Burstiness
Updated June 10, 2026
A measure of rhythm variation across a text — humans write in bursts, models write on a metronome.
Definition
Burstiness measures how much sentence-level statistics — primarily length, also structure and complexity — vary across a passage. Human writing alternates naturally: a long accumulating sentence, then a four-word one. Model output converges on a steady medium band, sentence after sentence. Low variance reads as machine.
Why models are flat
Generation optimizes for the locally probable: each sentence independently lands near the statistical center of "good sentence". Nothing pushes a model to follow a 40-word sentence with a 3-word one — that's an authorial impulse, born of emphasis and breath, not probability. The flatness is a side effect of the objective.
Using it
Burstiness is the most actionable detection concept for writers: scan your sentence lengths; if they picket-fence, vary them. It's also why crude tricks fail — random chopping creates incoherent variance that classifiers and readers both catch. A proper humanize pass rebuilds rhythm coherently, and the score shows the difference. Paired with perplexity in every explainer for a reason: together they sketch how detection actually works.