What is burstiness in writing? (And why AI detectors use it)

Burstiness measures how much your sentence lengths vary. A document with high burstiness mixes short sentences with long sentences. A document with low burstiness has sentences that all sit in roughly the same length range. AI-generated text is famously low-burstiness — most sentences land between 15 and 22 words — because the underlying model has no reason to alternate between a four-word sentence and a 25-word one. Human writers do this all the time, often without thinking about it.

That's why burstiness is the second-biggest signal AI detectors use, right behind perplexity.

Run a free burstiness check on our detector → No card. Plain-text paste.

A worked example

Same paragraph, written two ways.

Low burstiness (sounds AI):

The student arrived at the testing center on Tuesday morning. She had prepared for the exam for several weeks beforehand. Her preparation included practice tests, flashcards, and review sessions. She felt confident as she sat down at her assigned desk. The proctor distributed the test booklets at exactly 9 AM. She began working through the questions methodically.

Every sentence is between 9 and 13 words. The rhythm is uniform. The pacing is mechanical. Even though nothing is technically wrong, the paragraph reads as flat — and an AI detector will register it as low-burstiness.

High burstiness (sounds human):

The student arrived Tuesday morning, exhausted. She'd been preparing for weeks — practice tests, flashcards, four study sessions with her roommate that ended in tears. She felt ready anyway. At nine, the proctor distributed booklets. She started.

Sentence lengths: 5, 18, 4, 6, 2. Same content, more or less. The variance is the whole difference.

Why AI has low burstiness

LLMs generate one token at a time. They have no concept of pacing. They have no instinct to "vary it up" because they don't know what came two sentences ago — they're working from a sliding window of context that biases them toward producing the most-likely next sentence given what they just produced.

The result, statistically, is that AI-generated text drifts toward a comfortable middle length. Some sentences end up a little longer; some a little shorter. But there's no extreme variation. There's no four-word punchline followed by a 30-word run-on. The standard deviation of sentence lengths in AI output is consistently lower than the standard deviation in human writing.

That regularity is what burstiness detection picks up. The detector doesn't need to understand the content. It just measures the variance.

The math, briefly

Burstiness is calculated by taking the standard deviation of sentence lengths in a document and dividing it by the mean. A higher ratio means more variance — more bursty. A ratio below ~0.5 reads as suspiciously uniform on most detectors. A ratio above ~0.8 reads as comfortably human.

You don't need to compute this manually. But the intuition matters: detectors aren't looking at any one sentence. They're looking at the spread.

Why this gets messy

Same problems as perplexity, slightly different shape.

Problem 1: Some genres have legitimately low burstiness. Legal contracts. Technical specifications. The methods section of a scientific paper. Each of these is full of sentences that sit in the same length range because the genre demands it. Humans write them. They flag as AI.

Problem 2: Short documents are unreliable. Burstiness is a statistical measure, and statistical measures need a sample. A 100-word document doesn't have enough sentences to give a meaningful variance estimate. Detectors are less reliable on short text in both directions — more likely to false-positive on human work, more likely to false-negative on AI work.

Problem 3: Editing can flatten burstiness. A writer who edits aggressively often reduces burstiness in the process. The instinct to "tighten" your sentences usually means cutting the shortest and longest ones to the middle. Heavy revision can make human writing measure more like AI. This is part of why polished, professional prose sometimes flags.

How to raise your burstiness

Without changing what you're saying:

Add at least one very short sentence per paragraph. Not every sentence needs to be a four-word punchline, but one per paragraph breaks the uniformity. A short sentence anywhere in a paragraph spikes the variance for that whole paragraph.

Combine some adjacent sentences. If you have three sentences in a row that are each 15 words, merge two of them with a comma or a dash into one 30-word sentence. The other one stays at 15. Now you have variance.

Vary your openings. Most AI sentences open with the subject. Try a sentence that opens with a participle ("Walking into the room…"), a prepositional phrase ("In the back corner…"), or a fragment used as a setup ("The thing about it. The complicated part."). Variance in openings correlates with variance in length.

End some paragraphs short. A paragraph that closes on a four-word sentence reads as deliberately paced. It also raises burstiness noticeably.

Read out loud and listen for monotone. The fastest test for low burstiness is auditory. If your paragraphs sound metronomic when you say them aloud, they measure metronomic too.

Burstiness + perplexity = how detectors actually work

Almost every major AI detector — GPTZero, Originality.ai, Copyleaks, Turnitin's AI checker — combines perplexity and burstiness as their two primary signals. The combination is more reliable than either alone.

The intuition: AI writes with predictable word choices (low perplexity) and uniform pacing (low burstiness). To flag as human, you need to raise both. Raising one and not the other is sometimes enough to clear a threshold, but rarely enough to clear it confidently.

For the deep dive on the perplexity half: What is perplexity in AI detection? For how GPTZero specifically combines the two: How does GPTZero work?

What HumanWriteup does to burstiness

The HumanWriteup rewrite operates on both signals at once. The model identifies sentences that sit in the AI-typical length range and splits some of them, merges others, and reshapes openings — without changing your meaning or your argument. The rewrite also handles perplexity by swapping high-probability words for less-predictable alternatives.

The result is text that measures higher on both perplexity and burstiness, which is what detectors actually require. Conservative mode is the default for essays and academic writing; Aggressive mode does heavier restructuring.

FAQ

What is burstiness in writing?

Burstiness is a measure of how much sentence lengths vary in a document. High burstiness means a mix of short and long sentences; low burstiness means sentences that sit in roughly the same length range. It is the second-biggest signal AI detectors use.

Why is AI writing low-burstiness?

Because language models generate text one token at a time without any sense of pacing. The output drifts toward a comfortable middle length, with much lower variance than human writers naturally produce.

Can I have high burstiness and still get flagged as AI?

Yes, if your other signals (especially perplexity) still look AI-generated. Burstiness is one of several signals detectors combine. Raising burstiness alone is sometimes enough to clear a threshold but rarely enough to clear it confidently.

Is editing my writing for burstiness the same as humanizing it?

Burstiness is one part of humanizing. The other big part is perplexity (varying word-choice predictability). Most humanizer tools, including HumanWriteup, target both signals at once.

How do I check my burstiness?

Most AI detectors report a perplexity score and either an explicit burstiness score or a combined likelihood that includes burstiness. The HumanWriteup free detector shows both.

Check your text's burstiness free on the HumanWriteup detector → 500 words/month, no card.