ChatGPT-maker OpenAI has announced that it will add invisible watermarks to text generated by ChatGPT and Codex for users in the European Union, as the company moves to comply with new requirements under the EU AI Act.
The company also said that the rollout will being over the coming weeks and will apply to eligible ChatGPT and Codex text outputs across all plans in the EU. The move is designed to make AI-generated text identifiable in a machine-readable way while promoting greater transparency around AI-created content.
OpenAI said that over the coming weeks it will add the watermark to eligible text output from ChatGPT and Codex. The company said it is not making text watermarking a global default at launch and described the regional approach as a way to learn from real-world use and feedback. OpenAI has published a technical report on the method and said it plans to release the technology as open source.
The change covers all plans in the EU and nowhere else. OpenAI’s technology, called textGrain, adds a statistical signal to the model’s word choices. Readers cannot see it. A detector looks for that pattern to judge whether a passage carries an OpenAI watermark. OpenAI’s own evaluations show where the technology struggles: OpenAI listed several things a detection result cannot tell anyone:
OpenAI said it is working with cloud partners to offer watermarking for its models through their services in the coming weeks. The company is also making watermarking available to developers. Starting immediately, API customers worldwide can opt in to watermarked text for select models. It stays off by default in the API. Its tools for images and audio remain publicly available. These include a web verification tool at openai.com/verify and a Content Provenance API.
What it means for users
OpenAI said it found no meaningful quality difference in tests of its latest frontier model, Astra, with and without watermarking. For people using ChatGPT or Codex in the EU, the visible experience should not change, since the marker cannot be seen. On eight benchmarks, the scores were close, with the watermarked version slightly ahead on some and slightly behind on others.