On Oct. 5, 2026, OpenAI announced plans to add its textGrain watermark to eligible ChatGPT and Codex text across all plans in the European Union over the coming weeks. The signal is statistical and invisible to readers; OpenAI’s detector looks for a pattern in a model’s word choices.
OpenAI’s planned EU rollout
OpenAI said the planned consumer rollout covers eligible ChatGPT and Codex text across all plans in the EU. The announcement describes a rollout over the coming weeks; it does not give a more specific schedule.
The scope is eligible text output. OpenAI’s announcement does not say that every kind of Codex-generated code is covered.
How textGrain works
A text watermark is a signal built into generated text so a detector can look for evidence of its origin. OpenAI describes textGrain as an invisible statistical signal introduced through a model’s word choices—not a visible label added to the page.
That makes the detector’s result a limited provenance clue, not a verdict about a whole passage. OpenAI says a detected signal can indicate that an OpenAI system generated or processed part of the text. It does not identify the user, account, prompt, or conversation, or measure the human contribution.
What OpenAI’s tests found
OpenAI reported different detection rates for different text lengths and types. At a target false-positive rate of 1%, it reported about 80% detection for eligible 200-token psychology passages and about 95% for eligible 400-token psychology passages. Results for mathematics were substantially lower.
In a separate 400-token evaluation, replacing words with synonyms reduced detection from about 92% to 66% when 10% of the words were replaced, and to 17% when 25% were replaced.
| Evaluation condition | Text sample | OpenAI-reported detection |
| Target false-positive rate of 1%; 200 tokens | Eligible psychology passages | About 80% |
| Target false-positive rate of 1%; 400 tokens | Eligible psychology passages | About 95% |
| 10% of words replaced with synonyms | 400-token passages | From about 92% to 66% |
| 25% of words replaced with synonyms | 400-token passages | From about 92% to 17% |
These percentages describe OpenAI’s evaluations under specific conditions. They are not a universal detection rate for any text a reader might submit. A result also cannot verify a passage’s accuracy, establish ownership or responsibility, or show how much of it was written by a person.
What a detector result can establish
A detected watermark is a signal that an OpenAI system generated or processed part of a passage; it does not name a writer. Conversely, failing to detect a watermark does not prove that a person wrote the text. OpenAI notes that a passage may be too short or may have been edited or translated, among other reasons.
OpenAI said initial access to its text detector would be for approved researchers and expert organizations that apply.
The separate API option
The API arrangement is distinct from the planned EU rollout for ChatGPT and Codex. OpenAI said API customers worldwide could opt in to text watermarking for select models starting Oct. 5, and that it would remain off by default. OpenAI also said it would not make watermarking a global ChatGPT default at launch.
