OpenAI Launches textGrain Watermarking to Meet EU AI Act
OpenAI is introducing invisible textGrain watermarks for ChatGPT and Codex users in the European Union to comply with AI content identification mandates.
OpenAI is rolling out an invisible, machine-readable watermarking system called textGrain for text generated by ChatGPT and Codex. The initiative is a direct response to the European Union AI Act, which mandates that AI-produced content be detectable. The rollout will occur over the coming weeks for eligible users across all plans within the EU, where the feature will be enabled by default. Outside the EU, the feature remains disabled by default, though API users worldwide could opt in starting October 5.
Developed in collaboration with researchers from the Wharton School, textGrain works by subtly influencing word choices to create a statistical pattern detectable via a secret key. The company cautioned that the technology has significant limitations; detection rates drop sharply when text is translated, consists of short passages, or is edited. Technical tests showed that replacing 25% of words with synonyms reduced detection rates from 92% to 17%.
Due to the risk of false positives and missed detections, OpenAI is limiting access to the detection tool to approved researchers and expert organizations. The company stated that the watermark indicates a system processed the text but does not prove human authorship, establish ownership, or verify accuracy. This implementation follows a similar worldwide rollout by Anthropic for its Claude models in August.