OpenAI plans text watermarks for ChatGPT and Codex in the EU, while detection has clear limits
OpenAI says API customers can opt in to textGrain for select models, while eligible EU ChatGPT and Codex outputs will get it over coming weeks. Short, mathematical and edited text is harder to detect.
On October 5, OpenAI announced textGrain: global API customers can opt in for select models, with watermarking off by default. Eligible EU ChatGPT and Codex text outputs will receive invisible watermarks over the coming weeks; this is not a global default. Detector access initially goes to approved researchers and expert organizations. The approach responds to the EU AI Act’s provider marking requirement, using statistical word-choice signals to detect an OpenAI watermark.
In OpenAI’s English ELI5 evaluation, at a target 1% false-positive rate, detection was about 80% for 200-token psychology passages and 95% for 400-token passages, with lower results for mathematics. Replacing 10% of words in 400-token passages lowered detection from about 92% to 66%; replacing 25% lowered it to 17%. These figures do not establish performance across all languages or text. A watermark cannot prove accuracy, ownership or authorship; absence of detection does not prove human creation.
Sources
- Our approach to EU text provenance rules — OpenAI
- Code of Practice on Transparency of AI-generated Content — European Commission