OpenAI starts invisibly watermarking ChatGPT text in the EU

OpenAI is rolling out textGrain, an invisible watermark for eligible ChatGPT and Codex text in the EU, over the coming weeks and on all plans. Nothing visible changes. The model's word choices are nudged to leave a statistical pattern that a detector can pick up later, and it survives copy-paste. The driver is Article 50 of the EU AI Act, whose transparency rules took effect on 2 August.

Outside the EU, API customers can opt in for select models from 5 October. It's off by default, and OpenAI isn't making it a global default yet.

OpenAI says it's no silver bullet. It won't prove authorship or verify accuracy, and heavy editing breaks it. Reported figures put detection at about 80% for 200 tokens and 95% for 400 under ideal conditions. Swapping roughly 25% of words for synonyms drops that to around 17%. Detector access stays limited to vetted researchers for now, so teachers and employers can't use it yet.

Why it matters: this is the first big compliance-driven watermark on mainstream chat output, and it's a weak signal that casual rewording can defeat.