RegulationChatGPT text in the EU will carry an invisible watermark

OpenAI said on October 5 it is extending content provenance from images and audio to text to meet the EU AI Act. The model slightly adjusts the randomness in its word choices so a detector can find a pattern across a passage; no hidden characters or spaces are added.

API customers worldwide can opt in, and eligible ChatGPT and Codex output in the EU gets it in the coming weeks. OpenAI acknowledged the limits: short or edited text is hard to detect, and swapping a quarter of the words cuts detection from 92% to 17%. The detector is open only to approved researchers for now.