It is a word-pattern signal, not a pasted hidden character.
News coverage of Anthropic's announcement describes an imperceptible watermark applied at the model level. The important distinction is that a statistical watermark changes the distribution of generated word or token choices while keeping the text readable.
Searching for zero-width spaces cannot confirm whether this statistical signal is present.
A detector needs more than the visible words.
Reliable verification depends on a detector designed for the exact watermarking method and access to the relevant secret or service. A general AI-writing score is not the same thing as reading a watermark.
- Copying and pasting does not automatically prove removal.
- Light editing may weaken a statistical signal, but absence cannot be inferred by eye.
- A positive signal would indicate model involvement, not who authored or approved the final work.
What WatermarkScope can tell you today.
The text inspector can show observable Unicode formatting characters and remove low-risk anomalies. It does not assign a Claude-generated probability. For files, the C2PA checker can read signed provenance credentials where present.