OpenAI said on October 5 that it will add an invisible watermark, called textGrain, to eligible ChatGPT and Codex text for users in the European Union. API customers anywhere can turn it on for select models. At first, only approved researchers can use the detector.
New to this? Read it in simple words
- OpenAI will hide an invisible mark in ChatGPT and Codex text for EU users.
- The EU AI Act asks for AI-made text that machines can recognise.
- In OpenAI’s tests, swapping 25% of the words cut detection to 17%.
- At first, only approved researchers can use the tool that finds the mark.
- Watermark
- A hidden signal in content that shows which system made it.
- Token
- A word or a part of a word, the unit an AI model reads and writes.
- API
- A way for other companies’ software to use an AI model directly.
What OpenAI will do
Over the coming weeks, OpenAI will add an invisible watermark to eligible ChatGPT and Codex text for users in the European Union, on all plans. It is not making the watermark a global default.
From October 5, API customers anywhere can turn the watermark on for select models. It stays off by default, OpenAI says.
At first, only approved researchers and expert organisations can apply to use the detector. OpenAI’s public tools for checking images and audio stay open.
Detection of the textGrain watermark in 400-token passages, from OpenAI’s own tests. No outside test yet.
How it works, and how well
The method, called textGrain, adds a hidden statistical signal to the model’s choice of words. A detector then looks for that signal in a passage.
In OpenAI’s own tests, the detector found the watermark in about 95% of 400-token passages, and in about 80% of 200-token passages. A token is a word or a part of a word.
Editing weakens the signal. When 10% of the words were swapped for synonyms, detection fell from about 92% to 66%. When 25% were swapped, it fell to 17%.
Maths answers were much harder to detect, because there is less choice of words. OpenAI says the watermark did not change its Astra model’s test scores in a meaningful way.
What a watermark can’t show
OpenAI lists the limits itself. A detected watermark shows that an OpenAI system made or processed part of a text, but not how much a human added.
It does not identify the user, prove who owns the text, or show whether the text is true. A missing watermark does not prove that a human wrote it, because short, edited or translated text can escape detection.
OpenAI also says it plans to release the method as open source, but it gave no date.
Sources
Every fact in this story comes from the sources below. Open them to check our work.
- 1
- 2
- 3Research · October 5, 2026OpenAI is adding invisible watermarks to ChatGPT and Codex text in the EU BleepingComputer
- 4Research · October 6, 2026OpenAI plans to watermark ChatGPT and Codex text output in EU amid AI transparency push Business Today
All detection figures come from OpenAI’s own tests, and no outside group has checked them yet. The detector is not public, so nobody else can test it today.
A compliance plan still rolling out; OpenAI’s own untested figures show light editing hides most of the mark, so the benefit is unclear. How we rate the mood