tech
Anthropic shares more details about how Claude’s new watermarks will work
How will the watermarking actually work? Can it be hidden with editing? And how does this affect code?
TL;DR
- Anthropic is detailing its watermarking system for Claude to comply with the EU AI Act's Transparency Code.
- Watermarking involves embedding undetectable patterns in text during low-stakes word choices.
- The system does not impact the quality of Claude's output and is indistinguishable to readers.
- Anthropic plans to use the SynthID-Text approach and release a watermark detection API.
- Watermarking differs from AI detection methods that look for writing "tells".
- Light editing may not remove the watermark, but a complete rewrite will.
- The impact on edited text depends on the extent of Claude's revisions.
- Watermarking will have a negligible effect on actual code produced, though comments may be marked.
- Other major AI model developers are also implementing watermarks under the same Code of Practice.