Our approach to EU text provenance rules (openai.com)

OpenAI has announced its approach to text watermarking, aligning with the EU AI Act, which mandates that generative AI outputs be identifiable in a machine-readable format. This initiative aims to enhance content provenance, allowing users to trace the origin and authenticity of text produced by their models. Starting today, API customers can opt into text watermarking for specific models, with the feature being off by default. Over the coming weeks, an invisible watermark will be applied to outputs from ChatGPT and Codex within the European Union, while access to the text watermark detector will be granted to approved researchers for further evaluation. This development is significant for the AI/ML landscape as it promotes transparency and responsibility in AI-generated content, addressing concerns about misinformation and the provenance of digital creations. OpenAI's watermarking technology, known as textGrain, embeds a subtle statistical signal into the text, which can be detected to verify authenticity. Initial evaluations suggest that textGrain offers solid performance, but challenges persist, particularly with shorter or more tightly constrained text. As the technology evolves, OpenAI plans to make it available as open source, inviting further advancements and broader community engagement in ensuring responsible AI practices.