OpenAI announced on October 5 that it will begin embedding an invisible statistical watermark called textGrain in AI-generated text outputs from ChatGPT and Codex in the European Union. This move aligns with the EU AI Act requirements and will roll out over the coming weeks. API customers worldwide can opt in to watermarking for select models, although the feature remains off by default, according to medianama.com.
The watermark alters the statistical pattern of word selection without adding visible marks or hidden characters. OpenAI will restrict access to the watermark detector to approved researchers and expert organizations. The detector identifies whether a passage contains the OpenAI watermark but does not reveal user identities or prompt details. This approach aims to balance transparency with user privacy.
OpenAI’s tests showed that the watermark detection performs best on longer passages, identifying watermarks in about 80% of 200-token psychology texts and 95% of 400-token passages. Detection accuracy drops for shorter, heavily edited, or highly constrained texts, such as mathematical content where word choice is limited. Editing text by replacing 10% of words in a 400-token passage can weaken the watermark signal, highlighting the technology’s current limitations.
The watermarking rollout in the EU is part of OpenAI’s compliance with the EU AI Act, which seeks to increase transparency in AI-generated content. The company’s detector will initially be available only to select researchers and organizations, marking a cautious step in AI content identification under evolving regulatory frameworks.