OpenAI to introduce watermarks for ChatGPT and Codex in the EU — The Verge

OpenAI will begin rolling out invisible machine-readable watermarks in text responses from ChatGPT and Codex for users in the European Union over the coming weeks. The feature, called textGrain, will be available to eligible users across all plans in the EU.

The company will not make text labeling a global standard at launch. OpenAI explained its regional approach by the need to gain experience with the technology's practical use and receive user feedback.

Labeling for the API

As The Verge reports, OpenAI API customers worldwide can already voluntarily choose watermarks for text responses from individual models. The company says this will allow customers to independently determine how to use labeling based on their own transparency obligations and user experience.

More current news is available on the UA.News Telegram channel Telegram.

OpenAI is also working with cloud partners to make watermarks available over the coming weeks for model responses accessed through their services.

Technology limitations

According to OpenAI, textGrain matched or outperformed other approaches in testing, including Google DeepMind's SynthID for text. The company published AI benchmark results indicating similar model performance with and without labeling.

At the same time, OpenAI stressed that textGrain does not guarantee reliable detection. The watermark does not confirm the accuracy of text, identify its owner, show the extent of human input, or prove human authorship. Approved researchers and specialized organizations can apply for access to the detector.