textGrain
Announced October 5, 2026. OpenAI reports that at a 1 percent false positive rate its detector found the mark in about 95 percent of 400-token passages and about 80 percent of 200-token passages, with lower rates on constrained text such as mathematics. Detector access is limited to approved researchers and expert organisations. OpenAI says it plans to open-source the technique.
What it is
textGrain is OpenAI's own statistical watermark, announced with an unusually detailed set of failure numbers. At a 1 percent false positive target, OpenAI's detector found the mark in about 95 percent of 400-token passages and 80 percent of 200-token passages on open-ended topics, with much lower rates on constrained text such as mathematics. Replacing a tenth of the words with synonyms dropped detection on 400-token passages from about 92 to 66 percent; a quarter dropped it to 17 percent. The rollout is phased: opt-in for API customers everywhere, on by default only for ChatGPT and Codex users in the EU, and the detector only for approved researchers and expert organisations. OpenAI says the technique will be open-sourced.
Known limits
In OpenAI's own evaluation, replacing 10 percent of words with synonyms cut detection on 400-token passages from about 92 to 66 percent, and replacing 25 percent cut it to 17 percent.
When to use this entry
Use this entry for the most precise public numbers on how editing weakens a text watermark, and for the EU-only scope of ChatGPT marking.
Questions
Is ChatGPT output watermarked in the United States?
Not by default. As announced on October 5, 2026, the mark applies to eligible ChatGPT and Codex output in the EU; API customers anywhere can opt in.
Can I detect textGrain myself?
No public detector exists. Access is granted case by case to researchers and expert organisations.
Does textGrain change answer quality?
OpenAI published benchmark scores with and without the mark and reports differences within normal variation.