Text Watermarking in ChatGPT, Codex, and the API

OpenAI is adding an invisible marker to AI-generated text, rolling out for ChatGPT and Codex in the EU and available as an opt-in for API customers worldwide.

Details

Hides a statistical signal in text as the model writes it. A detector can look for that signal, and no visible marks or special characters are added.
It only answers whether text was likely generated by an OpenAI model. It doesn't identify the author, account, organization, conversation, or prompt.
Rolls out to eligible ChatGPT and Codex text in the EU over the coming weeks to comply with the EU AI Act.
API customers anywhere can turn it on for select models today.
OpenAI says watermarking hasn't changed model speed, capability, or how responses read in its testing, and notes that current text watermarking has significant limits.