“`html
Anthropic has announced plans to embed invisible watermarks into text and images generated by its Claude AI model. The company says it will include machine-readable data in all Claude outputs to help identify artificial content and comply with European Union transparency requirements. These watermarks will not be visible to human readers but will allow people and online platforms to detect Claude-generated material.
The implementation will roll out gradually. While new Claude models will include watermarks from their launch date, existing models are still in development for the feature. The EU’s AI Act, which took effect in August, provides a four-month grace period for AI products released before that date, giving Anthropic time to prepare its systems.
Anthropic will use two different marking techniques depending on content type. Images will be marked using C2PA, a provenance standard already adopted by Adobe, OpenAI, and Google. Text will receive an “imperceptible watermark” woven directly into the content that will survive copying and pasting. The company plans to release technical documentation explaining how users can detect these watermarks.
This marking system represents a broader industry shift toward identifying AI-generated content online. However, experts note that such watermarks can sometimes be stripped away accidentally when content is uploaded to various platforms, potentially limiting their effectiveness.
“`
Originally reported by Theverge →
About the Author
