tech
Anthropic defends and explains its AI watermarking feature
Anthropic addresses concerns about its new watermark, which will affect Claude users globally. It said the watermark is to comply with the EU AI Act.
TL;DR
- Anthropic is implementing watermarks on text generated by its Claude AI model.
- The watermarking is primarily to comply with the European Union's AI Act.
- The technology embeds subtle patterns in word choices, detectable with a specific 'key'.
- The watermark will be applied globally to new and older Anthropic models.
- Concerns about customer cancellations have been addressed, with Anthropic stating no significant trend has been observed.
- The watermark indicates Claude's involvement but does not identify AI in general.
- Limitations exist for factual passages and AI-generated code, where watermarking may be less frequent or effective.
- Watermarks do not affect user rights or ownership of the processed content.
- A complete rewrite of AI-generated text could remove the watermark.