Anthropic adds invisible watermarking to Claude texts for EU AI Act
Claude's hidden watermark survives copy-paste and stays detectable after edits
Deep Dive
Anthropic is introducing an invisible watermarking system for texts generated by its Claude AI model. This watermark is embedded at the model level, is imperceptible to readers, and persists even after text is copied and pasted, remaining detectable after certain modifications. This measure aligns with the European AI Act, which became largely applicable on August 2, 2026, introducing new transparency requirements for AI development and use.
Key Points
- Watermark is embedded at model level during token generation, not as post-processing
- Persists after copy-paste and survives certain modifications like small edits
- Aligned with EU AI Act, largely applicable August 2, 2026, for AI transparency compliance
Why It Matters
Anthropic's watermarking keeps AI content traceable, helping enterprises meet EU AI Act transparency duties.