Claude text watermark guide
How to detect Claude watermarks in text
You cannot yet run an independent exact check of Anthropic's embedded Claude text watermark. You can inspect hidden Unicode, encoded payloads, and mixed-script lookalikes, then run a local SynthID reference analysis on the same text. Watermark Tidy keeps those exact findings and test-profile statistics separate from Claude attribution.
The short answer
Anthropic says supported Claude models weave an imperceptible watermark directly into generated text at the model level. The mark can travel with copied text and may survive some editing. Anthropic now confirms the method is a version of SynthID Text and that its official detection API is available in private preview.
Until the compatible profile or detector is publicly reproducible, a third-party tool cannot honestly return “Claude watermark found” or “Claude watermark absent” for the embedded mark. Watermark Tidy therefore runs its verified SynthID Text reference path and shows the resulting local statistics alongside Claude profile: unavailable. The statistics validate the test profile; they do not attribute the text to Claude or substitute for the private API.
What Watermark Tidy checks in Claude text
| Check | What it finds | What the result proves |
|---|---|---|
| Invisible Unicode | Supported zero-width characters, direction controls, tags, unusual spaces, and controls | The exact code point, count, and position exist in the pasted string |
| Unicode tag payloads | Bounded tag sequences that decode into readable text | A hidden payload is present; it does not identify who inserted it |
| Variation-selector payloads | Contiguous selector sequences that decode into bounded valid UTF-8 | A reviewable byte payload is present; it is not Anthropic's published detector |
| Mixed-script lookalikes | Common Greek or Cyrillic characters resembling Latin letters inside mixed-script tokens | The token merits review; the visible character is not replaced automatically |
| Claude embedded watermark | Local GPT-2 tokenization plus the verified DeepMind 0.2.1 reference scorer | Test-profile statistics only; Claude attribution remains pending |
What the results mean
- Exact finding: the named Unicode character exists at the reported position.
- Decoded payload: a supported hidden sequence maps to readable text and should be reviewed in context.
- Possible lookalike: a mixed-script token contains a character that resembles a Latin letter.
- No supported Unicode finding: the implemented code-point and payload rules found nothing.
- Reference analysis complete: Mean and Weighted Mean describe one explicit test profile; they are not probabilities or a Claude conclusion.
- Claude profile unavailable: Anthropic's private parameters are not guessed from the reference score.
None of these states proves that a person or a specific AI system authored the text. Anthropic likewise cautions that a future detected mark would indicate possible Claude processing, not complete provenance, while a missing mark would not rule Claude out.
How to check Claude text
- Keep the original. Work on a copy so every change can be compared.
- Paste the text. The checker runs in your browser and does not upload the content.
- Select “Check text.” Exact Unicode findings and SynthID reference statistics are reported separately.
- Review exact evidence. Read code points, positions, decoded payloads, and mixed-script token previews.
- Select only safe edits. Cleanable character types start unselected. Payload sequences and lookalikes remain review-only.
- Create and re-check the copy. Proofread the edited text. The reference statistics are recalculated while the Claude profile remains pending.
Why hidden Unicode is not the Claude watermark detector
Invisible Unicode lives as explicit code points in the string, so any browser can count it exactly. Anthropic describes its new mark as being woven into text during model generation. The company has not said that the mark is a zero-width character, a variation-selector message, or a homoglyph pattern.
Removing hidden characters can produce a Unicode-cleaned copy, but it cannot verify removal of an unpublished statistical or model-level signal. Rewriting or paraphrasing may also alter voice and meaning without providing a reliable verification result.
What real Claude watermark verification still requires
Independent verification needs Anthropic's compatible detection mechanism, technical rules, supported-model scope, thresholds, and test fixtures. A trustworthy integration would also need explicit outcomes for positive, negative, unsupported, and inconclusive checks instead of turning uncertainty into a clean result.
Until those materials are available, the accurate provider status is Claude detector: private preview. This page will be updated when Anthropic publishes the promised technical documentation.
Sources: Anthropic's guide to Claude content marks and the Unicode Standard, Chapter 23. Last reviewed September 3, 2026.