OpenAI announced it will begin watermarking text produced by ChatGPT and Codex within the European Union to comply with the EU AI Act, effective since August 2, 2026. This law mandates that AI-generated content must be marked for easy identification by other systems. The watermark, called textGrain, is a subtle pattern embedded in word choices rather than a visible symbol, making it imperceptible to users but detectable with the right key. It will roll out in the coming weeks to all eligible EU users and will be available as an optional feature for developers worldwide through OpenAI's API.

The watermark does not degrade the models' performance or disclose user information but travels with the text when copied. However, OpenAI acknowledges some limitations, as editing the text—such as replacing 10% of words with synonyms—can reduce detection rates significantly, from 92% to 66%. It is also harder to detect the watermark in short passages, mathematical answers, or translated content. Consequently, OpenAI is limiting detector access to selected researchers and expert organizations to ensure responsible evaluation and use.

OpenAI cautions that the absence of a watermark does not conclusively prove human authorship, as heavily edited or short content and AI-generated text from other companies could lack the watermark. The company clarifies that the watermark only signals that an OpenAI system generated or processed part of a passage but does not indicate the level of human involvement or creativity. This move follows Anthropic's similar watermarking initiative for its Claude AI, though Anthropic applies it globally, while OpenAI is currently limiting it to the EU due to user concerns and competitive factors.

The watermarking rollout forms part of broader industry steps to align with the EU’s AI transparency code of practice, supported by major AI companies including Google, Meta, Microsoft, and Anthropic. OpenAI’s decision reflects balancing regulatory compliance and user acceptance, as watermarks could affect how customers interact with AI content but help address concerns about AI-generated disinformation and content attribution. The technical report co-authored by teams from OpenAI, the University of Pennsylvania, and Yale details the watermarking methodology designed to withstand manipulation while respecting user privacy.