OpenAI said on 5 October that it is adding text watermarking, citing the EU AI Act, which requires machine-identifiable AI text. API customers worldwide can opt in for select models now; it stays off by default. Over the coming weeks, eligible ChatGPT and Codex text in the EU gets an invisible mark. Approved researchers and expert groups can apply for the detector.
OpenAI says textGrain adds a statistical signal to word choices, which its detector looks for. The model leans toward certain words and the detector counts them, so more words and freer wording should give a stronger signal. OpenAI's own tests used ELI5 questions and a 1% false positive limit. It found marks in about 80% of 200-token psychology passages and about 95% of 400-token ones. Maths, with less freedom in wording, scored substantially lower. In 400-token English passages, swapping 10% of words for synonyms cut detection from about 92% to 66%; 25% cut it to 17%. Swaps remove the very words that carry the signal.
This is a compliance step, not a checking tool for buyers: the mark is opt-in on the API, EU-only in the apps and OpenAI-only, and the detector is closed to the public. Ask which of your text flows are covered, who can verify marks, and what editing does to them. Cloud-partner support is still pending.