A watermark is an embedded signal
In ordinary use, a watermark is information added to content so its origin or processing history can be checked later. For generated text, that signal may be introduced through patterns in token selection rather than through a visible label.
The important distinction is intent: a watermark is deliberately added by a system. A detector, by contrast, tries to infer whether text looks machine-generated after the fact.
Statistical does not mean visible
A statistical text watermark can influence which words or tokens are selected while keeping the passage natural to a reader. A compatible detector then looks for the expected pattern across enough text.
That is fundamentally different from inserting a zero-width character such as U+200B. Hidden Unicode is part of the text data itself; statistical watermarking can operate through word-selection patterns without adding an invisible character.
What a result can and cannot tell you
Watermark detection is usually specific to a technique, model family or provider. Editing, translating, shortening or heavily paraphrasing text may also change the strength of a signal.
- A supported signal can provide useful provenance context.
- No signal found does not prove that a person wrote the text.
- A signal alone does not establish who prompted, edited or published the final passage.
Use inspection as one piece of context
A responsible workflow separates visible wording, invisible characters, formatting artifacts and supported provenance signals. Review each finding on its own terms before deciding what it means.
TextTrace is designed around that separation: inspect the text layer, understand what was found, clean supported formatting issues and avoid turning a limited signal into a sweeping claim.
Further reading
Primary references used to keep this guide grounded.