OpenAI announced a new text-provenance rollout on October 5, 2026. Over the coming weeks, it will add an invisible watermark to eligible ChatGPT and Codex text output in the European Union across all plans. API customers worldwide can opt in on select models starting the same day, with watermarking off by default for API output. Support through cloud partners is scheduled for the coming weeks. Source
EU products and the API use different rollout models
For eligible ChatGPT and Codex output in the EU, OpenAI describes automatic watermarking rather than a setting each user turns on. The announcement does not list every eligible model or an account-by-account launch date; it gives a phased window of the coming weeks. It also gives no timetable for automatic use in ChatGPT or Codex outside the EU. Source
The API starts as an optional feature for customers worldwide on supported models. OpenAI says the feature is available on select models but does not provide model-specific setup steps in the announcement. API customers need to check their account and current documentation for the supported model and activation option. Source
The signal is placed in word choices
OpenAI calls the system textGrain. It adjusts generation so that word choices carry an invisible statistical signal, and a separate detector evaluates whether that signal is present in submitted text. OpenAI also says it plans to release the technology as open source. Source
The detector is not launching as a public lookup tool. OpenAI is accepting applications and plans to grant initial access case by case to approved researchers and expert organizations. Its existing public tools for image and audio provenance, openai.com/verify and the Content Provenance API, remain available. Source
Length and editing change the detection rate
At a targeted 1% false-positive rate, OpenAI reports detecting about 80% of 200-token passages and 95% of 400-token passages in psychology content. A token is a unit a model uses to process text. Performance was lower for mathematics content, and short passages can carry too little signal for confident detection. Source
Editing reduced detection sharply in OpenAI’s tests. For a 400-token passage, replacing 10% of the words reduced detection from about 92% to 66%. Replacing 25% reduced it to 17%. Translation, rewriting, passage length and whether the model supports the watermark can therefore affect the result. Source
A result does not identify the author or verify the text
When the detector finds a watermark, it does not reveal a user, account, prompt or conversation. It also does not measure the human share of the work, establish ownership or responsibility, decide whether use was lawful, or verify the accuracy of the text. A positive result indicates that the statistical signal associated with supported OpenAI output was detected. Source
A negative result does not establish human authorship. The passage may be short, edited or translated, produced by an unsupported model, created before watermarking began, or generated by another provider. OpenAI also says the detector does not decide a publisher’s disclosure duties or legal responsibility. Source