Burstiness (AI detection)

Burstiness describes how much a text varies — mostly in sentence length, sometimes in structure and rhythm. Human writing is naturally uneven: a long, winding sentence, then a short one. Fragments. Then another long one. Early language-model output tended toward uniform, medium-length sentences, so detectors adopted a rule of thumb: low burstiness reads as machine-generated; high burstiness reads as human.

Like perplexity, this heuristic penalizes exactly the people writing most carefully. Non-native English writers are often taught a safe sentence recipe — subject, verb, object, one idea per sentence, roughly the same length every time — because it minimizes grammar mistakes. Follow that advice consistently and your essay comes out rhythmically flat: correct, clear, and uniform in a way a burstiness check scores as suspicious.

Uniform: The experiment was conducted in three phases. Each phase lasted two weeks. The participants completed a survey after each phase. Bursty: The experiment ran in three phases, each two weeks long, each closing with the same short survey — and nothing about the design changed between them. That was deliberate.

Both passages are fine English. The first is what most ESL academic writing instruction produces; it is also what a detector expects from a machine.

This overlap is one mechanism behind documented detector bias: a Stanford study (Liang et al., published in Patterns, 2023) found AI detectors falsely flagged 61% of TOEFL essays by non-native English speakers.

Burstiness is almost always evaluated alongside perplexity (word-level predictability). Artificially chopping and stretching your sentences to game the metric is a losing strategy — it degrades your writing and detectors keep changing. The gap between honest, uniform learner prose and detector heuristics is a large part of what drives flagxiety; documenting your writing process addresses it more durably than stylistic tricks.