What's this about?
OpenAI's textGrain changes statistically equivalent word choices to embed an invisible signal. EU ChatGPT and Codex output will use it by default for eligible text; API users can opt in globally.
What exactly is a statistical watermark?
It is not visible metadata. The generator slightly favours particular acceptable words, and a detector later looks for that distributed pattern across a long passage.
What's the community saying?
OpenAI's announcement drew 5,833 likes, 401 reposts and 673 replies.
- Editing and paraphrasing can rapidly erase the signal.
- False positives are serious enough that detector access initially goes only to approved experts.
- Developers asked why EU consumer output is automatic while the API remains opt-in worldwide.