OpenAI to implement statistical watermark for AI-generated text in EU
OpenAI will begin applying a statistical watermark to text generated by its models in the European Union. The system aims to help identify AI-produced content, but the company notes limitations in its detection capabilities.
AI company OpenAI announced it will start applying an invisible statistical watermark to text generated by its models within the European Union. This approach, named textGrain, is intended to help identify content produced by systems like ChatGPT and Codex, with deployment expected in the coming weeks.
The watermark modifies the statistical patterns of word selection without adding visible marks or hidden characters. A separate detector checks for the presence of these patterns. Initially, access to the detector will be restricted to approved researchers and expert organizations. OpenAI stated the detector will only indicate whether an OpenAI watermark is detected and will not identify users or their prompts.
OpenAI acknowledged significant limitations to its watermarking technology. Tests indicate the detector is more reliable for longer text passages, with performance dropping for shorter, highly constrained, or heavily edited texts. Replacing just 10% of words in a 400-token passage with synonyms reduced detection rates from approximately 92% to 66%.
The company emphasized that a watermark does not establish ownership, accuracy, or authorship. Furthermore, the absence of a detected watermark does not prove human authorship. These limitations are cited as reasons for not making the detector publicly available at launch.
This move aligns with EU regulations, which require generative AI providers to make generated text identifiable in a machine-readable way. OpenAI currently positions the watermark as a signal that an OpenAI system may have generated or processed text, not as definitive proof of AI authorship.