NEW YORK--OpenAI has announced plans to begin placing invisible watermarks in eligible text generated by ChatGPT and Codex in the European Union as the company moves to comply with transparency requirements under the EU AI Act.
The company announced the move on Monday, October 5, saying its new text watermarking technology, calledtextGrain, places an invisible statistical signal within the word choices made by its AI models. A separate detector can then analyze text for signs that the watermark is present.
OpenAI said the EU AI Act requires providers of generative AI systems to make generated text identifiable in a machine-readable form. Beginning Monday, API customers around the world can choose to activate watermarking for select models, although the feature will remain switched off by default for API users. Eligible ChatGPT and Codex users in the European Union will begin receiving watermarked text over the coming weeks.
The technology could eventually provide another way for researchers, platforms and other organizations to assess whether OpenAI systems were involved in producing particular text. OpenAI, however, is cautioning against treating a detected watermark as definitive proof about how a piece of content was produced.
According to the company, a watermark cannot determine how much human editing or creativity was involved, who owns or is responsible for the material, who generated it, or whether the information itself is accurate. The absence of a watermark also does not prove that a person wrote the text without AI assistance.
OpenAI acknowledged that the technology has technical limitations. Short passages can be more difficult to detect, while editing can significantly weaken the watermark. In company testing, replacing portions of a watermarked passage with synonyms sharply reduced detection rates.
Because false positives and missed detections remain possible, OpenAI is not initially making the text detector generally available to the public. Approved researchers and expert organizations can apply for access, with applications being reviewed on a case-by-case basis. The detector will indicate whether an OpenAI watermark is detected without identifying a user or revealing their prompts or conversations.
OpenAI also said it plans to make the textGrain technology available as open source in the future so that others can build on the system.
The development adds another layer to the growing debate over how AI-generated material should be identified as increasingly capable systems produce text, images, audio and video that can closely resemble human-created content.
