OpenAI's textGrain watermarking system for ChatGPT text can detect its own marks in 80-95% of outputs, but detection collapses when one word in four is replaced with synonyms, dropping to 17%. The company is deploying the watermark by default in the EU to comply with Article 50 of the AI Act, which requires machine-detectable marking of synthetic text, while leaving it optional elsewhere.
OpenAI will add invisible watermarks to ChatGPT and Codex text in the EU starting in coming weeks to comply with the EU AI Act's transparency requirements. The watermark, called textGrain, subtly shapes word choices to create a detectable pattern without identifying users or significantly affecting model performance. The company notes limitations including vulnerability to editing and shorter passages, and will initially restrict detector access to approved researchers.