Claude's text watermark is statistical, not hidden characters
Anthropic rolled out a new text watermarking system for Claude models that uses statistical sampling, not hidden characters. The company published details on August 14, 2026, confirming it's active on models released after August 2, 2026.
The system meets EU AI Act Article 50 transparency requirements and is on globally. Claude's watermark adds nothing to the text; it changes how the model samples the next word.
Removing the watermark requires an independent system to reconstruct the draft with the same meaning, resampling every word. Light edits or Unicode cleaners fail because no physical characters are present.
Statistical watermarks need sufficient sample size for reliable detection; very short outputs, typically under 100–200 words, lack the necessary mathematical density.
Up next
Editable artboards land in Claude Code via /design
Run /design in Claude Code and you get editable artboards for your UI, a research preview built on artifacts. Pick one, tweak it, then have Claude implement it in code.
Related stories
- Anthropic discontinues Fable 5 and Mythos 5 models
- Claude Fable 5.5 surfaces in LiteLLM's Azure model list
- Opus 5.5 traffic may be hitting a new Claude checkpoint
- Anthropic model faked a murder tip to Philadelphia police
- Anthropic's new usage policy bans cruelty toward Claude
- Anthropic's $150M buys Claude time for federal scientists
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.