OpenAI will begin embedding an invisible, machine-readable watermark into ChatGPT and Codex text for European Union users over the coming weeks, in a bid to comply with the EU AI Act's transparency rules.
The technique, dubbed 'textGrain' and co-developed with researchers at the University of Pennsylvania and Yale, subtly shapes word choices via a secret key so the signal survives copy-paste without identifying the user. API developers worldwide can opt in for select models now, though it is off by default — it is not a global default.
OpenAI was candid about the limits: swapping synonyms in just 10 per cent of words drops detection from about 92 per cent to 66 per cent, and short passages, math and translations are harder to detect. Detector access will initially be limited to approved researchers.
The move makes OpenAI one of the first major AI labs to operationalize the EU AI Act's transparency obligations, which took effect August 2, and will test how well watermarking holds up against real-world evasion.
