OpenAI adds default ChatGPT watermarks in the EU

OpenAI plans to roll out automatic watermarking for text generated by ChatGPT and Codex in the European Union in the coming weeks. Driven by the EU AI Act that took effect in August, this feature will be enabled by default for European users but will remain opt-in for other regions. For developers using OpenAI's API, the watermarking feature will remain turned off by default but available as an option.

The company is deploying a proprietary methodology called textGrain, which embeds imperceptible patterns into the model's word choices. While these adjustments do not degrade the overall quality of the generated text, they allow a specialized detector to identify AI-generated content. OpenAI will initially restrict access to this detector to a select group of researchers and organizations, though it plans to establish an approval process for other entities over time.

However, the technology faces significant technical limitations. OpenAI's internal testing of textGrain shows a peak detection success rate of 92 percent under ideal conditions. That reliability drops sharply with minor edits: altering just 10 percent of the generated text reduces the detection rate by nearly 30 percent, while modifying 20 percent of the text slashes the success rate by almost 75 percent. The system also struggles to identify shorter passages or translated text.

For AI practitioners and developers, this regional rollout highlights the fragmented regulatory landscape. While competitor Anthropic enabled global watermarking for its models in August, OpenAI is taking a more localized approach. Developers must navigate these differing defaults, especially when deploying applications globally. The fragility of textGrain also suggests that while watermarking helps meet compliance mandates, it remains an imperfect defense against deliberate evasion.