OpenAI is introducing invisible watermarking for ChatGPT and Codex text in the European Union to meet transparency requirements under the EU AI Act. The rollout will occur over the coming weeks across all plans in the region. Globally, API customers can now opt in to use this feature for select models.
The technology, named textGrain, modifies the statistical patterns of word choices during text generation rather than adding metadata or visible labels. A specialized detector analyzes passages to estimate the presence of the watermark. OpenAI plans to release more technical details and eventually open-source the technology.
The effectiveness of textGrain varies significantly based on text length and editing. For 400-token passages, the detector identified watermarks with 92% accuracy under test conditions. However, replacing 10% of words with synonyms reduced detection to 66%, and replacing 25% dropped it to 17%.
Access to the detection tool is currently restricted to approved researchers and expert organizations to mitigate risks of false positives and negatives. OpenAI notes that the watermark cannot identify specific users, prompts, or accounts, nor can it confirm the accuracy or copyright status of the generated content.
Internal testing on the Astra model showed no significant performance declines due to textGrain. While the EU rollout is mandatory, OpenAI maintains that the watermark is not a definitive proof of AI origin, as edited, translated, or shortened text may bypass detection, and the absence of a signal does not confirm human authorship.
