Overview
- OpenAI announced Tuesday that it will roll out textGrain to ChatGPT and Codex in the European Union in the coming weeks, with the feature disabled by default outside the EU.
- TextGrain works by nudging the model's word choices to create a subtle statistical pattern that is invisible to readers and requires a detector to spot.
- OpenAI published test results showing detection varies by text type and length, for example a psychology text hit rates rose from 78.5% at 200 tokens to 94.3% at 400 tokens while mathematical texts scored as low as 36.5% at 200 tokens.
- The company will limit detector access at launch to vetted researchers and specialist organisations who must apply, and paying API customers may opt in to enable the watermark globally.
- OpenAI warns the watermark is probabilistic and fragile: simple edits, translations, or replacing 10% or 25% of words can cut detection from about 92% to 66% or 17%, and a detected mark does not identify users, ownership, human contribution, or truthfulness of content.