OpenAI Begins Phased Text Watermarking Under EU AI Act Rules

From October 5, 2026, API customers worldwide can opt in to text watermarking for select models, with the setting off by default. In the EU, invisible textGrain watermarks will be added to eligible ChatGPT and Codex output across all plans in the coming weeks, while detector access stays limited to approved researchers.

• API customers opt in to watermarking; it stays off by default
• EU rollout covers eligible ChatGPT and Codex users on all plans
• At 1% false positives, detector finds marks in ~80% of 200-token passages
• Replacing 25% of words with synonyms cuts detection from 92% to 17%