OpenAI will watermark ChatGPT text in the EU but makes it optional for API users worldwide
decoder ·
OpenAI is implementing invisible watermarks for ChatGPT text in the European Union to comply with the EU AI Act, using a technology called textGrain that embeds statistical signals in word choices. Unlike Anthropic's mandatory global watermarking for Claude, OpenAI makes watermarking optional for API users worldwide, allowing them to turn it off. The company reports that detection rates vary significantly based on text length and subject matter, with lower accuracy for math content compared to psychology topics. Editing text substantially reduces detection capabilities, as replacing just 25% of words can drop identification rates below 20%. OpenAI claims that enabling watermarks does not negatively impact model performance or output quality across various benchmarks. Initially, access to the detector will be restricted to selected researchers and organizations under the EU Code of Practice to prevent misuse. The system is designed to only confirm the presence of a watermark without identifying users, assigning responsibility, or verifying content accuracy.
decoder