OpenAI is introducing invisible watermarking for text generated by ChatGPT and Codex within the European Union. This initiative aligns with the transparency requirements set by the EU AI Act. The company plans to roll out this feature to eligible users in the region over the coming weeks.
The technology, known as textGrain, embeds statistical signals into the word selection process of the models. A dedicated detector can then identify these signals to verify if the text originated from an OpenAI model. This feature will not be enabled as a global default at launch.
For API customers, OpenAI is offering an opt-in watermarking feature starting today. This allows developers to decide how to implement watermarks to meet transparency obligations. The company is also working with cloud partners to extend this functionality to models accessed through their platforms.
Additionally, OpenAI is providing researchers and expert organizations with access to detection tools. These entities can apply for access to evaluate text provenance. The tool confirms the presence of a watermark but does not disclose user identities, prompts, or specific conversation content.
