Burstiness

Burstiness measures how much writing varies across a document, most visibly in sentence length and structure. Human writing tends to alternate short and long sentences; AI-generated text is typically more uniform, producing low burstiness.

The numbers

Definition

Burstiness describes how much a piece of writing varies as it unfolds. QuillBot defines it as a measure of "how factors like sentence length, sentence structure, and word use shift throughout a document". A high-burstiness passage mixes short, punchy sentences with long, winding ones and shifts construction as it goes; a low-burstiness passage holds one steady rhythm from the first sentence to the last.

There is also a stricter technical reading. Pangram defines burstiness as "the change in perplexity over the course of a document": how much the moment-to-moment predictability of the text rises and falls as a language model reads it. In practice, most tools approximate burstiness with directly observable features, chiefly the variance of sentence lengths, which needs no model at all.

Why it became an AI detection signal

Human writers drift. GPTZero observes that people naturally "vary our writing patterns", partly because of the limits of short-term memory, while "language models have a significant 'AI-print' where they write with a very consistent level of AI-likeness". QuillBot makes the same point from the other side: generative AI "will often use more uniform, repetitive sentence structure".

That contrast made burstiness attractive to early detectors. If uniformity is the machine's tell, then measuring variation looks like a cheap authorship test, and GPTZero notes that perplexity and burstiness are "actually the main applications behind dozens of other AI detector apps".

Burstiness versus perplexity

The two are usually named together but operate at different levels. QuillBot's summary: "Perplexity measures predictability at the level of words, whereas burstiness measures predictability at the level of sentences and text structure." Perplexity asks how surprising each next word is to a language model; burstiness asks how much that surprise, and the surface rhythm of the text, changes across the whole document.

The two can move independently. Consistently unusual vocabulary delivered in identical sentence shapes gives high perplexity with low burstiness; plain vocabulary in wildly varying sentence lengths gives the reverse. Neither number implies the other, which is why detectors that relied on them tracked both.

Why burstiness alone is a weak AI detector

Burstiness is a real, measurable property of text, but it is weak evidence about authorship. Pangram points out that the metric is "relative to a particular language model", so the same passage scores differently depending on which model measures it, and that "text written by English language learners is lower perplexity and lower burstiness in general", which inflates false positives for non-native writers. A careful human writing in a deliberately plain, even register scores low too.

The field has moved accordingly. GPTZero states: "As of autumn 2023, GPTZero no longer uses perplexity and burstiness for its AI detection because we migrated to a deep-learning based architecture"; the pair survive only as one of the seven indicators in its newer model. Low burstiness tells you a text is rhythmically uniform. It cannot tell you why.

How ScriptGrain measures it

ScriptGrain treats burstiness as measurement, not verdict. Sentence rhythm and variation are among the stylometric attributes ScriptGrain analyses, and the free client-side analyser at scriptgrain.com/tools/writing-style-analysis computes them in your browser from your actual sentences rather than inferring them through a model.

ScriptGrain sells neither AI detection nor evasion, so it has no stake in what your burstiness implies. The number exists to describe your style: whether your rhythm is even or varied, and whether that matches the voice you are trying to keep.

Worked example

Two four-sentence paragraphs, word counts done by hand. Paragraph A: "The team reviewed the results. The data showed a clear trend. The manager approved the plan. The project moved to the next stage." Sentence lengths: 5, 6, 5, 7 words. Paragraph B: "Results came in. The team spent the whole afternoon arguing about what the numbers actually meant. Nobody agreed. By Friday the manager had approved a plan anyway." Sentence lengths: 3, 13, 2, 9 words.

Paragraph A: mean = (5 + 6 + 5 + 7) / 4 = 5.75 words. Squared deviations: 0.5625 + 0.0625 + 0.5625 + 1.5625 = 2.75. Variance = 2.75 / 4 = 0.6875; standard deviation = 0.83 words. Paragraph B: mean = (3 + 13 + 2 + 9) / 4 = 6.75 words. Squared deviations: 14.0625 + 39.0625 + 22.5625 + 5.0625 = 80.75. Variance = 80.75 / 4 = 20.1875; standard deviation = 4.49 words.

Both paragraphs average roughly six words per sentence, yet Paragraph B's rhythm varies more than five times as much (standard deviation 4.49 against 0.83). That spread, not the average, is what burstiness captures.

Sources

More terms

ScriptGrain · Why ScriptGrain · Sample voice profiles · The Journal · Support