tech
Claude's new Scarlet Letter watermark is invisible—for now
The mark flags anything Claude processed, even human writing it only edited.
TL;DR
- Anthropic will watermark all content processed by its AI models, including edits to human-written text, to comply with the EU's AI Act.
- The watermarks are invisible to users but can be detected by specific tools, though they are easily bypassed through editing or by pasting content into other systems.
- The broad application of watermarks may lead to misinterpretations, potentially labeling lightly edited content the same as fully AI-generated content.
- The EU's AI Act requires labeling for AI-generated content, especially on matters of public interest, but Anthropic's approach marks content even when exemptions apply.
- Anthropic plans to release a text detection API for users to identify watermarks, but details on detection methods and testing results are not yet public.