Anthropic, a leading AI developer, has announced the introduction of a text watermarking scheme designed to comply with the European Union's AI Act. This innovative approach modifies the word choices of its AI models to embed a detectable identifier, ensuring that generated text can be traced back to its source without compromising its meaning or integrity.
Key Insights
The watermarking scheme relies on altering inconsequential word choices in AI-generated text, allowing for the embedding of a digital signature that can be detected using a specific key. Anthropic emphasizes that this watermarking has a negligible impact on the speed and cost of its models, making it an efficient solution for regulatory compliance. While the scheme is semi-effective and can be defeated with significant editing, it represents a crucial step towards meeting the EU AI Act's requirements without incurring additional costs.
