- Anthropic has detailed how its new invisible watermarking system for Claude text outputs operates following widespread user criticism.
- The system subtly influences low-stakes word choices using a cryptographic key to create a pattern detectable only by specialised software.
- Introduced to comply with new European AI regulations, Anthropic plans to roll out the watermarking technology globally across all Claude responses.
- The company clarified that light editing or proofreading will not generate enough altered words to trigger a false positive watermark detection.
- Tech commentators and users remain critical of the update, arguing that altering model outputs for compliance compromises the quality for end users.
IN FULL
Anthropic explains how Claude’s controversial new ‘watermarking’ feature works
