does not identify users, establish ownership, verify accuracy, or measure human contribution. OpenAI plans to open-source textGrain and will expand detector access as performance evidence and regulatory standards mature.