OpenAI to add invisible watermarks to ChatGPT and Codex text output in the EU

On October 5, 2026, OpenAI announced that it will add a machine-readable invisible watermark to eligible ChatGPT and Codex text output in the EU over the coming weeks, in order to comply with the content transparency requirements of the EU AI Act. The watermarking technique is called textGrain. According to OpenAI, it embeds an invisible statistical signal in the model's word-selection process, and a detector can read that signal to determine whether a piece of text carries an OpenAI watermark. The company said the feature will initially roll out only in the EU, across all plans, to eligible ChatGPT and Codex users.

OpenAI also said watermarking will not be on by default globally at launch. For API customers, the feature is opt-in: starting October 5, API customers worldwide can choose to add watermarked text output for some models, so they can decide how watermarking fits their own transparency obligations and the experience they offer their users. OpenAI added that it is working with cloud partners to make watermarking available for OpenAI model output accessed through their services in the coming weeks.

On the detection side, OpenAI will give researchers and expert organizations access to a detector. Approved researchers and expert organizations could begin applying on October 5, and the report said access under the code of conduct will initially be handled on a case-by-case basis. The report did not disclose further technical details of textGrain, its detection accuracy, or whether watermarking will later become a default outside the EU.