- 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
