EmbeddingGemma 2: an open, lightweight multimodal embedding model
deepmind.google·6d ago
AI Engineering
Hacker News·Aug 10·5 min read
How Claude marks AI-generated content
TL;DR
Anthropic is implementing machine-readable markings for AI-generated content in compliance with the EU AI Act, enhancing transparency for users. This initiative will apply to Claude models launched after August 2, 2026, and will also retroactively support earlier models.
✦ Why It Matters
Ensure your AI-generated content complies with transparency standards by implementing machine-readable markings today.
Key Takeaways
- Implement machine-readable marks in your AI-generated content.
- Understand the limitations of Claude's marking system.
- Stay updated on technical guidance from Anthropic.
- Assess compliance with the EU AI Act for your applications.
How It Works
Claude employs two techniques for marking content: embedded watermarks in text and signed provenance metadata for files. The watermarks are woven into the text, remaining imperceptible and intact during copying and editing.
For file types like .svg or .jpg, signed provenance metadata is attached, following the C2PA standard, which helps verify the content's origin and integrity.
⚠ The Catch
A detected mark indicates that content may have been processed by Claude but does not confirm full provenance. Content can carry a Claude mark even if it was not originally authored by Claude or has been modified after processing.