Content provenance helps people understand where content came from, how it was created or edited, and whether it contains signals associated with our models. Today, we’re sharing our approach to text watermarking in response to the EU AI Act, and how it fits into our broader work.
Our phased approach reflects both the EU AI Act requirements as well as the technology’s limitations: Starting today, API customers globally will be able to opt in to text watermarking for select models. Over the coming weeks, we will add an invisible watermark to eligible ChatGPT and Codex text output in the European Union. We’re opening applications to access our text watermark detector.
Our text watermarking technology, text Grain, adds an invisible statistical signal to the model’s word choices. In our evaluations, text Grain matched or exceeded the performance of other approaches we tested, including Synth ID for text. Even so, strong performance under ideal conditions does not guarantee reliable detection in everyday use. Shorter or more constrained text is harder to detect. Editing can weaken the watermark. These limitations contribute to our decision to provide initial detector access only to approved researchers and expert organizations.
Source: OpenAI Blog · Summarized by HeadlinesBriefing