On October 5, 2026, OpenAI published how it plans to make generated text machine-identifiable in response to the European Union AI Act. The system is called textGrain. Readers do not see it: it is a statistical signal embedded in the model's word choices.
The rollout is phased. From the announcement, API customers worldwide can turn the watermark on for selected models. It stays off by default. Over the coming weeks, the company expects to add the signal to eligible ChatGPT and Codex output in the European Union, on all plans. OpenAI says it is not making the watermark a global default at launch.
What the detector can and cannot do
The detector looks for the textGrain signal to estimate whether a passage contains an OpenAI watermark. Initial access is limited to approved researchers and expert organizations, who can apply from the announcement. The tool reports whether it detected the mark. It does not identify a user, account, prompt, or conversation.
In the company's own tests, at a target false-positive rate of 1%, the detector found the mark in about 80% of 200-token passages and about 95% of 400-token passages for content such as psychology. In mathematics, where wording is less flexible, the rate drops clearly. Editing also weakens the signal: replacing 10% of words with synonyms cut detection from about 92% to 66% on 400-token passages; replacing 25% dropped it to 17%.
OpenAI says that, on the benchmarks used to evaluate Astra, it did not see a meaningful performance difference with and without the watermark. It also says it plans to open-source the technology so others can build on it. The technical report is due to gain more detail in the coming weeks.
Why it matters
OpenAI images and audio already had public verification tools, including Content Credentials and SynthID. Text is a different problem: it can be rewritten, translated, or shortened. The company lists what the mark does not prove: human contribution, authorship, legal responsibility, who generated the text, or whether the content is true. A missed detection also does not prove that a text is human.
The official post is Our approach to EU text provenance rules. The textGrain report is in the technical PDF.
Sources
Transparency: This content was created, edited, or reviewed with the assistance of artificial intelligence. The information was cross-checked with public posts on X and sources available on the internet. Check the original sources for the full context.
Por GeekikiBot