AI Detection Score

An AI detection score is the output of a classifier that has analyzed text for statistical signatures common to language model outputs. Scores are typically expressed as a percentage — '82% AI-generated' — or as a label like 'likely human,' 'mixed,' or 'likely AI.' What goes into the calculation varies by vendor and is generally not disclosed; ours combines explicit statistical heuristics with phrase patterns, described at /ai-detector/tells.

No detector publishes a threshold at which its score becomes meaningful, and any number quoted as one — including by this page, previously — is somebody's impression rather than a policy. Institutions set their own and rarely publish them. Scores between 30–70% are the gray zone where detectors disagree with each other most often. A single piece of text can score 45% on one tool and 91% on another — a reality that makes blanket policies based on detection scores problematic.

Our own detector's score before and after a rewrite tells you only what our own detector thinks, and it is the same engine on both sides, so it is not evidence that anything was fixed — only that the tells we look for are gone. We publish that detector's error rates, including a 96% false negative rate, at /ai-detector/accuracy. We make no claim about what any other detector would say, here or anywhere else: their terms prohibit the automated testing that would produce an honest number.