Anthropic implements statistical watermarking in Claude text
August 12, 2026
Anthropic is embedding imperceptible watermarks in Claude's outputs by biasing statistical word choices. While the watermarks persist through simple copying, they are vulnerable to removal via manual rewriting or canonical regeneration.
HOW THIS AFFECTS YOU
●
designerYou may encounter subtle shifts in linguistic style due to these statistical biases.
●
policyYou should account for the technical limitations of these watermarks in content provenance frameworks.