OpenAI has initiated the deployment of an invisible, machine-readable watermarking system for text generated by ChatGPT and Codex. This feature is currently being introduced exclusively for users located within the European Union across all subscription plans. The company plans to roll out this update over the coming weeks to gather feedback and observe real-world performance.
The new technology, branded as textGrain, is designed to align with the transparency requirements mandated by the EU AI Act. OpenAI reports that the performance of textGrain matches or exceeds existing solutions, such as Google DeepMind’s SynthID. Internal benchmarks indicate that the presence of a watermark does not significantly impact the quality or performance of the generated text content.
Beyond the EU rollout, OpenAI is offering an opt-in watermarking feature for global API customers starting today. This allows developers to integrate watermarking based on their specific transparency needs. The company is also collaborating with cloud partners to extend this functionality to model outputs accessed through third-party services in the near future.
Access to detection tools is currently restricted to approved researchers and expert organizations. OpenAI is granting access on a case-by-case basis to facilitate the evaluation of text provenance. The detection tool identifies the presence of an OpenAI watermark but does not reveal user prompts, conversation history, or personal identity information.
