OpenAI to Add Invisible Watermarks to ChatGPT and Codex Text in Europe

API customers globally will also be able to opt into text watermarking for select models, OpenAI said. The company is accepting applications for its text watermarking detector, with access initially limited to approved researchers and expert organizations to help test the technology.

OpenAI said its watermarking technology, called textGrain, adds a statistical signal to the model's word choices. A detector then looks for that signal to verify whether text was generated using ChatGPT or Codex.

OpenAI said textGrain matched and even exceeded the performance of all the approaches it tested, but it does not always work as intended. The company said detectors can make two kinds of errors: reporting a watermark where none is present, known as a false positive, or missing a watermark that is present, known as a false negative. OpenAI also said shorter text is harder to detect and editing can weaken the watermark.

The detection system cannot determine how much a human contributed to the text, who owns it, who created it, or whether the information it contains is accurate, according to OpenAI.

For now, OpenAI will introduce text watermarking to all users across all plans in Europe. The system is not going global just yet. OpenAI said its regional approach will give it room to learn from real-world feedback.

The change comes after years in which AI tools have made it easy to generate everything from images and audio to human-like text, Android Authority reported. While AI-generated images, videos and audio can often be identified using dedicated detection tools, AI-written text has largely escaped the same level of scrutiny. ChatGPT has become one of the most widely used tools for generating text that can closely mimic human writing, and OpenAI says it plans to change that with watermarks for text generated by ChatGPT and Codex.

Editor's Summary

OpenAI will begin adding invisible watermarks to eligible ChatGPT and Codex text in Europe under the EU AI Act, while letting API customers globally opt into watermarking for select models and limiting detector access to approved researchers and expert organizations. The textGrain system embeds a statistical signal in word choices, but OpenAI says short or edited text can be harder to detect, and the detector cannot determine human contribution, ownership, authorship or accuracy. The Europe-first rollout is intended to gather real-world feedback before any broader global deployment.