Burstiness and perplexity

Also called: perplexity score, burstiness score, AI detection metrics

Perplexity measures how surprising a text is to a language model and burstiness measures how much that surprise varies from sentence to sentence, and AI-detection tools use both because model output tends to be both unsurprising and uniformly so.

Low perplexity plus low burstiness is the signature detectors look for; neither number identifies an author.

The two metrics answer different halves of one question. Perplexity asks how predictable the words are. Burstiness asks whether predictability rises and falls the way it does in human prose, where a plain sentence often sits next to a strange one.

Neither is evidence about a person. A detector that flags a text says the text is statistically regular, which is also true of careful non-native writing, of legal boilerplate and of anyone writing under a tight house style. That is the source of most false accusations.

MetricMeasuresHigh value meansCannot establish
PerplexityHow unexpected each token is to a reference modelText is surprising to that modelWho wrote it
BurstinessVariance of perplexity across sentencesMix of plain and unusual sentencesWhether variation was deliberate
Both togetherRegularity of the text overallReads as human-irregularIntent, authorship, or tool use
What each metric measures and what it cannot establish

Frequently asked questions

Can I raise burstiness to beat a detector?
Mechanically varying sentence length changes the number without making the text yours. It also tends to damage the writing, which is a poor trade for a metric no reader sees.
Why do non-native writers get flagged more often?
Careful second-language writing is often more regular than native writing — a narrower vocabulary band, more conventional constructions — which is the same statistical shape detectors treat as machine-like.
Is stylometry the same as perplexity-based detection?
No. Stylometry compares a text to a specific author's habits; perplexity-based detection compares a text to a general model's expectations. Different reference points, different questions.

Read further

Sources and how to cite this page

Cite this page

MyWritingTwin. “Burstiness and perplexity.” Last updated 2026-09-15. https://www.mywritingtwin.com/glossary/burstiness-and-perplexity