OpenAI to watermark ChatGPT and Codex text with textGrain in EU

The rollout is tied directly to the EU AI Act, which requires generative AI providers to make AI-generated text machine-identifiable. OpenAI is answering that with textGrain, a technique that avoids stamping characters or metadata onto the text and instead bakes a statistical signal into how the model picks its words.

OpenAI's own numbers show where the approach holds and where it frays. Detection is strongest on prose with wide word choice, but weakens when the text is short, when the subject offers few alternatives, or when a person edits it — swapping 10% of words in a 400-token English passage cut detection from about 92% to 66%, and swapping 25% dropped it to 17%. The company also stresses that a detected watermark cannot establish human involvement, ownership, responsibility, or user identity, and that a missing watermark does not prove human authorship.

That is why the detector will go first to approved researchers and institutions rather than the general public. The practical stakes look likely to hinge on how much access those outsiders get, and whether detection holds up outside OpenAI's tests — especially on text that has passed through editing tools.