On October 5, 2026, OpenAI announced plans to add its textGrain watermark to eligible ChatGPT and Codex text output in the EU. The company said the rollout would take place over the following weeks for eligible users on all plans. A separate option lets API customers worldwide opt in for selected models, with watermarking off by default. OpenAI’s announcement ties the phased approach to EU transparency obligations.
OpenAI’s EU product plan
The planned watermark applies to eligible text output from ChatGPT and Codex for eligible users in the EU, across all plans. OpenAI said it would not make watermarking a global default for its consumer products at launch.
The API option is separate
API customers worldwide can opt in to watermarking for selected models. The feature is off by default, so this is a separate choice from the planned EU rollout for ChatGPT and Codex.
How textGrain works—and what OpenAI’s tests found
OpenAI describes textGrain as an invisible statistical signal encoded in a model’s word choices, rather than a visible label or special character. A detector assesses the text for that pattern.
In an OpenAI evaluation of 400-token passages, replacing 10% of words with synonyms reduced detection from about 92% to about 66%; replacing 25% reduced it to 17%. These results apply to that test condition, not every passage.
What a watermark can establish
OpenAI says a detected watermark does not identify a user, account, prompt, or conversation, and cannot measure how much human judgment or editing contributed to a passage. A detector can also return false positives or false negatives.
No detected watermark does not prove that a person wrote the text. Short, edited, translated, unsupported, pre-watermark, or other providers’ model outputs may lack a detectable OpenAI signal.
