What is burstiness in writing

Burstiness definition

Burstiness measures how much a writer's sentences differ from one another. The term comes from statistics, where it describes events that arrive in clusters instead of at a steady rate. In writing, the clusters are spikes in the rhythm, such as a 4-word sentence after a 38-word one or a question after a run of statements. Writers rarely plan for it. It appears because people move fast through easy points and slow down to explain hard ones.

Burstiness vs perplexity

Perplexity scores how predictable each word is. Burstiness scores how much that predictability, and the shape of the sentences, changes across the whole text. A language model gives every possible next word a probability, so text the model would have written itself gets low perplexity. A 2024 technical report from Checkfor.ai describes the original GPTZero as built on both measures, with burstiness defined as the variation of perplexity across a document. So a draft can be predictable sentence by sentence and still read as human, as long as those sentences vary.

Why AI scores low

AI-generated text has low burstiness because a language model picks the likely next word again and again. Each likely choice pulls the sentence toward an average length and a familiar structure. Training then rewards clear, balanced answers, and that pushes the model further toward the tidy middle. You end up with paragraphs where every sentence runs a similar length and opens with a subject and a verb. When I check a model draft, that sameness is visible before I count a single word.

A worked example

Here's a low-burstiness paragraph, with 4 sentences of 7 to 8 words each:

> Coffee contains caffeine, which affects the nervous system. Caffeine blocks adenosine receptors in the brain. This reduces feelings of tiredness during the day. Office workers drink it for better focus.

Here's the same content rewritten with more variation:

> Coffee keeps you awake for a chemical reason. Caffeine settles into the brain's adenosine receptors, the sites that collect a tiredness signal over the day, and blocks that signal from landing. You feel alert. Until it wears off.

To raise burstiness in an article, read the draft aloud and mark where the rhythm goes flat. Split a long sentence where a point should land hard. Then merge two short ones that explain a single mechanism. Swap one statement for a question or a fragment. Let the content decide the length, because random short lines added for effect read as forced.

How detectors use it

Detectors calculate a burstiness score and combine it with perplexity to estimate whether a model wrote the text. GPTZero, which Princeton student Edward Tian built in January 2023, made the pairing well known by scoring text on its "perplexity and burstiness", according to Business Insider. The signal is weak on short passages. It also misfires on formal or technical prose, where human writers keep their sentences even on purpose. Treat any detector score as a probability.

How Snoika measures it

Snoika scores burstiness on every draft as one of twenty linguistic measures. It sits under rhythm because it describes the same pattern as sentence length variance and paragraph length. All 3 work at different scales. Snoika's linguists set a human range for these measures, and a draft that falls outside it is rewritten before anyone reviews it. The goal is prose that reads like a person wrote it, which is a different goal from beating a detector.

Related terms

Burstiness overlaps with several other writing measures. Here's how each one differs from it.

Perplexity

Perplexity measures how predictable each word is to a language model. Burstiness measures how much that predictability and the sentence shape change across a text.

Sentence length variance

Sentence length variance is a plain count of how much sentence lengths spread around their average. It's the simplest proxy for burstiness, but it ignores structure and word choice.

Paragraph length

Paragraph length tracks rhythm at the block level. A page can vary its sentences and still feel flat if every paragraph runs to the same size.

AI detection

AI detection is the wider practice of estimating whether a model wrote a text. Burstiness is one input to that estimate, and modern detectors weigh it alongside other signals.