OpenAI is hiding an invisible watermark in ChatGPT text for EU users, like a fingerprint only its scanner can see
By Brett Gavaghan, Founder of Adaptd ·
OpenAI will add an invisible watermark to ChatGPT and Codex text in the European Union, and API customers anywhere can now opt in. Here is what it can and can't tell you.
The short version
- OpenAI's textGrain adds an invisible statistical watermark to the words ChatGPT picks, and a separate detector can test for it.
- Eligible ChatGPT and Codex users on all plans in the EU get it over the coming weeks; API customers worldwide can opt in now.
- Australian businesses don't need to change anything yet, but shouldn't treat AI detectors as proof of who wrote something.
The EU AI Act wants AI-written text to be identifiable by machines, and on 5 October OpenAI explained how it will do that. The answer is a watermark you can't see, tucked into which words the model chooses.
What is OpenAI's text watermark?
OpenAI's text watermark, called textGrain, is an invisible statistical signal added to the word choices ChatGPT makes, which OpenAI's detector can look for later.
Nothing changes on screen. The text reads the same, and OpenAI says it saw no meaningful difference in quality on its benchmarks with the watermark switched on. The detector itself is only open to approved researchers and expert organisations, who have to apply.
Who gets it, and when?
ChatGPT and Codex users on every plan in the European Union will get the watermark over the coming weeks, and only there for now.
OpenAI isn't making it a global default. Everywhere else, including Australia, API customers can opt in for select models, but it stays off unless they do. OpenAI is upfront about the limits, too. Short passages are harder to detect, and in its tests, swapping 10% of the words for synonyms cut detection from about 92% to 66%.
Should your business do anything about it?
Most Australian businesses don't need to change anything yet, unless they serve EU customers through ChatGPT or build products on OpenAI's API.
My take: the most useful part is OpenAI's own fine print. A watermark doesn't measure how much a person contributed, doesn't say who is responsible for the text, and a missing watermark doesn't prove a human wrote it. So if anyone in your business is running staff or supplier work through an AI detector, stop. Judge the work on whether it's right, not on whose fingerprints might be on the glass.
Sources
Wondering what this means for your business?
That's exactly what the discovery call is for.