OpenAI to add invisible text watermarks to ChatGPT and Codex in the EU

OpenAI says it will begin adding invisible watermarks to eligible ChatGPT and Codex text output in the European Union over the coming weeks. The company announced the change on October 5.

OpenAI’s textGrain system subtly changes word choices to add a statistical signal. A detector checks for that signal and reports whether it finds an OpenAI watermark. OpenAI says the detector will not identify users or reveal their prompts or conversations.

At launch, access to the detector will be limited to researchers and expert organizations. OpenAI says approved groups can apply for access, and that it is withholding public access because the detector could miss watermarks or produce false positives. The company also says shorter or more constrained text is harder to detect, and editing can weaken the watermark.

The change applies to eligible users in the EU; the source says users in the US and UK will not be affected. TechRadar’s Graham Barlow describes the move as a response to the EU AI Act’s requirement that AI-generated text be identifiable in a machine-readable way. Barlow questions whether a watermark that can be weakened through editing and checked only by a restricted detector will make AI-generated text meaningfully more transparent.