Our approach to EU text provenance rules

OpenAI has announced its approach to text provenance and watermarking in response to the EU AI Act, which requires generative AI providers to make generated text identifiable in a machine-readable way. Starting immediately, API customers globally can opt in to text watermarking for select models, though it remains off by default. Over the coming weeks, OpenAI will add invisible watermarks to eligible ChatGPT and Codex text outputs in the European Union. Access to the text watermark detector is initially limited to approved researchers and expert organizations to help evaluate and improve the technology, given its early stage and limitations. The technology, called textGrain, adds an invisible statistical signal to a model's word choices. Evaluations show that text watermarking faces challenges: shorter or constrained text is harder to detect, and editing text by replacing words with synonyms significantly weakens the watermark signal. OpenAI emphasizes that watermarks have strict limitations, noting that they do not measure human contribution, establish ownership or responsibility, identify users, or verify accuracy. Furthermore, the absence of a detected watermark does not prove human authorship. Alongside text watermarking, OpenAI continues to provide publicly accessible verification tools for audio and images, such as the openai.com/verify web tool and the Content Provenance API. OpenAI plans to make the textGrain technology available in open source and will revisit its phased approach as technology, standards, and evidence evolve.