OpenAI has announced its compliance strategy for the European Union AI Act’s text provenance mandate, introducing an invisible text watermarking system known as textGrain. The tool injects statistical signals into model word selections, enabling specialized detectors to determine if a passage was generated by OpenAI models.

Beginning immediately, API customers can opt into text watermarking, though the feature remains off by default globally. Over the coming weeks, OpenAI will automatically apply invisible watermarks to eligible ChatGPT and Codex text outputs generated within the European Union. Access to the corresponding detection tool will initially be limited to select researchers and organizations.

OpenAI highlighted significant technical limitations with current text watermarking capabilities. In testing, short passages, technical text like mathematics, and edited content significantly degraded detection accuracy. Despite these constraints, OpenAI plans to eventually open-source textGrain to assist broader industry development.

Why it matters

  • EU AI Act enforcement is forcing frontier AI labs to deploy watermarking, altering enterprise API default settings.

  • Technical limits in text watermarking leave enterprise applications vulnerable to evasion through simple editing.

  • Open-sourcing provenance technology creates opportunities for developers to build standards-compliant verification tools.

Source: openai.com