A watermark is added on purpose
A text watermark is introduced during generation according to a particular scheme. A compatible verifier looks for that planned signal later.
The verifier and generator therefore share knowledge of a technique, key or expected statistical pattern.
A detector makes an inference
A detector examines text after it has been written and estimates whether it resembles machine-generated examples. It may operate without any relationship to the original generation system.
Because styles overlap and text can be edited, detector results can include both false positives and false negatives.
The outputs answer different questions
Watermark verification asks whether a supported signal is present. Detection asks whether text looks statistically similar to known generated writing.
- Check whether the method supports the claimed model.
- Look for a clear explanation of confidence and limitations.
- Never infer authorship from one automated result.
Choose language that matches the evidence
Report an observed character, verified signal or detector estimate by its correct name.
Precise language protects readers from conclusions that the underlying test cannot support.
Further reading
Primary references used to keep this guide grounded.