AI WATERMARKS

Claude Text Watermarks Explained

Claims about Claude watermarks often mix together statistical signals, AI-detector scores and invisible Unicode. These are different mechanisms and should be evaluated separately.

01

Start with the claim being made

A statement that a model watermarks text can refer to several very different things. Ask whether the claim concerns an announced provider system, a statistical research method, metadata or merely a third-party detector result.

Without that distinction, ordinary patterns in writing can easily be mistaken for a provider-specific signal.

02

Hidden characters are a separate question

Zero-width spaces and other format characters can occur in copied text, but finding one does not identify the model or service that produced the passage.

Statistical watermarking generally concerns patterns across token choices. It does not require a character that can simply be selected and deleted.

03

Treat detector output carefully

AI detectors estimate whether writing resembles patterns in their training data. That estimate is not the same as checking a provider-controlled watermark.

  • Check the source and scope of the claim.
  • Separate visible text from invisible code points.
  • Keep uncertainty visible when reporting a result.
04

Use evidence, not assumptions

Provider documentation and reproducible technical evidence are stronger than screenshots, anecdotes or unexplained scores.

A careful review can describe exactly what was inspected without making unsupported claims about authorship.

REF

Further reading

Primary references used to keep this guide grounded.