What are perplexity and burstiness?
Most AI-text detectors lean on two ideas: perplexity and burstiness. Understanding them tells you a lot about why some writing reads as "AI" and some doesn't.
Perplexity: how predictable the words are
Perplexity measures how "surprised" a language model is by a piece of text — essentially, how predictable the word choices are. Because a language model generates text by choosing likely words, its own output tends to have low perplexity: everything is what the model expected.
Human writing is messier. People pick words for reasons a model doesn't optimise for — memory, rhythm, humour, a very specific fact — so human text usually scores higher perplexity. As a rough rule of thumb, detectors treat higher perplexity as more likely to be human.
Burstiness: how much sentences vary
Burstiness measures variation in sentence structure across a passage. The classic version is the standard deviation of sentence length (often combined with variation in complexity):
burstiness = standard deviation of sentence length ÷ mean sentence length
High burstiness = aggressive alternation between short, punchy sentences and long, complex ones — very human. Low burstiness = uniform length and structure — very AI.
How detectors combine them
If a document is consistently low perplexity (predictable words) and consistently low burstiness (uniform sentences), a detector flags it as likely AI. Neither signal alone is decisive; it's the combination and consistency that matter.
What this means for your writing
You can't (and shouldn't) chase a detector, but you can write better, which naturally moves both metrics:
- Raise perplexity honestly — use specific, concrete words instead of generic ones; add a real detail only you would know.
- Raise burstiness — follow a short sentence with a long one. Let some thoughts land in three words. Let others unfold across a full line when the idea genuinely needs the room.
Naturable reports a Sentence rhythm score based on burstiness — pure arithmetic, no model required — so you can see your variation at a glance. See how it works.
An honest caveat
Detectors are imperfect and change often. Perplexity and burstiness are useful lenses, not a verdict, and no tool can guarantee how any specific detector will score a given text.
Related: How to remove AI tells · Do AI humanizers actually work?