How do AI detectors work?
By Sourabh Singh18 September 20267 min read
Strip away the marketing and every detector in this market reduces to two measurements.
Every AI detector on the market, GPTZero, Turnitin, Originality.ai, Copyleaks, is built on the same two ideas. Vendors describe them differently and weight them differently, but nobody has found a third signal that works better.
Signal one: perplexity
Perplexity is a measure of how surprising each word is given the words before it. Language models generate text by repeatedly picking a likely next word, so their output is, by construction, less surprising than human writing. A person reaching for an odd but exact word produces a spike a model rarely produces.
This is also why heavy editing genuinely lowers a detection score rather than merely fooling the detector. When you replace a predictable word with a specific one, the text really has changed in the dimension being measured.
Signal two: burstiness
Burstiness is the variance in sentence length across a passage. Human writing is lumpy: a long sentence, then a short one, then a medium one. Model output clusters around a central length and stays there. Measured across a few hundred words, the difference is stark.
What detectors add on top
- A dictionary of stock phrases that appear disproportionately in model output.
- Vocabulary range, usually measured length-stably so long and short texts stay comparable.
- Repeated sentence openers, which models produce far more often than people do.
- Punctuation profile, including the much-discussed em dash.
Why they are unreliable
Both core signals describe a style, not an origin. A careful writer producing plain, even, professional prose scores as machine-written because their writing genuinely has low perplexity and low burstiness. This is why non-native English speakers are flagged at materially higher rates: writing in a second language tends to mean a narrower vocabulary and steadier sentence construction.
No amount of model improvement fixes this, because the signal is not a mistake. The detector is measuring exactly what it claims to measure, the measurement just does not mean what institutions want it to mean.
See both signals measured on your own text, per sentence.
Measure your draftCommon questions
They are built on two measurements: perplexity, how surprising each word is, and burstiness, how much sentence length varies. Detectors then add things like stock phrase lists, vocabulary range, repeated sentence openers and punctuation profile.
Perplexity measures how surprising each word is given the words before it. Model output is less surprising than human writing, because models pick likely next words.
Burstiness is the variance in sentence length across a passage. Human writing is lumpy, with long and short sentences mixed, while model output clusters around a central length.
Both core signals describe a style, not an origin. A careful writer producing plain, even prose scores as machine-written, and non-native English speakers are flagged at materially higher rates.
