Understand the difference between AI text watermarks and statistical AI-likelihood analysis — and how this tool approaches both.
Paste the text you want to analyze.
The system checks supported watermark and provenance signals.
Receive a transparent report explaining what was detected and how confident the system is.
Modern text watermarking is typically statistical rather than visible. It does not insert hidden characters or invisible Unicode marks in the common sense. Instead, it works at the level of token selection inside the language model.
During generation, the model assigns probabilities to possible next tokens. A watermarking algorithm can slightly bias those probabilities according to a secret key or pattern. The resulting text still looks natural, but a matching detector that knows the same scheme can later test whether the observed token sequence is more consistent with the watermarked distribution than with the unwatermarked one.
This produces a confidence score or hypothesis test result — not a binary “yes/no” that is always correct. Short texts, paraphrasing, translation, and heavy editing can weaken or destroy the signal.
Watermark detection looks for a specific, intentionally embedded machine-readable signal that a generator chose to insert. It is provider- or scheme-specific. If the detector and the generator do not share the same watermarking method (and usually a key), detection fails.
AI-likelihood analysis is a statistical estimate based on surface features of the text — vocabulary diversity, sentence-length consistency, repetition patterns, and similar indicators. It does not look for a watermark. It only estimates how “AI-like” the writing appears. It can be wrong, and it is never proof of authorship.
This product keeps these two concepts separate. A missing watermark does not prove that text was written by a human.