OpenAI is introducing invisible watermarking for text generated by ChatGPT and Codex within the European Union. This measure aligns with the EU AI Act, which mandates that AI developers label machine-generated content so that other systems can detect it. The regulation officially entered into force on August 2.
The watermark functions by subtly adjusting the model's word selection to create a detectable pattern. This marker persists even when text is copied or moved. OpenAI states that the process does not identify individual users and does not negatively impact model performance. The company also released a technical report on its textGrain method.
The feature will roll out to all ChatGPT and Codex users in the EU over the coming weeks. For API developers globally, the option is available starting today, though it remains disabled by default. OpenAI currently has no plans to make text watermarking a global default for its services.
Internal testing shows that the watermark can be bypassed through editing. Replacing 10% of words with synonyms reduced detection accuracy from 92% to 66%. Additionally, the company noted that short passages, mathematical solutions, and translated text are more difficult to identify, limiting initial detector access to approved researchers.
