AI Detection Scoring

The score is built from five signals. The first is burstiness: the variance in sentence length across the document. Human writing tends to swing widely — a long subordinated sentence followed by a fragment. Most LLM output settles into a uniform mid-length cadence. Low variance, high score.

The second is token-level repetition density, computed over bigrams. Models tend to lean on the same two-word phrases (“it is important,” “in order to,” “moreover, the”) at characteristic rates. We measure how often a document falls into those grooves. The third is hedge-word density— the frequency of qualifiers like “may,” “could,” “arguably,” “tends to.” LLM defaults skew higher than most human writing.

The fourth is vocabulary entropy, measured as the n-gram predictability of the document under a baseline language model. Predictable token sequences score higher; surprising ones score lower. The fifth is the function-word distribution— the ratios of articles, prepositions, conjunctions. Function-word fingerprints are surprisingly stable within a corpus, and AI text has a recognizable one.

The five signals are combined into a weighted sum, normalized to 0–100, calibrated against an internal corpus of known-AI and known-human samples drawn from public datasets and our own test materials. Calibration is recalculated as the corpus grows.

Here is what we are explicit about: this is not Turnitin and it is not GPTZero. We do not have parity with commercial detectors. The score does not measure quality, factuality, or whether your professor will accept the work. It measures statistical patterns that correlate with machine generation under our heuristic, nothing more. Use it as a feedback loop, not as proof.