OpenAI will launch the textGrain technology for ChatGPT and Codex users in the EU in the coming weeks, enabling AI-generated texts to be marked with invisible watermarks and their origin to be identified.

The watermark is embedded as a statistical pattern without hidden characters

The marking will be integrated directly into the text structure through a statistical pattern in word and token selection, without using hidden characters or special spaces.

The detector identifies texts created or processed by OpenAI models

A dedicated detector will be able to determine whether a text was generated or processed by OpenAI models.

Paraphrasing and translation reduce accuracy; launch linked to EU AI Act

Paraphrasing, translation, and word substitution can significantly reduce detection accuracy. The launch is connected to EU AI Act requirements for labeling AI-generated content.