OpenAI says EU ChatGPT and Codex text will receive invisible watermarks

OpenAI said on Monday 5 October that eligible ChatGPT and Codex users in the European Union will begin receiving text with invisible watermarks over the coming weeks. The company said the move is intended to meet transparency requirements under the European Union AI Act.

OpenAI said its textGrain system places a statistical signal in word choices made by its AI models. A separate detector can analyse text for that signal. API customers worldwide can choose to turn on watermarking for select models, OpenAI said, though the feature will be off by default.

OpenAI said a detected watermark is not proof of how content was produced. According to the company, the system cannot show the extent of human editing or creativity, who generated or is responsible for material, who owns it, or whether its information is accurate. The absence of a watermark also does not show that text was written without AI help.

Short passages may be harder to detect, and editing can weaken the watermark, OpenAI said. The company said it is not making its detector available to the public at first because false positives and missed detections remain possible. Approved researchers and expert organisations can apply for access, with applications reviewed case by case.

OpenAI said the detector will report whether it finds an OpenAI watermark without identifying users or revealing prompts or conversations. The company also said it plans to make textGrain open source in the future.