OpenAI announced on October 5, 2026 its approach to text provenance in response to the EU AI Act, which requires generative AI providers to make generated text identifiable in a machine-readable way. Starting the same day, API customers globally can opt in to text watermarking for select models, though it remains off by default in the API. Over the coming weeks, OpenAI said it will add an invisible watermark to eligible ChatGPT and Codex text output in the European Union. The company also opened applications for access to its text watermark detector, initially limited to approved researchers and expert organizations. The watermarking technology, called textGrain, adds a statistical signal to the model's word choices; OpenAI said it plans to release the technology in open source. In its evaluations, OpenAI reported that at a 1% target false positive rate, its detector identified watermarks in about 80% of 200-token passages and about 95% of 400-token passages for content such as psychology, with substantially lower rates for mathematics. Editing reduced reliability: replacing 10% of words with synonyms cut detection from about 92% to 66%, and replacing 25% dropped it to 17%. OpenAI stated a watermark does not measure human contribution, establish ownership, identify the user, or verify accuracy. Existing image and audio verification tools remain publicly accessible.
- API text watermarking is opt-in and off by default; ChatGPT and Codex output in the EU will get an invisible watermark over the coming weeks
- Detector access is initially limited to approved researchers and expert organizations
- Detection dropped from ~92% to 17% when 25% of words in a 400-token passage were replaced with synonyms
What it means for you
OpenAI is now tagging text from ChatGPT and Codex in the EU with a hidden signal so it can later be identified as AI-generated, because EU law requires it. But OpenAI is unusually blunt that this doesn't work well: short text, math-heavy text, or text you edit even slightly becomes much harder to detect, and a watermark tells you nothing about who wrote it or whether it's true. In practice, you can't rely on these watermarks to prove whether any given document was AI-written.
Try this
If you edit AI drafts before publishing, know that light editing already breaks the watermark — so don't assume your output is traceable, and don't assume anyone else's is either.
Who should care
Businesses operating in the EU, developers using OpenAI's API, and anyone whose job involves detecting or disclosing AI-generated text (compliance, education, publishing).
Skip this if
You don't operate in the EU and you aren't trying to detect or verify whether text was machine-generated.
Sources: OpenAI — read the original