Overview
- OpenAI announced on Tuesday that it will enable a new invisible watermark called Textgrain by default in ChatGPT and the Codex programming model inside the European Union in the coming weeks.
- Paying API customers can opt in to add Textgrain worldwide, though the company will not enable the mark by default outside the EU.
- OpenAI released example detection rates showing stronger performance on longer, freer-form text and much weaker results on short or technical text, for example psychology passages hit about 94% at 400 tokens while mathematical text reached about 61% at the same length.
- The company warned the watermark cannot identify who generated a text, measure how much a human edited it, assign rights or prove truth, and common edits such as synonym replacement, translation or mixing with other text can sharply reduce detection rates.
- OpenAI will give detector access first to selected researchers and organizations on application, plans to collect feedback before widening access, and intends to open-source Textgrain eventually to allow broader scrutiny.