OpenAI will begin rolling out technology to identify AI-generated text in the European Union as new transparency requirements under the EU AI Act take effect.
The company said in a blog post that API developers worldwide can opt into an invisible watermark for eligible models starting Monday, while OpenAI plans to begin adding the watermark to text generated by ChatGPT and Codex for users in the EU over the coming weeks.
The watermark is designed to make AI-generated text more traceable, although OpenAI acknowledged that the technology can become less effective when text is edited or rewritten.
OpenAI's watermarking system, called textGrain, uses statistical patterns in word selection to embed a signal that can later be detected. The company said developers using its API will have the option to enable the technology, but it will remain off by default.
OpenAI is also accepting applications for access to a detector that can determine whether text contains an OpenAI watermark. Initially, access will be limited to vetted researchers and specialist organizations.
The detector is designed to identify the presence of an OpenAI watermark without linking the text to a specific user, account or prompt.
OpenAI said its testing found the watermark was more reliable on longer passages than shorter ones. The technology also performed less consistently in areas such as mathematics, where there is generally less flexibility in word choice.
The company said rewriting text can further weaken the signal. In internal testing, replacing some words with synonyms significantly reduced the detector's ability to identify the watermark.
Those limitations are part of the reason OpenAI is initially restricting access to the detection system rather than making it broadly available, the company said.
The rollout comes as the European Union moves ahead with transparency requirements for AI-generated and manipulated content under the EU AI Act, which requires providers of certain generative AI systems to make machine-readable markings available for synthetic content.
OpenAI plans to publish additional technical details about its approach in an updated report and eventually open-source the watermarking technology.
The company is also keeping its existing verification tools for AI-generated images and audio broadly available, including its web-based verification tool and content provenance API.
The move gives OpenAI a way to distinguish AI-generated text at the system level while highlighting one of the central challenges facing AI-content detection: signals can become harder to identify once generated material is modified.
