Archive — US — techcrunch
Anthropic shares more details about how Claude’s new watermarks will work
Anthropic has unveiled details on watermarking text generated by its chatbot Claude to comply with the EU AI Act’s Transparency Code, which mandates identifying AI-generated content. The watermark remains undetectable to readers but can be identified through a special key. Users have voiced concerns, with some canceling subscriptions due to these changes, but Anthropic asserts that watermarking will not affect the quality of output.
Key points
- Anthropic published details on watermarking to comply with EU AI Act's Transparency Code.
- Watermarking is undetectable to readers but identifiable to those with a key.
- Users are concerned about the implications, with some threatening to cancel subscriptions.
- Watermarking does not affect the quality of the text produced by Claude.
- The watermark can be lost through substantial editing of the text.
- Watermarked responses are similar to unwatermarked ones for readers.
- Codes generated will have minimal watermarking due to the nature of coding.
Why it matters
- The move aligns with growing regulatory demands on AI technologies, especially in the EU.
- The EU AI Act reflects wider concerns regarding transparency in AI-generated content.
- Similar watermarking practices may emerge across other AI developers as compliance becomes standard.
- This trend parallels increasing scrutiny of AI technologies in the context of misinformation and authenticity.
Anthropic's introduction of watermarking illustrates a critical shift towards transparency in AI-generated content, catering to regulatory demands while sparking user concerns. This approach sets a precedent for other AI developers, emphasizing the need for accountability in the tech landscape.