Perplexity and burstiness
By Sourabh Singh18 September 20266 min read
Two words the detection industry uses constantly and explains badly.
If you have read anything about AI detection you have met these two words. They are usually defined in a sentence and then left alone. They are worth understanding properly, because they are the whole mechanism.
Perplexity
Perplexity measures how surprising each word is given the words before it. "The cat sat on the" is very likely followed by "mat", low perplexity. A human writer might follow it with "radiator", which is a spike.
Language models generate by choosing likely continuations, so their output is systematically less surprising than human writing. That is not a flaw in the models; it is what they are for.
Burstiness
Burstiness is variance in sentence length. Humans write unevenly, a long, winding sentence with several clauses, then four words. Models cluster around a central length and stay near it.
Measured over a few hundred words this is one of the most reliable differences between human and model prose, and it is also the easiest thing for a writer to change deliberately.
Changing them honestly
- Burstiness: deliberately follow long sentences with short ones. Read aloud, flat rhythm is audible.
- Perplexity: add specific concrete detail. Names, numbers, dates, examples. Not rare synonyms, which raise perplexity while destroying readability.
The limitation worth remembering
Both describe a style, not an origin. A careful human writer producing even, predictable prose scores low on both, which is the entire reason false positives exist and why no detector should be treated as evidence.
Measure both on your own writing, per sentence.
See your numbersCommon questions
Burstiness is the variance in sentence length across a text. Humans write unevenly, mixing long winding sentences with short ones, while models cluster around a central length and stay near it.
Perplexity measures how surprising each word is given the words before it. Language models pick likely continuations, so their output is systematically less surprising than human writing.
For burstiness, follow long sentences with short ones and read aloud to hear flat rhythm. For perplexity, add concrete detail such as names, numbers, dates and examples, not rare synonyms, which hurt readability.
Yes. Both describe a style, not an origin, so a careful human writer producing even, predictable prose scores low on both. That is why false positives exist and why no detector should be treated as evidence.
