OpenAI to Embed Watermarks in AI-Generated Text

OpenAI will embed watermarks in AI-generated text. OpenAI announced on the 5th (local time) that it will introduce 'TextGrain,' a technology for identifying AI-generated text, in a move responding to the EU AI Act's transparency obligations for generative AI.

TextGrain works by embedding subtle statistical signals during the AI model's word selection process as it generates text. OpenAI explained that the technology causes no meaningful difference in the quality, performance, or readability of AI responses.

There are technical limitations, however. Detection rates drop for short sentences or content with little flexibility in word choice, such as mathematical formulas. OpenAI also noted that post-editing, such as replacing words with synonyms, significantly reduces detection rates. According to OpenAI, replacing one-tenth of the words with synonyms in a 400-token text lowers the detection rate from 92% to 66%, while replacing a quarter drops it to 17%.

The technology will be applied to ChatGPT and Codex outputs in the EU within the coming weeks. It will be disabled by default for API outputs, with individual organizations free to enable it themselves. Watermark detection tools will first be provided to approved researchers and professional institutions upon application for access.