Text watermarking changes how a model selects among acceptable next tokens. A secret or defined rule slightly favors some choices, causing a statistical pattern to accumulate across a sufficiently long passage without adding a visible label.
A compatible detector measures whether the expected pattern appears strongly enough to be unlikely by chance. Watermarking can support disclosure, but removal tools may rewrite or transform content until the detectable signal becomes too weak.
ELI5
Text watermarking gives AI-written text a hidden statistical pattern. The words still look ordinary, but a detector that knows what pattern to check can look across a long passage for signs that a model helped produce it.
For example, a model may quietly prefer certain equally suitable words more often than it normally would. Finding that preference can be evidence of model involvement, but it cannot prove that every word was generated or identify how much a person edited.
