AI writing detectors are probabilistic, not proof: how to read their scores and use them honestly

AI writing detectors return a probability, and probability is not proof. A high score does not establish that a text was written by AI, and a low score does not establish that it was written by a human. This method explains what the scores can and cannot show, and how to use them without making a false accusation.

What the scores can show: statistically, some text patterns resemble machine output more than typical human writing. What they cannot show: authorship. The same pattern can come from a translated text, a non-native draft, or a heavily edited document, and detectors are known to produce false positives on those.

The honest workflow: treat the detector as a triage flag, not a verdict. A high score starts a human review of the source and the writing history; it does not end one. Keep the raw score, the model and version used, and the date, because scores are not stable across detector versions.

The failure signals that mean the score is being misused: a decision based on the number alone, no documented review, or a score presented as certainty. Each of these turns a probabilistic tool into an unreliable accusation.