OpenAI introduces watermarking for AI-generated text

In response to the EU AI Act requirements mandating that AI-generated text be identifiable in a machine-readable way, OpenAI introduced textGrain, a text watermarking system that adds an invisible statistical signal to the model’s word choices.

Rather than inserting hidden characters, textGrain influences which words the model chooses at each step of generation. The system slightly alters the statistical pattern of word selection, creating a signal that is imperceptible to humans but detectable by a detector. Because the signal is embedded in the words themselves, it does not depend on metadata and survives ordinary copying. However, editing and translation can significantly weaken the watermark, so detection reliability depends on the length and content of the text.

The technology has several limitations, which is why OpenAI is not making the detector publicly available: • shorter or more constrained text is harder to detect; • replacing 10% of words with synonyms reduces the detection rate from about 92% to 66%, while replacing 25% of words reduces it to 17%; • the absence of a detected watermark does not prove human authorship. The text may have been edited, translated, generated by another model, or created before watermarking was introduced.

API customers globally can opt in to text watermarking for select models. Text watermarking is off by default in the API. Over the coming weeks, OpenAI will add an invisible watermark to eligible ChatGPT and Codex text output in the European Union. Access to the text watermark detector will initially be limited to approved researchers and expert organizations.