What Is Burstiness in AI Writing?

Burstiness is a measure of variation. Specifically, it captures how much sentence length and structural complexity fluctuate across a passage of writing.

Human writing is bursty. We write a long, winding sentence that accumulates clauses and qualifications as the thought develops. Then a short one. Then something in between, before doubling back to add the caveat we forgot.

Model output typically isn't. It tends to settle into a comfortable range — many sentences of similar length, similar construction, similar weight — and stay there for paragraphs at a stretch.

Published analyses put human writing in a burstiness range of roughly 0.6 to 1.2, with model output clustering around 0.2 to 0.4. The exact numbers vary by how the metric is computed, but the direction of the gap is consistent.

Why the gap exists

It falls out of how models generate text.

At each step, a language model predicts a probability distribution over what could come next, then samples from it — weighted toward the likely options. Repeat a few hundred times and you have a paragraph. The process is uniform, so the output is too. Nothing in it corresponds to a writer deciding a section needs a jolt, or getting bored of their own rhythm, or cutting a sentence in half on the third read.

Human writing carries the residue of a non-uniform process: revision, mood, uneven attention, a particular argument needing a particular shape. Burstiness is, in effect, a measurement of that residue.

Burstiness and perplexity are different things

They get mentioned together constantly and they measure different properties.

Perplexity operates at the word level: how surprising is each word, given the ones before it? Low perplexity means predictable word choice.

Burstiness operates at the passage level: how much does the structure vary from sentence to sentence? Low burstiness means uniform rhythm.

Text can be low on one and not the other. A technical manual has low perplexity — the vocabulary is constrained and predictable — but may have perfectly normal burstiness, with short procedural steps between long explanatory passages. A stylistically monotonous piece of creative writing could show unusual word choices with very even structure.

Detectors typically combine both signals, which is why neither number alone tells you much.

What burstiness looks like in practice

Consider a paragraph with sentences of 18, 21, 19, 22, and 20 words, each following subject-verb-object with a subordinate clause attached. Low burstiness. Every sentence has the same shape.

Now: 31 words, then 6, then 24, then 4, then 15. Some starting with the subject, one with a dependent clause, one a fragment. That's a bursty passage. It's also, generally, easier to read — the variation gives the reader somewhere to breathe and signals which points carry weight.

This is the part worth holding onto. Burstiness isn't an arbitrary detector artefact. It correlates with something real about readable prose. Uniform rhythm is genuinely flatter to read, whoever produced it.

The limits of the metric

Burstiness is a weaker signal than the discourse around it suggests, for several reasons.

It's easy to manipulate. Sentence length variation can be introduced mechanically, without changing anything meaningful about the text. Any tool — including ours — can raise a burstiness score without making the writing better. That's a reason to be skeptical of burstiness-based claims from humanizer vendors generally.

Genre moves it around. Legal writing, technical documentation, and formal academic prose are all structurally uniform by convention. A well-written contract has low burstiness because contracts are supposed to be uniform. That's not a signal of anything except genre.

It penalises constrained writers. This is the serious one. Writing in a second language, or writing to a rubric, or writing in a discipline with rigid conventions all tend to produce more uniform structure. A Stanford study found detectors falsely flagged 61.3% of human-written TOEFL essays — and burstiness is part of the mechanism. We've written about that finding in detail, because it's the strongest argument against treating any of these metrics as evidence.

Detection researchers have moved past it. Some detector vendors now argue publicly that perplexity and burstiness are insufficient as a detection basis and that their newer classifiers work differently. Take that with the appropriate amount of salt given who's saying it, but the underlying point stands: burstiness is a first-generation signal.

What this means if you're editing AI-drafted text

The useful version of this, stripped of detector framing:

Vary your sentence lengths deliberately. Read a draft and note where five sentences in a row have the same shape. Cut one to four words. Let another run long.

Don't manufacture variation. Chopping a sentence in half at a random point raises the metric and hurts the writing. The variation should track the argument — short sentences for the load-bearing claims, longer ones for development and qualification.

Watch your paragraph openings. Model output frequently starts consecutive paragraphs the same way. It's one of the most visible tells and one of the easiest things to fix.

Read it aloud. Genuinely the most reliable test available. Monotonous rhythm is immediately obvious to your ear and nearly invisible on the page.

Where we stand

RehumanizeText targets sentence-level structure, and burstiness is one of the patterns it changes. We're not going to tell you that raising a burstiness score makes text undetectable — the metric is manipulable, detectors weigh it alongside other signals, and no honest tool in this category can promise a detection outcome.

The claim we'll defend is smaller: rhythm variation makes writing better to read, and that's true whether or not anyone runs a detector over it.