OpenAI will watermark ChatGPT and Codex text in the EU, and opens opt-in watermarking to API customers worldwide

OpenAI is adding an invisible watermark, which it calls textGrain, to ChatGPT and Codex text in the European Union over the coming weeks. It announced the plan today, as its answer to the EU AI Act rule that generated text be identifiable by machine.

WHAT CHANGES, AND WHERE

- ChatGPT and Codex: eligible text output in the EU only, on all plans, over the coming weeks. OpenAI says it is not making watermarking a global default at launch. - API: from today, API customers anywhere can opt in to watermarked output for select models. It stays off by default. - The detector: applications open today, but access goes only to approved researchers and expert organizations, case by case. There is no public detector.

HOW WELL IT WORKS, BY OPENAI'S OWN NUMBERS

textGrain hides a statistical signal in the model's word choices. At a 1% false-positive target, OpenAI says its detector caught about 80% of watermarked 200-token passages and about 95% of 400-token ones on psychology answers, and substantially fewer on maths, where word choice is tighter. Editing weakens it: swapping 10% of the words for synonyms cut detection on 400-token passages from about 92% to 66%, and swapping 25% cut it to 17%.

DOES IT HURT QUALITY

OpenAI's own table shows GPT-6 Astra (max) scoring about the same with the watermark as without, for example 49.76 against 49.57 on the Artificial Analysis Intelligence Index and 93.94% against 94.44% on GPQA Diamond.

WHAT WE DO NOT KNOW

Which API models support it, the exact EU start date, and which ChatGPT and Codex outputs count as eligible. Every figure here is OpenAI's; no independent test has been published.