OpenAI Adds Invisible ChatGPT Watermark in EU Under AI Act

TLDR

OpenAI revealed on October 5 that it will include an undetectable watermark to ChatGPT and Codex outputs throughout the European Union. The business states the relocation responses openness requirements under the EU AI Act.

We’re broadening our technique to material provenance to consist of text in action to EU regulative requirements, while acknowledging the considerable restrictions of present text watermarking innovation.

Our tools currently assist validate whether an image or audio file was developed with our …

— OpenAI (@OpenAI) October 5, 2026

The watermark system is called textGrain. It works by pushing the design’s word options to leave a surprise analytical pattern.
A different detector tool then scans text for that pattern. OpenAI states the approach includes no surprise characters, undetectable areas, or odd punctuation.
This indicates a typical workaround, scrubbing covert characters from text, will not get rid of the watermark. The signal is constructed into the phrasing itself.
How Well Does the Watermark Hold Up?
Detection precision depends greatly on text length. OpenAI procedures text in tokens, which are word pieces that AI designs procedure. One token is approximately three-quarters of an English word.
In business tests, the detector captured about 66% of unedited responses around 150 words long. At approximately 300 words, that rate reached about 92%.

Modifying the text deteriorates the watermark quick. For 300-word passages, switching one word in 10 for a synonym cut detection from 92% to 66%.
Changing one word in 4 pressed detection to simply 17%. OpenAI’s own information reveals the system has a hard time as soon as text gets modified, reduced, or reworded.
The business likewise stated the watermark had little impact on benchmark ratings for its most current design, GPT-6 Astra.
Why Isn’t the Detector Public?
OpenAI is not launching the detector to the public immediately. Rather, gain access to will go to authorized scientists and professional companies.
The business states a favorable match from the detector does not expose who the user was, what they typed, or what the discussion consisted of.
An unfavorable outcome does not show much either. Short, modified, or equated text can slip past the detector without being flagged.
Output from other AI tools will likewise not set off the system, because it just look for OpenAI’s own pattern. OpenAI specified clearly that the lack of a found watermark does not show a human composed the text.
Images and audio are managed in a different way than text. Anybody can currently submit those files to OpenAI’s public confirmation tool to look for its SynthID watermark.
Outside the European Union, the ChatGPT watermark will remain off by default for routine users. API clients all over the world, nevertheless, can choose in for choose designs through their task or company settings.
The rollout follows the EU started imposing AI Act openness guidelines back in August. OpenAI framed the watermark as part of its continuous action to that enforcement.
The business likewise indicated a larger information point in its statement. Bad guy usage of AI tools climbed up 40% year-on-year, according to research study company TRM Labs.
OpenAI did not state when or if the detector tool would open to the public. In the meantime, the EU rollout and restricted scientist gain access to stay the existing state of the job.