Guide · Updated 19 November 2026

Claude text watermark

Claude marks its text worldwide with an invisible watermark, a version of Google DeepMind's SynthID-Text, across the API, the Claude apps, Claude Code, Claude Cowork, and Claude Tag. Detection runs through an API in private preview for regulators, media, researchers, and other approved groups.

Claude models and surfaces

Anthropic announced the watermark on 14 August 2026 and applied it globally at launch. Its help center now lists every current model: Fable 5.2, 5.1, and 5; Mythos 5.2, 5.1, and 5; Opus 5.5 through 4.5 and Opus 3; Sonnet 5.5 through 4.5; and Haiku 5.5 and 4.5. Older models joined on 3 and 16 November 2026, and the mark also applies through AWS, Google Cloud, and Microsoft Foundry.

The SynthID-Text method

SynthID-Text runs a keyed knockout tournament between candidate words as Claude writes, and the winners leave a statistical pattern only the key holder can detect. The signal grows with the number of real word choices, so longer, more varied text carries a stronger mark.

Claude watermark limits

  • Short samples carry too little signal to detect reliably.
  • Factual passages and code carry a sparser mark.
  • Light edits keep the mark, and a complete rewrite removes it.
  • Claude's own translations carry the mark, and text translated afterward can lose it.
  • On the 4.x models and Opus 3, non-default sampling settings may leave text unmarked.
  • A missing mark doesn't prove a human wrote the text.

Claude watermark detection

Anthropic's detection API is in private preview, by request form, for regulators, law enforcement, media, fact-checkers, researchers, education, EU civil society, and select enterprises.

Marketing copy

Readers never see the watermark. What they notice is Claude's default voice drifting back into every section. See why AI can't hold a brand voice and AI text watermarks compared.

Claude watermark FAQ

Does Claude watermark its text?

Yes. Anthropic marks text from every current Claude model worldwide, across the API, the Claude apps, Claude Code, Claude Cowork, and Claude Tag.

Can I turn off Claude's watermark?

Anthropic's help center lists no opt-out. On the 4.x models and Opus 3, non-default temperature, top_p, or top_k settings may leave text unmarked.

Does Claude Code watermark code?

Claude Code is a covered surface, and Anthropic says code carries less watermark than prose.

Can I check whether text came from Claude?

Only through Anthropic's detection API, in private preview for regulators, law enforcement, media, fact-checkers, researchers, education, EU civil society, and select enterprises.

3 agency-grade versions of your site, with conversion copy and art, for a one-time $399.

Generate 3 versions

Sources, checked 19 November 2026

  • anthropic.com/news/claude-text-watermark (14 Aug 2026, updated 1 Sep 2026)
  • support.claude.com (How Claude marks AI-generated content)
  • platform.claude.com/docs (release notes; text watermark)
  • Dathathri et al., Scalable watermarking for identifying LLM outputs, Nature 634, 2024