what’s burstiness?
Burstiness measures the variance of perplexity across a document: how much the difficulty and surprisingness of the writing fluctuates from sentence to sentence.
if perplexity is the average surprise per word, burstiness is the standard deviation of that surprise across a piece of writing. it asks: does the difficulty of the prose fluctuate?
human writing is bursty. we write one elegant clause, then a clunky one, then a flash of jargon, then a colloquialism, then a long flat passage of exposition. the rhythm is uneven because humans get tired, switch gears, hedge, dig in, abandon ideas mid-sentence. perplexity from word to word jumps around.
LLM-generated text tends to flow at a more uniform pace. every sentence sits in roughly the same probability band, because the same sampling temperature is producing all of them. the rhythm is flatter. burstiness is lower. detectors combine low perplexity with low burstiness to flag AI text.
this is also why non-native English speakers get flagged at much higher rates than the marketing copy admits. ESL writing runs flatter and more rhythmically uniform than native-speaker prose: the writer is working within a smaller vocabulary and a more cautious grammar. a 2023 Stanford study found AI text detectors flagged 61 to 97% of TOEFL essays from non-native English writers as AI-generated, for this structural reason.
three things a layperson needs to know:
burstiness + perplexity are the two main statistical legs of AI-text detection. almost every text classifier leans on some combination of these.
flat-rhythm writing gets flagged whether it's AI or not. detectors penalize ESL students, technical writers, legal-copy writers, and people writing in constrained formats.
burstiness is a signal of style, not provenance. a paraphraser pass can introduce burstiness into AI text and defeat the signal entirely. a heavy editing pass can do the same. burstiness measures how rhythmic the prose is. it doesn't measure where the prose came from.
bursty-equals-human and flat-equals-AI is a probability, and the false-positive rate against ESL writing is one of the field's most documented failure modes. if you're making a consequential call based on a verdict, the breakdown next to the headline number matters more than the burstiness signal alone.
questions
what is burstiness in AI detection?
burstiness measures how much the difficulty of writing fluctuates from sentence to sentence. statistically it is the variance of perplexity across a document. human writing tends to be bursty: an elegant clause, then a clunky one, then jargon, then a flat passage. LLM text runs flatter because one sampling temperature produces every sentence. detectors pair low perplexity with low burstiness to flag AI text.
why does ESL writing get flagged as AI?
non-native English prose runs flatter and more rhythmically uniform than native-speaker writing, because the writer works within a smaller vocabulary and more cautious grammar. that matches the low-burstiness pattern detectors hunt for. a 2023 Stanford study (Liang et al.) found detectors misclassified more than half of TOEFL essays from non-native writers as AI (a 61% average false-positive rate), and 97% got flagged by at least one detector.
can burstiness detection be fooled?
a paraphraser pass or a round of heavy editing introduces burstiness into AI text and defeats the signal. burstiness measures how rhythmic the prose reads, not where it came from. it tracks style, not provenance.
does flat writing mean text is AI?
bursty-equals-human and flat-equals-AI is a probability. detectors flag flat-rhythm writing regardless of origin, which penalizes ESL students, technical writers, and people working in constrained formats. this ranks among the field’s best-documented false-positive modes.
sources.
- 01Liang et al. — GPT detectors are biased against non-native English writers (Cell Patterns 2023, arXiv:2304.02819)Documented ESL false-positive rates of 61-97% on legacy detectors.
- 02
- 03