OpenAI announced Monday that it will add invisible watermarks to text generated by ChatGPT and Codex for users in the European Union, effective over the coming weeks. The move complies with the EU AI Act's transparency rules, which took effect August 2 and require AI companies to mark AI-generated content in ways that other systems can identify.
The watermark operates by subtly shaping the model's word choices to create a pattern that humans cannot perceive but detection software can identify. OpenAI said the watermark does not identify users and causes no meaningful degradation in model performance. The company published a technical report for the method, called textGrain, co-written with researchers from the University of Pennsylvania and Yale. The approach uses a secret key to sort next-word predictions, and combining hundreds of these nudges allows a detector to flag AI-generated content using only the text and the key.
OpenAI noted significant limitations. Tests showed that replacing 10% of words with synonyms dropped detection accuracy from approximately 92% to 66%. Short passages, math answers, and translated text proved harder to detect. The company said it will restrict detector access initially to approved researchers and expert organizations to evaluate reliability and responsible uses.
The company cautioned that a missing watermark does not prove human authorship. Text may be too short, too heavily edited, or generated by a different AI system. OpenAI stated that watermarks can indicate its system generated or processed part of a passage but cannot measure how much human judgment, editing, or creativity contributed.
API developers worldwide can enable watermarking for select models starting today, though it remains off by default globally. The EU-specific rollout reflects regional legal requirements. The announcement follows Anthropic's decision two months ago to watermark Claude-generated text globally, a move that drew criticism from some users who argued they provided instructions and context while Claude functioned as a tool. OpenAI had previously built a text watermark but withheld release, partly over concerns users would switch to unwatermarked competitors, according to the Wall Street Journal reporting in 2024.
Anthropic, Google, Meta, Microsoft, and OpenAI have committed to following the EU's code of practice on AI-generated content.