To comply with the transparency rules set out in the AI Act, OpenAI is rolling out textGrain, an invisible watermark embedded in the choice of words, exclusively for European users.
OpenAI announced on October 5 that it would integrate an invisible watermark into texts produced by ChatGPT and Codex for users located in the European Union. The measure meets the transparency obligations of the AI Act, which came into force on August 2, which require AI-generated content to be marked in a way that can be detected by a machine. The deployment is expected to take place over the next few weeks, all subscriptions combined, but only in Europe; on the API side, developers around the world can activate it now on certain models, with the option remaining disabled by default.
Marking in words, not in layout
Called textGrain and described in a technical report written with researchers at the University of Pennsylvania and Yale, the method subtly guides the choice of words using a secret key. Each inflection is imperceptible, but their accumulation allows a detector with the key to recognize the text, even if it is copied and pasted. OpenAI claims that the watermark does not identify the user and does not affect performance.
However, the company acknowledges its limitations: replacing 10% of words with synonyms drops the detection rate from about 92% to 66%, and short passages, mathematical answers or translations are more difficult to spot. The detector will therefore initially be reserved for researchers and approved organizations, and OpenAI insists that the absence of a watermark does not prove that a human wrote the text. Two months earlier, Anthropic had announced a comparable measure, applied on a global scale.
Read the full article on the original source
Go to the source →