AI Generated Image
Anthropic defends and explains its AI watermarking feature
Business Insider • • business
Key Points:
- Anthropic is implementing a watermarking system in its Claude AI model to comply with the EU's AI Act, applying it globally due to the lack of regional scoping methods, with plans to extend watermarking to older models in the coming months.
- The watermark works by subtly influencing the AI's word choice through a specialized random number generator, making AI-generated text detectable without altering readability or introducing visible markers; an API will be provided for verification once the watermark "key" is released.
- Anthropic emphasizes the watermark's limitations, noting it only identifies text likely generated by Claude specifically, with longer texts easier to detect and factual or code outputs less frequently watermarked to preserve accuracy and functionality.
- Despite initial customer concerns and some subscription cancellations, Anthropic reports no significant increase in cancellations and clarifies that watermarking does not affect user ownership or authorship rights over the content.
- The watermark may not be completely removed by light editing but can be eliminated by thorough rewriting, raising questions about the future detection and user interaction with AI-generated content.
